Repository navigation
Conversation
Local-Gate: v1 tree=6703d5a7b82b1192cc68aa8bb32fb8a5f8b1543a secs=31 lanes=lint:run
Ci-Attestation: {"at":1792526371,"gate":"python/all/lint","host":"linux-x86_64","key":"faac07e785ec1e972423557bb7cb2eca9b94e7a6fcf90b0689fc9110dbe1b148","lane":"lint","parents":["ee3623f6b7a485754a0a0aa3462d8cba43390e8b"],"secs":31,"stamp":"c07bd0cdfb89dc89d89c28e9bc518744","tree":"6703d5a7b82b1192cc68aa8bb32fb8a5f8b1543a","v":1,"via":"run"}
Local-Gate: v1 tree=20d414c77126a74712c5597a430e88a2ce1c6622 secs=40 lanes=lint:run
Ci-Attestation: {"at":1792526844,"gate":"python/all/lint","host":"linux-x86_64","key":"8b72b190f62addb335eee8fef10fb3cb1f7626ca542907a9438ed081ff60e25a","lane":"lint","parents":["f631fd48f79377d1738b8cafe380c40a3d8d7ab1"],"secs":40,"stamp":"7fcf2a6f9e500cca4e56b56f21080b95","tree":"20d414c77126a74712c5597a430e88a2ce1c6622","v":1,"via":"run"}
) Local-Gate: v1 tree=39069b2fd2eb2b32a95301905f562c4370416584 secs=45 lanes=lint:run Ci-Attestation: {"at":1792526986,"gate":"python/all/lint","host":"linux-x86_64","key":"acc1290a4c74a80341ac15c2697de2c7b80edc356c6d8db88fb4ef032c692712","lane":"lint","parents":["1d4d4ba4a50976e85cb27915cbd0c03aa90e3111"],"secs":45,"stamp":"219c5e7f901beb08a334be8c4ecb0d4e","tree":"39069b2fd2eb2b32a95301905f562c4370416584","v":1,"via":"run"}
Local-Gate: v1 tree=185a2b2a427e55aee2ed4db0827572e4c8faa085 secs=71 lanes=lint:run
Ci-Attestation: {"at":1792527471,"gate":"python/all/lint","host":"linux-x86_64","key":"dd2d975bf1c252bb5eb89e902bfca56e93ad6f746359921e24b974dc7bc30457","lane":"lint","parents":["d3e043675b2d374f3643e4c8d99481548386ecc5"],"secs":71,"stamp":"298b93679c3f27a691eb38571a369605","tree":"185a2b2a427e55aee2ed4db0827572e4c8faa085","v":1,"via":"run"}
zackees
added
the
ci-full
…ts (#585) Local-Gate: v1 tree=ae356b3f151f53ade04098d16d3a7b8d48f3a6d8 secs=32 lanes=lint:run Ci-Attestation: {"at":1792528190,"gate":"python/all/lint","host":"linux-x86_64","key":"ca75dca180e65fd2e2c8efc8baed969d3d309afa476a228a6b4ec50a41482745","lane":"lint","parents":["8b41221106917ea66cceec4d08e01e1f8736c998"],"secs":32,"stamp":"47b752957e8c7cae7c3a04a8b8d22edc","tree":"ae356b3f151f53ade04098d16d3a7b8d48f3a6d8","v":1,"via":"run"}
Local-Gate: v1 tree=72f146723f4e459def99022e96846c349876e48f secs=42 lanes=lint:run
Ci-Attestation: {"at":1792528576,"gate":"python/all/lint","host":"linux-x86_64","key":"d01156bffae95535acf8d582ba8ebb0a41e0a3a473b61f1e6046beea40d4a6a6","lane":"lint","parents":["26281e248fcd8569e5f09142f6ae37d678482aab"],"secs":42,"stamp":"45b3f9478fde95870d8a46ddd10304b1","tree":"72f146723f4e459def99022e96846c349876e48f","v":1,"via":"run"}
#585 progress: historical attribution and rejected geometry experimentThis is an interim research report, not candidate acceptance or completion. Frozen candidate baseline remains Historical paired comparisonsEach run used 15 paired repetitions plus a base-vs-base null arm, four logical CPUs, and the unchanged trace. Eight-worker cells are 2:1 oversubscribed. p99 is sampled stream-operation latency from a separate replay, leaving throughput/RSS runs uninstrumented. CPU/op below is process CPU divided by the fixed stream-operation count, including the harness’s existing drain CPU accounting; it is not active-worker-only CPU. Raw samples are in each linked run’s #570: run 39489801571Baseline
Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
#569: run 39489808894Baseline
Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
#572: run 39489815787Baseline
Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
Interpretation
Fresh local resident-byte attribution (5 repetitions, untimed)
These are resident snapshots, not a proof that every free byte is cold. #531 already refutes simply reducing the retention window or assuming first-fit fragmentation. Retired-body discard remains rejected because of refault CPU. Rejected: trim span geometry to the last complete blockA scratch experiment removed trailing whole slices without losing block capacity (e.g. 384-KiB blocks: 16 → 12 slices). Capacity checks and Resident-claim retry-budget screen (untimed, 5 repetitions each)I also tested the existing Current candidate and gatesDraft #609 contains exact-fit search termination and reduced page-map rewriting, plus the measurement gaps requested here. Local debug suite: 89/89; fast-path identity against the fixed baseline and pinned upstream: passed. The combined candidate has not passed performance acceptance: run 39491166866 tripped the first-sample chart-ephemeral drain-RSS gate (9.9 → 76.7 MiB). Run 39492581314 retains the same sanity threshold and now preserves all raw first-sample arms if it fails, so a failure cannot be hidden by a rerun. Full platform validation also found a missing Windows ARM64 Rust standard-library target ( |
Reuse-distance evidence and chart-ephemeral diagnosticLocal slice-lifetime trace: diagnostic onlyTemporary instrumentation at arena claim/free recorded address, slice count, timestamp, and event order into a bounded static buffer (no hooked allocations). Source policy matched candidate
Limits: these lifetime quantiles count completed reuse intervals, weighted by slices, and exclude intervals still free when tracing ends. They do not prove that all queued memory is hot, and the tracing overhead affects wall time. The allocation-count distances are also reported so the inference does not depend only on instrumented clocks. Never-reused free memory and OS residency require separate treatment. The same claim/free streams were replayed through two offline interval allocators, both respecting 512-slice chunk boundaries and preserving every request/free. First-fit versus best-fit virtual high-water footprints (MiB) were 164.06 / 168.13, 168.13 / 164.06, and 164.06 / 160.00 across the three sparse traces. Best-fit changes sign across runs and saves at most one 4.06-MiB increment here. This model ignores OS purges and physical residency; it is not a performance benchmark or proof of an implementation benefit. It gives no basis to revive #531’s best-fit proposal as the solution to the ~95-MiB snapshot queue. The traces support preserving hot reuse. They do not justify a discard policy, and none was added. Temporary tracing patch for reproductiondiff --git a/src/arena.c b/src/arena.c
index 57db5935..6cf90517 100644
--- a/src/arena.c
+++ b/src/arena.c
@@ -231,6 +231,51 @@ static size_t mi_page_full_size(mi_page_t* page) {
Arena Allocation
----------------------------------------------------------- */
+// Scratch diagnostic only: record arena-slice lifetimes without hooked allocations.
+#include <time.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <fcntl.h>
+#include <unistd.h>
+#ifndef MI_SCRATCH_TRACE_CAPACITY
+#define MI_SCRATCH_TRACE_CAPACITY (1U << 20)
+#endif
+typedef struct mi_scratch_trace_event_s {
+ uint64_t address;
+ uint64_t time_ns;
+ uint32_t slices;
+ uint32_t freed;
+} mi_scratch_trace_event_t;
+static mi_scratch_trace_event_t mi_scratch_events[MI_SCRATCH_TRACE_CAPACITY];
+static _Atomic(size_t) mi_scratch_event_count;
+static void mi_scratch_record(mi_arena_t* arena, size_t index, size_t count, bool freed) {
+ const size_t slot = mi_atomic_add_relaxed(&mi_scratch_event_count, (size_t)1);
+ if (slot >= MI_SCRATCH_TRACE_CAPACITY) return;
+ struct timespec now;
+ clock_gettime(CLOCK_MONOTONIC, &now);
+ mi_scratch_events[slot].address = (uint64_t)(uintptr_t)mi_arena_slice_start(arena, index);
+ mi_scratch_events[slot].time_ns = (uint64_t)now.tv_sec * UINT64_C(1000000000) + (uint64_t)now.tv_nsec;
+ mi_scratch_events[slot].slices = (uint32_t)count;
+ mi_scratch_events[slot].freed = freed ? 1 : 0;
+}
+__attribute__((destructor)) static void mi_scratch_dump(void) {
+ const char* path = getenv("MI_SCRATCH_TRACE");
+ if (path == NULL) return;
+ const size_t count = mi_atomic_load_relaxed(&mi_scratch_event_count);
+ if (count > MI_SCRATCH_TRACE_CAPACITY) { fprintf(stderr, "SCRATCH TRACE OVERFLOW %zu\n", count); return; }
+ const int fd = open(path, O_CREAT | O_TRUNC | O_WRONLY, 0600);
+ if (fd < 0) return;
+ const char* data = (const char*)mi_scratch_events;
+ size_t remaining = count * sizeof(mi_scratch_events[0]);
+ while (remaining > 0) {
+ const ssize_t wrote = write(fd, data, remaining);
+ if (wrote <= 0) break;
+ data += wrote;
+ remaining -= (size_t)wrote;
+ }
+ close(fd);
+}
+
// #493 (strategy 9): resident-first claiming. A freed range sits in `slices_purge` (young) or
// `slices_purge_aged` for the #486 retention window: free, and still resident. The plain search
// (`mi_bbitmap_try_find_and_clearN`) knows nothing of that and, measured on short-lived threads,
@@ -372,6 +417,7 @@ static mi_decl_noinline void* mi_arena_try_alloc_at(
size_t slice_index;
const bool resident = mi_arena_try_claim_resident(arena, slice_count, &slice_index);
if (!resident && !mi_bbitmap_try_find_and_clearN(arena->slices_free, tseq, slice_count, &slice_index)) return NULL;
+ mi_scratch_record(arena, slice_index, slice_count, false);
if (!arena->memid.is_pinned) { mi_arena_unqueue_purge(arena, slice_index, slice_count); }
#if MI_DIAGNOSTICS
mi_arena_count_claim(arena, slice_index, slice_count, resident); // before the dirty bits are set below
@@ -1956,6 +2002,7 @@ static void mi_arenas_free_ex(mi_subproc_t* subproc, void* p, size_t size, mi_me
}
// and make it available to others again
+ mi_scratch_record(arena, slice_index, slice_count, true);
bool all_inuse = mi_bbitmap_setN(arena->slices_free, slice_index, slice_count);
if (!all_inuse) {
_mi_error_message(EAGAIN, "trying to free an already freed arena block: %p, size %zu\n", mi_arena_slice_start(arena,slice_index), mi_size_of_slices(slice_count));Chart-ephemeral diagnostic: still not acceptedRun 39492581314 compared frozen baseline This run did not reproduce run 39491166866’s extreme first-sample drain jump (9.9 → 76.7 MiB). Across these 25 samples, drain RSS ranged 35.8–66.0 MiB in the baseline, 35.2–63.0 in the candidate, and 30.0–71.4 in the null. That variability is evidence to investigate the original failure, not permission to erase it or declare a rerun sufficient.
Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
The overall verdict remains INCONCLUSIVE, because the p99 interval slightly exceeds its null-derived allowance. No memory credit is used for latency. The broader candidate/control run is still in flight. |
Local-Gate: v1 tree=2a860088093b035658f80003fdf54ca50f204512 secs=42 lanes=lint:run
Ci-Attestation: {"at":1792529727,"gate":"python/all/lint","host":"linux-x86_64","key":"952590386b98d9a41d039814c586d336303ba383051ce1306abf12e022169e7b","lane":"lint","parents":["9b67e411fc8c8899e07e05d528a4c5b37ab20417"],"secs":42,"stamp":"f082fcdd98116d222fc4528c2dbef13c","tree":"2a860088093b035658f80003fdf54ca50f204512","v":1,"via":"run"}
Local-Gate: v1 tree=6f35648090e34aee7fc5f7eef5b3957514301069 secs=39 lanes=lint:run
Ci-Attestation: {"at":1792530060,"gate":"python/all/lint","host":"linux-x86_64","key":"1d9079fff997c8f9566db08126eb75d6b20caf8008a81e5afa7f5e46b921362c","lane":"lint","parents":["6310bebeaea2961170bf83175cef0065ca7d3ecd"],"secs":39,"stamp":"52b1d7992e5081160c35fd0d21029ab4","tree":"6f35648090e34aee7fc5f7eef5b3957514301069","v":1,"via":"run"}
Exact-fit-only acceptance failure; compact-singleton experiment rejectedExact-fit-only paired run: FAIL, despite workflow successRun 39489674226 compared frozen baseline The small control fails: CPU/op +20.3% [+16.7%, +26.9%] against a 4.8% null allowance; throughput −18.1% [−22.6%, −14.1%] against a 5.2% allowance. Fast-path byte identity does not override this measured regression. I have not established its cause, and the exact-fit-only revision is not accepted. Large-class/8 does improve: CPU −1.4% [−1.8%, −0.9%] and throughput +2.5% [+1.2%, +3.5%], with that row’s full ledger PASS. Several ephemeral/chart/profiler latency or control intervals remain INCONCLUSIVE. Those results cannot be called a global pass.
CPU/op is the artifact’s existing process CPU measurement divided by fixed stream operations, including the existing short-drain CPU accounting. It is not active-worker-only CPU. The ledger below is computed from all recorded paired samples with the same-run null noise; latency earns no memory credit. Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
Compact-singleton prototype: structural success, memory failureThe page layout currently reserves an inline-header slice for singleton allocations even when arena page metadata is separate. A prototype reduced only arena singleton reservations to their block extent, while keeping the original OS-fallback reservation and excluding large-alignment/secure cases. The structural probe failed on the original geometry (all five tested sizes had one extra slice) and passed with compaction, including last-byte lookup/write/free and an OS-fallback allocation. However, five untimed byte-attribution repetitions per arm showed the wrong memory direction:
Rejected: the sparse snapshot rose about 6.6%, from 177.07 to 188.82 MiB, despite a smaller reservation per singleton. The queue grew too. Tighter geometry is insufficient evidence of lower resident memory; placement/reuse/THP interactions need attribution before a redesigned policy. No defaults were changed, no discard policy was introduced, and the prototype has been restored out of the worktree. These local byte probes are not timing acceptance. Rejected compact-singleton patch for reproductiondiff --git a/src/arena.c b/src/arena.c
index 57db5935..719f8985 100644
--- a/src/arena.c
+++ b/src/arena.c
@@ -941,16 +941,22 @@ static uint8_t* mi_arenas_page_alloc_fresh_area(mi_theap_t* theap, size_t slice_
uint8_t* start = NULL;
*memid = _mi_memid_none();
const size_t alloc_size = mi_size_of_slices(slice_count);
+ size_t arena_slice_count = slice_count;
+ #if MI_PAGE_META_IS_SEPARATED && MI_SECURE < 2 && defined(MI_SINGLETON_COMPACT_ARENA) && MI_SINGLETON_COMPACT_ARENA
+ if (!os_align && block_size > MI_LARGE_MAX_OBJ_SIZE) {
+ arena_slice_count = mi_slice_count_of_size(block_size);
+ }
+ #endif
if (!mi_option_is_enabled(mi_option_disallow_arena_alloc) && // allowed to allocate from arena's?
!os_align && // not large alignment
slice_count <= mi_arena_max_object_size()/MI_ARENA_SLICE_SIZE) // and not too large
{
- start = (uint8_t*)mi_arenas_try_alloc(heap, slice_count, page_alignment, commit, allow_large, req_arena, tld->thread_seq, numa_node, memid);
+ start = (uint8_t*)mi_arenas_try_alloc(heap, arena_slice_count, page_alignment, commit, allow_large, req_arena, tld->thread_seq, numa_node, memid);
if (start != NULL) {
mi_arena_pages_t* const arena_pages = mi_heap_ensure_arena_pages(heap, memid->mem.arena.arena);
*parena_pages = arena_pages;
if (arena_pages==NULL) {
- _mi_arenas_free(heap->subproc, start, mi_size_of_slices(slice_count), *memid); // roll back
+ _mi_arenas_free(heap->subproc, start, mi_size_of_slices(arena_slice_count), *memid); // roll back
start = NULL;
}
else {
@@ -1025,11 +1031,14 @@ static void mi_arenas_free_ex(mi_subproc_t* subproc, void* p, size_t size, mi_me
static mi_page_t* mi_arenas_page_alloc_fresh(mi_theap_t* theap, size_t slice_count, size_t block_size, size_t block_alignment, bool commit)
{
const bool os_align = (block_alignment > MI_PAGE_MAX_OVERALLOC_ALIGN);
- const size_t alloc_size = mi_size_of_slices(slice_count);
mi_memid_t memid = _mi_memid_none();
mi_arena_pages_t* arena_pages = NULL;
uint8_t* const slice_start = mi_arenas_page_alloc_fresh_area(theap,slice_count,block_size,block_alignment,os_align,commit,&memid,&arena_pages);
if (!slice_start) return NULL;
+ #if MI_PAGE_META_IS_SEPARATED && MI_SECURE < 2 && defined(MI_SINGLETON_COMPACT_ARENA) && MI_SINGLETON_COMPACT_ARENA
+ if (memid.memkind == MI_MEM_ARENA) { slice_count = memid.mem.arena.slice_count; }
+ #endif
+ const size_t alloc_size = mi_size_of_slices(slice_count);
// guard page at the end of mimalloc page?
#if MI_SECURE>=5Cross gate and ongoing evidenceCross run 39496948660 passed, including the Windows ARM64 build, after disabling only its solo toolchain cache. All toolchain/Cargo operations remain through soldr. This validates the documented fresh-toolchain mitigation for this run; the upstream cache defect remains tracked separately. Combined-candidate target/control/random-sparse runs 39495904856 / 39495909935 / 39495916158 and same-run competitor scaling 39494686018 remain active. They measure different candidate C code from this exact-fit-only run. Their results are still required; none of the failed or inconclusive evidence above is erased. The original combined-candidate drain-RSS sanity failure also remains unresolved. |
Local-Gate: v1 tree=d5da40744be8c8b1cadd5f4816e4b21a4e882420 secs=40 lanes=lint:run
Ci-Attestation: {"at":1792531183,"gate":"python/all/lint","host":"linux-x86_64","key":"a85874d7c9d12ba8608efddba2a3d24899f43c6fc895c4f8478d67ac1477fff4","lane":"lint","parents":["2adb4c009882f752f5f6146f871f85133c05e61a"],"secs":40,"stamp":"c0bdfb136b38cbd62216b72d1d29a49f","tree":"d5da40744be8c8b1cadd5f4816e4b21a4e882420","v":1,"via":"run"}
#585 combined-candidate controls and random/sparse resultsFrozen baseline Run 39495909935Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
Run 39495916158Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
Interpretation and attribution
|
#585 combined-candidate large-class resultsRun 39495904856 completed against the frozen baseline with the original sanity gate and same-run null. Four of five rows pass. Main large-class/8: paired CPU/op -2.0% [-4.2%, -0.1%], throughput +3.3% [+0.3%, +4.3%], p99 -1.1% [-1.7%, +0.0%]; no established peak-memory reduction. The chart ephemeral row remains INCONCLUSIVE on minor faults, p99 and peak RSS. It is being narrowed to one row in 39500745637, 25 initial/up to 45 paired samples, unchanged thresholds. Candidate is not accepted yet: control/random/sparse follow-ups, competitor scaling and the earlier failed-run attribution remain outstanding. The older exact-fit-only small-control failure is retained. Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
|
Local-Gate: v1 tree=b29bf3ca7c1551169059eb68f52a73ee7b1c1a44 secs=35 lanes=lint:run
Ci-Attestation: {"at":1792532275,"gate":"python/all/lint","host":"linux-x86_64","key":"cf99b163b801d3af55d58955f96afcc2dd5f608104268a1718524862df58dead","lane":"lint","parents":["3bbe4e727b6b2364b3c8c4b3e4cf219666718014"],"secs":35,"stamp":"0b49181e657c4af4500410ae3dad0379","tree":"b29bf3ca7c1551169059eb68f52a73ee7b1c1a44","v":1,"via":"run"}
Local-Gate: v1 tree=9c3ad8c5176f7fc64846dfaad6ef2ff9725c34de secs=43 lanes=lint:run
Ci-Attestation: {"at":1792532501,"gate":"python/all/lint","host":"linux-x86_64","key":"0da51912007b2d69f8bed1d4a1df35ac25a25869688d5232c8edf4247570d8e7","lane":"lint","parents":["acd8300e5e8c7bdbc2293456c13741502ecb60f3"],"secs":43,"stamp":"8cad2435a94ce537259768a4e5b1dfd1","tree":"9c3ad8c5176f7fc64846dfaad6ef2ff9725c34de","v":1,"via":"run"}
#585 measurement-order correction and exact executable preservationA concrete harness defect was found while investigating the failed exact-fit-only small control: The regression test ran the actual measurement loop: RED head-position counts CI now also preserves the exact timing and latency executables (per arm/build kind) and SHA256 digests in Review: clean from the same single primary reviewer. Local gate on new head Corrected-order measurements retain baseline
Older live follow-ups 39500299084, 39500303123 and 39500745637 remain intact and will be reported under the legacy protocol. No previous sample/run was cancelled to replace an unfavorable result. Acceptance remains pending. |
#585 same-host competitor scaling and memory prioritiesRun 39494686018 completed successfully: 15 paired blocks, four patterns, worker counts 1/2/3/4/6/8, five allocators, 1,800 samples. Status is diagnostic/not publishable; nothing was published to benchmark-stats. Host: AMD EPYC 9V74, two physical/four logical cores; eight workers are 2:1 oversubscribed on logical CPUs. The temporary harness merge References are pinned: TCMalloc Audit: zero operation-count, operations-per-worker, checksum, worker-seed or size-histogram mismatches across all 360 paired blocks. All 120 allocator/cell schedules place each allocator in each of five execution positions exactly three times; this scaling scheduler does not have perf-ab's legacy fixed-middle defect. Findings
This run measures memory and throughput; CPU/op, faults and p99: not measured by scaling. It is not a rule-12 acceptance ledger. Corrected-order perf-ab runs supply those separate measurements and remain pending (order correction: #585 (comment)). Per-cell paired ratiosMedians and 95% sign-test intervals on paired ratios; ratio-of-medians need not equal the median paired ratio. Raw samples and telemetry replays are in the linked run artifacts. Trace parity audit: 0 mismatches across 360 paired blocks.
|
#585 corrected exact-fit control and baseline-only drain-RSS investigationCorrected-order run 39502542281 collected 60 actual paired samples per arm on EPYC 7763 / four logical CPUs. This is the same CPU model as the earlier +20.3% exact-fit-only CPU failure, not the same physical VM. Head remains CPU/op passes its same-run allowance at -0.3% [-2.2%, +1.0%]; the 20% CPU increase was not reproduced. Overall the row remains INCONCLUSIVE on throughput and p99. It is not standalone exact-fit acceptance. Old and corrected protocols are kept separate and cannot be pooled. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
The actual timed executables in Baseline-only reproduction of drain variationA local memory diagnostic used one unchanged frozen-baseline chart binary (pprof and memory-events compiled ON, runtime OFF), identical arguments, and identical two-CPU affinity for six consecutive fresh processes. The binary SHA256 is
Consecutive repetitions 1 and 2 changed drain RSS 8.934 → 68.867 MiB, while bound RSS stayed 5.871 → 5.867 MiB. Applying the existing Code context: Reproduction: build frozen baseline with the current taskset -c 0,1 ./perf_ab 8 8 98304 524288 400000 0 0 uniform 8 1300Choose two allowed CPUs on the host, keep All nine exploratory observations (same frozen baseline, varied affinity)[
{
"rep": 0,
"cpus": [
0
],
"peak_MiB": 58.83984375,
"drain_MiB": 2.94140625,
"bound_MiB": 2.94140625,
"release_ms": 450.0
},
{
"rep": 0,
"cpus": [
0,
1,
2,
3
],
"peak_MiB": 69.69140625,
"drain_MiB": 69.69140625,
"bound_MiB": 4.75390625,
"release_ms": 1030.0
},
{
"rep": 0,
"cpus": [
0,
1
],
"peak_MiB": 67.640625,
"drain_MiB": 64.1875,
"bound_MiB": 4.75,
"release_ms": 710.0
},
{
"rep": 1,
"cpus": [
0,
1
],
"peak_MiB": 72.76953125,
"drain_MiB": 63.44140625,
"bound_MiB": 4.75390625,
"release_ms": 1090.0
},
{
"rep": 1,
"cpus": [
0,
1,
2,
3
],
"peak_MiB": 71.49609375,
"drain_MiB": 70.90234375,
"bound_MiB": 4.75,
"release_ms": 1060.0
},
{
"rep": 1,
"cpus": [
0
],
"peak_MiB": 60.1953125,
"drain_MiB": 56.30859375,
"bound_MiB": 4.74609375,
"release_ms": 1090.0
},
{
"rep": 2,
"cpus": [
0
],
"peak_MiB": 62.23828125,
"drain_MiB": 59.1875,
"bound_MiB": 4.75,
"release_ms": 720.0
},
{
"rep": 2,
"cpus": [
0,
1,
2,
3
],
"peak_MiB": 71.515625,
"drain_MiB": 52.4375,
"bound_MiB": 4.75,
"release_ms": 1020.0
},
{
"rep": 2,
"cpus": [
0,
1
],
"peak_MiB": 72.6171875,
"drain_MiB": 6.87109375,
"bound_MiB": 5.87109375,
"release_ms": 450.0
}
]Independent truncated-prefix attributionTable values are medians across three independent untimed runs per prefix. Current C source unchanged from
Persistent unformed resident capacity is already ~27 MiB after 2,000 operations per worker. A policy targeting it still has to preserve cross-bin reuse and pass the full per-cell ledger; the earlier smaller-span and compact-singleton prototypes did not meet those requirements. No such policy was retained. Corrected candidate runs and corrected historical bisections remain live. All numerical results, including failed/inconclusive ones, remain in the investigation. |
#585 corrected historical #572 bisection: latency debt remainsRun 39504129574 completed on EPYC 9V74 / four logical CPUs, comparing exact pre-merge The plain large-class/8 row has a p99 regression of +5.2% [4.4%, 5.2%], beyond its 0.7% same-run noise allowance. Throughput, CPU/op, peak RSS, faults and drain RSS pass that historical cell's allowances. The profiler-on row passes all six metrics. This identifies historical latency debt at the #572 merge; it does not isolate a particular function and is not a current-candidate-versus-frozen-baseline failure. Current candidate acceptance remains pending in the three separately listed corrected runs. Earlier fixed-middle-order bisections remain diagnostic and are not pooled with this result. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
Actual sample and latency medians: Both |
#585 corrected historical #569 bisection: both measured rows passRun 39504125686 compares exact pre-merge Both rows pass their full six-metric ledgers. Plain peak RSS improves 5.3%, with CPU/op +0.7% [0.1%, 1.3%] inside its 1.6% allowance; p99 is within same-run noise. This historical result does not validate the combined current candidate, whose acceptance runs are still live. Earlier fixed-middle-order results remain diagnostic and are not pooled. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
The run retains |
#585 corrected historical #570 bisection: CPU improves, full ledger remains inconclusiveRun 39504122084 compares exact pre-merge CPU/op is lower beyond noise: -2.1% [-2.5%, -1.8%] plain, -1.0% [-1.4%, -0.7%] profiler-on. Throughput is also higher. This supports the narrow conclusion that moving the retired-page clock read off the owner path paid down CPU cost in these measured rows. It does not establish full-metric equivalence: plain p99 and 0.5-second drain RSS, and profiler-on faults and p99, remain INCONCLUSIVE at the sample cap. Equal rounded p99 medians are not proof of equivalence. This is not a standalone acceptance result. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
The run retains the exact measured timing/latency executables in Corrected historical attribution so far
These are separate same-run comparisons, not one common physical VM or a pool: #570 ran on EPYC 7763; #569/#572 on EPYC 9V74. Their deltas cannot be added into a causal decomposition of the whole fork/upstream gap. #544 is the coordinating issue for the span follow-ups; #569 and #570 are its merged changes. Earlier fixed-middle-order bisections remain diagnostic only. The three corrected combined-candidate-versus-frozen-baseline runs remain live. Those results are required for candidate acceptance; historical attribution does not substitute for them. |
#585 corrected combined-candidate controls: two PASS, two INCONCLUSIVERun 39502534567 compares frozen Actual samples per arm: small/8 60; larson/1 60; larson/8 30; larson/8 chart 35. Eight workers are 2:1 oversubscribed. Both Larson/8 rows PASS all six metrics. Small/8 and Larson/1 remain INCONCLUSIVE and are not accepted. Small/8 throughput and CPU pass, but p99, peak and drain RSS remain unresolved; Larson/1 remains unresolved on peak RSS. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
Do not interpret rounded zeros as equivalenceSmall/8 p99 medians are 91 ns for base, head and null, yet the full-precision paired head interval is [0%, 1.111111%], with a zero null median interval. Peak/drain head intervals are [0%, 0.0237023%], with a zero null interval. Peak-RSS medians are base 16.48046875 MiB, head 16.484375 MiB and null 16.484375 MiB; the one-page difference is unresolved, not waived. Larson/1 peak interval is [-0.0356532%, 1.2222090%] against 1.1770021% noise. All eight actual executable SHA256 hashes were verified. In the plain timing binaries, head text is 64 bytes larger; writable data and BSS are identical, and all LOAD segment addresses/page counts match. That excludes a different mapped segment page count as an explanation; it does not prove which physical pages were resident or explain latency. Two focused follow-ups retain the same exact source pair, null arm, original work traces and thresholds, requesting 60 samples with adaptive maximum 120:
The original result remains visible. Pooling, if applicable, must first pass the harness compatibility checks for CPU model, logical CPUs, refs, flags, latency protocol and arm-order protocol. No different-host model or legacy-order samples will be pooled. Candidate acceptance is still pending, including the two original live target/random runs. Untimed residency attribution using the actual control binariesA local diagnostic replayed the downloaded plain timing executables, unchanged and hash-verified, for three alternating paired blocks with the original small/8 arguments and the same four allowed CPUs. It sampled
The unchanged base itself had 71, 71 and 60 present executable-file pages. In the third base process, file offsets 0x0–0x2000 (read-only) and 0x18000–0x1f000 (executable) were absent from the final snapshot, unlike the first base process. This reproduces baseline process-residency variation; it does not establish the source of the original CI one-page interval. Anonymous mapping RSS is not a measurement of allocator memory alone. No threshold was waived and the control remains INCONCLUSIVE pending the focused CI follow-ups. |
#585 focused combined-candidate Larson/1: PASS on EPYC 7763Run 39509440439 collected 60 actual samples per arm, including null, comparing frozen Larson/1 PASSes all six metrics. This is an independent same-host paired result, not a pool with the earlier EPYC 9V74 control run. That earlier 60-sample row remains INCONCLUSIVE on peak RSS and stays visible: #585 (comment) . It is not reclassified as a pass on EPYC 9V74. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
The run preserves raw samples and all four actual timing/latency binaries. Each SHA256 digest was verified and each binary matches its counterpart from the original corrected control run byte for byte. Thus no allocator, benchmark binary or work-trace change was made for this follow-up; the result does not establish why the earlier host's RSS interval remained unresolved. Candidate acceptance still awaits the original large-class and random/sparse runs and the focused small/8 run. The original control ledger and its limitations remain part of the record; no memory credit or threshold exception was given to this control. |
#585 focused combined-candidate small/8: PASS on EPYC 7763Run 39509436788 collected 60 actual samples per arm, including null, comparing frozen Small/8 PASSes all six metrics. Full-precision paired p99 interval is exactly [0%, 0%], against zero null noise. Peak and drain RSS intervals are [-0.02387205%, 0%], also against zero null noise. CPU/op is -0.7% [-2.0%, +0.5%] and throughput -0.9% [-1.9%, +1.9%], each within 2.3% same-run noise. No memory credit or quantization exception was applied. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All four actual timing/latency executable hashes were verified and match the original corrected control run byte for byte. Raw samples and exact executables are retained as this run's artifacts. This is an independent same-host result, not a pool with the earlier EPYC 9V74 run; that run remains INCONCLUSIVE on small p99/RSS and Larson/1 peak RSS: #585 (comment) . The new result does not establish the cause of the earlier host's unresolved intervals or the legacy discarded-binary CPU failure. The focused Larson/1 run also independently PASSed at 60 samples per arm: #585 (comment) . Both Larson/8 rows PASSed in the original corrected control run. Thus every control has a conclusive PASS in a corrected run, with the earlier unresolved rows retained rather than reclassified. These are per-cell results from separate VMs; no pooled or aggregate cross-host acceptance claim is made. Combined-candidate acceptance still awaits the original large-class and random/sparse runs. No allocator or benchmark binary was changed to obtain these control follow-ups. |
#585 corrected combined-candidate random/sparse: two PASS, two INCONCLUSIVERun 39502538433 compares frozen Actual samples per arm: random-large/8 30; bursty/8 60; random-large/1 60; sparse 64-slot/8 60. Random-large/8 and random-large/1 PASS all six metrics. Bursty/8 remains INCONCLUSIVE on p99; sparse 64-slot/8 remains INCONCLUSIVE on CPU/op and faults. All their peak/drain memory metrics pass their same-run noise allowances, but no row earns memory credit. None of the unresolved cells is accepted. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All four preserved actual timing/latency executable hashes were verified and match the original corrected control run byte for byte, despite the different CPU model. These results are not pooled across models and do not establish an explanation for other hosts' intervals. Focused single-row follow-ups are predeclared at 60 samples, adaptive maximum 120, with the same exact source pair, work traces, original sanity/ledger thresholds, balanced order, null arm and binary preservation:
Only compatible metadata permits pooling; otherwise report independent same-host paired results. The original 60-sample inconclusive rows remain in this record. The large-class target run is still live. No source or benchmark executable was changed for these follow-ups. |
Chart ephemeral: standard row PASS; short generations remain INCONCLUSIVERun 39520756150 completed on AMD EPYC 7763 / four logical CPUs. The standard chart ephemeral row PASSes all six metrics at 80 samples per arm, CPU -1.4%. Short generations remain INCONCLUSIVE at the maximum 120 per arm, on minor faults, peak RSS and after-drain RSS. CPU, throughput and p99 PASS. Both rows have zero memory credit. Fixed baseline base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
Focused exact short-generation follow-up: https://github.com/zackees/mimalloc-pprof/actions/runs/39526576845, 120-to-240 samples/arm, same source pair, original trace and null/sanity gates. This raises the sample budget without changing an acceptance threshold. The prior INCONCLUSIVE result remains visible; no cross-host pooling. Actual executable hashes: {
"base/chart/perf_ab": "306d5c8a089d29a7ff276df621226fe2d08ffe7013669627f38dc5c0020d33f7",
"base/chart/perf_ab_latency": "2f883b21f4d48e9e1b404c3110ca058902c68cba23e99a82edd58026c7495090",
"head/chart/perf_ab": "51b9dde3f2ce5e446eb8443af7ab56136af1b146bbdb61acad221d2a652a0bff",
"head/chart/perf_ab_latency": "41d91d58314a4ca202ec6b0e4a8a3d62995a39a3aac891edab560f22c0500240"
}Full-precision cells and actual arm counts{
"large-class-ephemeral/8 (chart build)": {
"n": {
"base": 80,
"head": 80,
"null": 80
},
"saved_percent": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 8090087.300000001,
"head_median": 8236055.25,
"change": [
1.5976350413375746,
1.0741514386580004,
2.021272121039053
],
"regression": [
-2.021272121039053,
-1.0741514386580004
],
"allowance": 0.7038137898227369,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 1.5591499999999998,
"head_median": 1.5343499999999999,
"change": [
-1.4428242792762553,
-1.7187400166123523,
-1.2178619756427695
],
"regression": [
-1.7187400166123523,
-1.2178619756427695
],
"allowance": 0.1352134440795822,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 252.0,
"head_median": 252.0,
"change": [
0.0,
-0.4,
0.7905138339920948
],
"regression": [
-0.4,
0.7905138339920948
],
"allowance": 0.796812749003984,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 1853.0,
"head_median": 1843.5,
"change": [
-0.5396654074473827,
-1.1154489682097044,
0.48543689320388345
],
"regression": [
-1.1154489682097044,
0.48543689320388345
],
"allowance": 1.6008537886872998,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 76.466796875,
"head_median": 74.6328125,
"change": [
-0.028803626883975326,
-2.5002551280742935,
0.07141399714344011
],
"regression": [
-2.5002551280742935,
0.07141399714344011
],
"allowance": 2.482251391797334,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 8.7890625,
"head_median": 7.85546875,
"change": [
-0.05361540852510731,
-11.297440423654017,
0.0
],
"regression": [
-11.297440423654017,
0.0
],
"allowance": 12.729985082048731,
"verdict": "PASS"
}
]
},
"large-class-ephemeral/8 short generations (chart build)": {
"n": {
"base": 120,
"head": 120,
"null": 120
},
"saved_percent": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 7722059.1,
"head_median": 7802026.4,
"change": [
1.2275152243972343,
0.6723567289540237,
1.6046831424223438
],
"regression": [
-1.6046831424223438,
-0.6723567289540237
],
"allowance": 0.7108754433641795,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 0.423,
"head_median": 0.4168,
"change": [
-1.4744072052498967,
-1.7009213323883732,
-1.1461988304093604
],
"regression": [
-1.7009213323883732,
-1.1461988304093604
],
"allowance": 0.6850933144342106,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 250.0,
"head_median": 251.0,
"change": [
0.1984126984126984,
-0.3968253968253968,
1.1627906976744187
],
"regression": [
-0.3968253968253968,
1.1627906976744187
],
"allowance": 0.7843137254901961,
"verdict": "INCONCLUSIVE"
},
{
"metric": "p99 operation ns",
"base_median": 1834.0,
"head_median": 1824.0,
"change": [
-0.05437778098439455,
-1.0341261633919339,
0.5254860746190226
],
"regression": [
-1.0341261633919339,
0.5254860746190226
],
"allowance": 1.0309278350515463,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 74.6015625,
"head_median": 74.6015625,
"change": [
0.0,
-0.03058727569331158,
2.6056804854418436
],
"regression": [
-0.03058727569331158,
2.6056804854418436
],
"allowance": 2.544478680051685,
"verdict": "INCONCLUSIVE"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 74.6015625,
"head_median": 74.6015625,
"change": [
0.0,
-0.005380104374024856,
2.6056804854418436
],
"regression": [
-0.005380104374024856,
2.6056804854418436
],
"allowance": 2.544478680051685,
"verdict": "INCONCLUSIVE"
}
]
}
}Overall acceptance remains pending; PR #609 remains draft. Platform CI is 104 successes / five intentional skips; no source or memory setting changed. |
Current candidate: main row PASS, profiler p99 FAILRun 39520753532, frozen base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All eight actual measured executable hashes verify. Plain binaries match all other ordinary new-candidate runs byte for byte; baseline binaries match the previous original-target run. Row parameters, balanced arm positions, separate p99 replay and thresholds are unchanged. Eight workers on four logical CPUs is 2:1 oversubscription. No cross-host pooling. The failed frozen-baseline comparison will not be reclassified or repeated to seek a pass. The next comparison isolates old prefix-reuse candidate {
"base/plain/perf_ab": "7f0c70ce2450dfa09126c1674bddab53187614f12d5bd648b4fda33ccbd3c712",
"base/plain/perf_ab_latency": "ef2d5d273681a473e253d1dd2080307f1ae9f09d34977bda2ad4cab9cbfe7823",
"base/pprof/perf_ab": "c6784c2c2cb2619d31ce4bc679f840ad7d9a2cf04aba6ce4dda8c374d5c8ba37",
"base/pprof/perf_ab_latency": "f4e223e4c1489619ef2cd9831bb88a204931fa207d7ec18a17411e2899eef12f",
"head/plain/perf_ab": "9ee3617e371e93195de1374ef403455e7ae660848d689ce4bf0ae0ce2f8565b1",
"head/plain/perf_ab_latency": "c14766d194fd7f3c07652b148d1de18e3e4db90b4414bd563398f5f44ee9b982",
"head/pprof/perf_ab": "ef1f0135f09bec19ff3cba3c26a211359243cbcb40825818f9bddb3c512771ba",
"head/pprof/perf_ab_latency": "1f432e1f3a4c857bde03dd44859eac39b92a510806cceb4841087d16e98d7a8e"
}Full-precision cells and actual per-arm counts{
"large-class/8": {
"n": {
"base": 60,
"head": 60,
"null": 60
},
"saved_percent": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 6667628.9,
"head_median": 6678621.199999999,
"change": [
0.3412312144181615,
-0.5394161115517625,
1.1655071010572247
],
"regression": [
-1.1655071010572247,
0.5394161115517625
],
"allowance": 1.0053539940084428,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 1.8851,
"head_median": 1.8780999999999999,
"change": [
-0.9400186412634518,
-1.5480674957428158,
0.04700229789011389
],
"regression": [
-1.5480674957428158,
0.04700229789011389
],
"allowance": 0.7734279811410684,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 224.5,
"head_median": 222.0,
"change": [
0.0,
-2.727272727272727,
2.2522522522522523
],
"regression": [
-2.727272727272727,
2.2522522522522523
],
"allowance": 15.916955017301039,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 2053.0,
"head_median": 2033.0,
"change": [
-0.9862472489756378,
-1.872299567930869,
0.0
],
"regression": [
-1.872299567930869,
0.0
],
"allowance": 1.4541929229277752,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 74.43359375,
"head_median": 72.51171875,
"change": [
-0.02623501326673862,
-2.658486707566462,
0.07017165065313613
],
"regression": [
-2.658486707566462,
0.07017165065313613
],
"allowance": 2.6075550891920254,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 63.111328125,
"head_median": 62.298828125,
"change": [
-0.9369928771022332,
-3.7815889946063876,
1.0305269298075053
],
"regression": [
-3.7815889946063876,
1.0305269298075053
],
"allowance": 3.0231217575097675,
"verdict": "PASS"
}
]
},
"large-class/8 (profiler on)": {
"n": {
"base": 120,
"head": 120,
"null": 120
},
"saved_percent": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 4729425.15,
"head_median": 4753452.4,
"change": [
0.16241125103154308,
-0.2786631451692534,
0.7931977420037137
],
"regression": [
-0.7931977420037137,
0.2786631451692534
],
"allowance": 0.6542309515856304,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 2.65735,
"head_median": 2.6506499999999997,
"change": [
0.0077255990942455555,
-0.5575048732943462,
0.27628434886499675
],
"regression": [
-0.5575048732943462,
0.27628434886499675
],
"allowance": 0.6716220805191795,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 1632.5,
"head_median": 1629.0,
"change": [
-0.09200881441042014,
-0.30731407498463426,
0.24479804161566704
],
"regression": [
-0.30731407498463426,
0.24479804161566704
],
"allowance": 0.676923076923077,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 23811.0,
"head_median": 26675.5,
"change": [
11.288528431618904,
4.211122651307714,
15.721972731318527
],
"regression": [
4.211122651307714,
15.721972731318527
],
"allowance": 3.946504758779127,
"verdict": "FAIL"
},
{
"metric": "peak RSS MiB",
"base_median": 75.6328125,
"head_median": 74.02734375,
"change": [
-0.021476510871790477,
-2.471042471042471,
0.4027130139889784
],
"regression": [
-2.471042471042471,
0.4027130139889784
],
"allowance": 2.218438838157148,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 72.4296875,
"head_median": 72.029296875,
"change": [
-0.27955039925074177,
-0.9455248737509402,
0.6826302597194817
],
"regression": [
-0.9455248737509402,
0.6826302597194817
],
"allowance": 1.3112415647116915,
"verdict": "PASS"
}
]
}
} |
Profiler p99 failure: local cycle attribution, not acceptanceCurrent frozen-baseline candidate failed profiler p99 in 39520753532, +11.3% [4.2%,15.7%] against 3.9% noise. The six completed local samples below use the exact failed baseline/candidate latency executables, three alternating pairs on Ryzen 7 3700X with four-CPU affinity, profiler enabled and the original profiler row. User-mode sampled cycles exclude kernel work and perturb execution. No local latency is an acceptance result. Fresh-page cycle shares are lower in all three candidate runs (3.85/2.79/4.30% baseline vs 1.88/1.83/2.99% candidate); pthread mutex-lock shares are mixed (1.45/2.02/1.42% vs 1.96/1.88/1.22%). Fractions are not per-operation costs. These records do not establish the cause of the CI p99 regression. No source change follows from this local screen. The source-isolation comparison 39526717341 measures old prefix-reuse candidate Manifest: {
"cpus": "0,1,2,3",
"environment": {
"MIMALLOC_PROF": "1"
},
"purpose": "Local sampled cycles only; perturbed p99 is not acceptance",
"hashes": {
"base": "f4e223e4c1489619ef2cd9831bb88a204931fa207d7ec18a17411e2899eef12f",
"head": "1f432e1f3a4c857bde03dd44859eac39b92a510806cceb4841087d16e98d7a8e"
},
"runs": [
{
"rep": 0,
"arm": "base",
"command": [
"perf",
"record",
"-q",
"-e",
"cycles:u",
"-c",
"100000",
"-o",
"/tmp/mimalloc-585/profiler-cycle-attribution/0-base.data",
"--",
"taskset",
"-c",
"0,1,2,3",
"/tmp/mimalloc-585/balanced-byte-threshold-main/builds/base/pprof/perf_ab_latency",
"8",
"1",
"98304",
"524288",
"400000",
"0",
"0",
"uniform",
"8",
"1300"
]
},
{
"rep": 0,
"arm": "head",
"command": [
"perf",
"record",
"-q",
"-e",
"cycles:u",
"-c",
"100000",
"-o",
"/tmp/mimalloc-585/profiler-cycle-attribution/0-head.data",
"--",
"taskset",
"-c",
"0,1,2,3",
"/tmp/mimalloc-585/balanced-byte-threshold-main/builds/head/pprof/perf_ab_latency",
"8",
"1",
"98304",
"524288",
"400000",
"0",
"0",
"uniform",
"8",
"1300"
]
},
{
"rep": 1,
"arm": "head",
"command": [
"perf",
"record",
"-q",
"-e",
"cycles:u",
"-c",
"100000",
"-o",
"/tmp/mimalloc-585/profiler-cycle-attribution/1-head.data",
"--",
"taskset",
"-c",
"0,1,2,3",
"/tmp/mimalloc-585/balanced-byte-threshold-main/builds/head/pprof/perf_ab_latency",
"8",
"1",
"98304",
"524288",
"400000",
"0",
"0",
"uniform",
"8",
"1300"
]
},
{
"rep": 1,
"arm": "base",
"command": [
"perf",
"record",
"-q",
"-e",
"cycles:u",
"-c",
"100000",
"-o",
"/tmp/mimalloc-585/profiler-cycle-attribution/1-base.data",
"--",
"taskset",
"-c",
"0,1,2,3",
"/tmp/mimalloc-585/balanced-byte-threshold-main/builds/base/pprof/perf_ab_latency",
"8",
"1",
"98304",
"524288",
"400000",
"0",
"0",
"uniform",
"8",
"1300"
]
},
{
"rep": 2,
"arm": "base",
"command": [
"perf",
"record",
"-q",
"-e",
"cycles:u",
"-c",
"100000",
"-o",
"/tmp/mimalloc-585/profiler-cycle-attribution/2-base.data",
"--",
"taskset",
"-c",
"0,1,2,3",
"/tmp/mimalloc-585/balanced-byte-threshold-main/builds/base/pprof/perf_ab_latency",
"8",
"1",
"98304",
"524288",
"400000",
"0",
"0",
"uniform",
"8",
"1300"
]
},
{
"rep": 2,
"arm": "head",
"command": [
"perf",
"record",
"-q",
"-e",
"cycles:u",
"-c",
"100000",
"-o",
"/tmp/mimalloc-585/profiler-cycle-attribution/2-head.data",
"--",
"taskset",
"-c",
"0,1,2,3",
"/tmp/mimalloc-585/balanced-byte-threshold-main/builds/head/pprof/perf_ab_latency",
"8",
"1",
"98304",
"524288",
"400000",
"0",
"0",
"uniform",
"8",
"1300"
]
}
]
}Reproduction script: import hashlib,json,os,subprocess
from pathlib import Path
root=Path('/tmp/mimalloc-585');out=root/'profiler-cycle-attribution';out.mkdir(exist_ok=True)
exes={arm:root/f'balanced-byte-threshold-main/builds/{arm}/pprof/perf_ab_latency' for arm in ('base','head')}
cpus=','.join(map(str,sorted(os.sched_getaffinity(0))[:4]));results=[]
manifest={'cpus':cpus,'environment':{'MIMALLOC_PROF':'1'},'purpose':'Local sampled cycles only; perturbed p99 is not acceptance','hashes':{a:hashlib.sha256(p.read_bytes()).hexdigest() for a,p in exes.items()},'runs':results}
for rep in range(3):
arms=list(exes) if rep%2==0 else list(reversed(exes))
for arm in arms:
data=out/f'{rep}-{arm}.data'
cmd=['perf','record','-q','-e','cycles:u','-c','100000','-o',str(data),'--','taskset','-c',cpus,str(exes[arm]),'8','1','98304','524288','400000','0','0','uniform','8','1300']
env=os.environ.copy();env.update(manifest['environment'])
proc=subprocess.run(cmd,capture_output=True,text=True,env=env,check=True)
(out/f'{rep}-{arm}.stderr').write_text(proc.stderr);(out/f'{rep}-{arm}.stdout').write_text(proc.stdout)
report=subprocess.run(['perf','report','--stdio','--percent-limit','0.1','-i',str(data)],capture_output=True,text=True,check=True)
(out/f'{rep}-{arm}.report').write_text(report.stdout)
results.append({'rep':rep,'arm':arm,'command':cmd});(out/'manifest.json').write_text(json.dumps(manifest,indent=2)+'\n')
print(rep,arm,'saved',flush=True)0-base.report — top user-symbol cycle shares0-head.report — top user-symbol cycle shares1-base.report — top user-symbol cycle shares1-head.report — top user-symbol cycle shares2-base.report — top user-symbol cycle shares2-head.report — top user-symbol cycle sharesStatic follow-upActual old/new profiler executables differ. The listed profiler and generic-allocation symbol sizes match across old/new candidates; fresh-page size is also 1,218 bytes for both, versus 1,162 at frozen baseline. Source isolation changes only the byte predicate in C. Equal symbol sizes do not prove equal instruction sequences, call graphs or performance. This inspection provides no basis for dismissing the p99 failure or changing source before CI attribution. {
"symbols": {
"base": {
"_mi_prof_stack_capture": {
"address": "0000000000039990",
"size": 89
},
"_mi_prof_on_alloc": {
"address": "000000000002f700",
"size": 855
},
"mi_page_queue_find_free_ex": {
"address": "00000000000230c0",
"size": 959
},
"_mi_malloc_generic": {
"address": "0000000000023e50",
"size": 984
},
"_mi_prof_stack_intern": {
"address": "00000000000399f0",
"size": 984
},
"run_ops": {
"address": "00000000000050d0",
"size": 1051
},
"mi_page_fresh_alloc": {
"address": "0000000000022050",
"size": 1162
}
},
"old": {
"_mi_prof_stack_capture": {
"address": "00000000000399d0",
"size": 89
},
"_mi_prof_on_alloc": {
"address": "000000000002f740",
"size": 855
},
"mi_page_queue_find_free_ex": {
"address": "0000000000023100",
"size": 959
},
"_mi_malloc_generic": {
"address": "0000000000023e90",
"size": 984
},
"_mi_prof_stack_intern": {
"address": "0000000000039a30",
"size": 984
},
"run_ops": {
"address": "00000000000050d0",
"size": 1051
},
"mi_page_fresh_alloc": {
"address": "0000000000022050",
"size": 1218
}
},
"head": {
"_mi_prof_stack_capture": {
"address": "00000000000399d0",
"size": 89
},
"_mi_prof_on_alloc": {
"address": "000000000002f740",
"size": 855
},
"mi_page_queue_find_free_ex": {
"address": "0000000000023100",
"size": 959
},
"_mi_malloc_generic": {
"address": "0000000000023e90",
"size": 984
},
"_mi_prof_stack_intern": {
"address": "0000000000039a30",
"size": 984
},
"run_ops": {
"address": "00000000000050d0",
"size": 1051
},
"mi_page_fresh_alloc": {
"address": "0000000000022050",
"size": 1218
}
}
},
"actual_executable_sha256": {
"old": {
"perf_ab": "9249fe01572f8989a024d0f6ee72b6ded65ea6feff2343e2274ae800b23a5cc3",
"perf_ab_latency": "43e680d121436d9e4d8f61d66e11a0d162c166a8d7bccaa69209c264355f2022"
},
"head": {
"perf_ab": "ef1f0135f09bec19ff3cba3c26a211359243cbcb40825818f9bddb3c512771ba",
"perf_ab_latency": "1f432e1f3a4c857bde03dd44859eac39b92a510806cceb4841087d16e98d7a8e"
}
}
} |
Exact random-large/8 follow-up: still INCONCLUSIVERun 39526533882, frozen base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All four actual timing/latency executable hashes verify and are byte-identical to the original new-candidate random batch. Exact row parameters and balanced arm/sampling protocols match. The earlier Intel 8573C run remains INCONCLUSIVE on faults/peak; it is not pooled with or reclassified by this EPYC result. Eight workers on four logical CPUs is 2:1 oversubscription. No trace, threshold or sampling change. No further acceptance rerun of this already rejected candidate is being launched while the real profiler p99 failure is under source attribution. This control remains unresolved; a successful workflow conclusion is not a performance PASS. PR #609 remains draft and the failed profiler comparison remains in the record. {
"base/plain/perf_ab": "7f0c70ce2450dfa09126c1674bddab53187614f12d5bd648b4fda33ccbd3c712",
"base/plain/perf_ab_latency": "ef2d5d273681a473e253d1dd2080307f1ae9f09d34977bda2ad4cab9cbfe7823",
"head/plain/perf_ab": "9ee3617e371e93195de1374ef403455e7ae660848d689ce4bf0ae0ce2f8565b1",
"head/plain/perf_ab_latency": "c14766d194fd7f3c07652b148d1de18e3e4db90b4414bd563398f5f44ee9b982"
}Full-precision cells{
"actual_counts": {
"base": 120,
"head": 120,
"null": 120
},
"saved_percent": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 1035977.7,
"head_median": 1039670.1,
"change": [
0.07535061440704756,
-0.7616121749211847,
0.9318871090325886
],
"regression": [
-0.9318871090325886,
0.7616121749211847
],
"allowance": 0.7041781786584234,
"verdict": "INCONCLUSIVE"
},
{
"metric": "cpu s",
"base_median": 1.2324000000000002,
"head_median": 1.2218,
"change": [
-0.11101859806526174,
-1.1700273836196167,
0.3362001563721808
],
"regression": [
-1.1700273836196167,
0.3362001563721808
],
"allowance": 1.2820512820512813,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 263.0,
"head_median": 262.0,
"change": [
-0.18587360594795538,
-1.1111111111111112,
0.3861003861003861
],
"regression": [
-1.1111111111111112,
0.3861003861003861
],
"allowance": 0.3861003861003861,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 14848.0,
"head_median": 14917.5,
"change": [
-0.10560582149508456,
-1.8236444941307994,
1.8134527852033813
],
"regression": [
-1.8236444941307994,
1.8134527852033813
],
"allowance": 2.4077350859453994,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 182.380859375,
"head_median": 180.45703125,
"change": [
-0.0032476670924719074,
-1.1469596813028062,
1.0516976477462077
],
"regression": [
-1.1469596813028062,
1.0516976477462077
],
"allowance": 1.1141827929965653,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 20.578125,
"head_median": 20.63671875,
"change": [
0.018982536066818528,
0.0,
0.28468400075915734
],
"regression": [
0.0,
0.28468400075915734
],
"allowance": 0.018986140117714066,
"verdict": "INCONCLUSIVE"
}
]
} |
Three-control replacement: CI timeout, no performance verdictRun 39521491231 is authoritatively completed / cancelled. Its check annotation states: "The job has exceeded the maximum execution time of 1h30m0s". Job started 2026-10-21 00:28:09 UTC, completed 01:58:32 UTC. It used frozen baseline Both always-run artifact steps executed, but their logs say no files were found for This exposes a harness gap: raw data is written only at completion/sanity failure and exact builds are copied in The same three-row group will not be restarted. Future control measurements will use exact individual rows after evidence preservation is fixed and the candidate's separate profiler p99 failure is addressed. That failure remains visible in the current candidate ledger; PR #609 remains draft and unaccepted. |
Local-Gate: v1 tree=512cd0a71b7c127b40b0e2ffd9970635c4e40a68 secs=37 lanes=lint:run
Ci-Attestation: {"at":1792548431,"gate":"python/all/lint","host":"linux-x86_64","key":"a41da66c1fe96de5db75bca08c56bace07da7be78b8f505267c7d76be42fb4c5","lane":"lint","parents":["15bfcba0c3e59aea6b22dd04f91db9d95cd63a0a"],"secs":37,"stamp":"bbb8e91cd2ac15392b4353e4a67ec167","tree":"512cd0a71b7c127b40b0e2ffd9970635c4e40a68","v":1,"via":"run"}
Profiler source isolation: byte-threshold step remains INCONCLUSIVERun 39526717341 compares old prefix-reuse candidate base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All four actual timing/latency binary hashes verify. The attribution baseline matches the preserved old-candidate profiler binary byte for byte; the head matches the actual failed new-candidate profiler binary. Row parameters, balanced arm positions and sampling definition are unchanged. Eight workers on four logical CPUs is 2:1 oversubscription. This does not establish the cause of the full candidate p99 regression and is not acceptance against frozen {
"base/pprof/perf_ab": "9249fe01572f8989a024d0f6ee72b6ded65ea6feff2343e2274ae800b23a5cc3",
"base/pprof/perf_ab_latency": "43e680d121436d9e4d8f61d66e11a0d162c166a8d7bccaa69209c264355f2022",
"head/pprof/perf_ab": "ef1f0135f09bec19ff3cba3c26a211359243cbcb40825818f9bddb3c512771ba",
"head/pprof/perf_ab_latency": "1f432e1f3a4c857bde03dd44859eac39b92a510806cceb4841087d16e98d7a8e"
}Full-precision cells{
"actual_counts": {
"base": 120,
"head": 120,
"null": 120
},
"saved_percent": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 4197780.0,
"head_median": 4204376.449999999,
"change": [
0.16305753947899565,
-0.06370095224841385,
0.40023848597696116
],
"regression": [
-0.40023848597696116,
0.06370095224841385
],
"allowance": 0.5953655172004731,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 2.9408,
"head_median": 2.9341,
"change": [
-0.0878065580051814,
-0.3093788882536865,
0.22150281138183509
],
"regression": [
-0.3093788882536865,
0.22150281138183509
],
"allowance": 0.4663829787234058,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 1626.0,
"head_median": 1625.0,
"change": [
0.061595324598065375,
-0.36968576709796674,
0.30959752321981426
],
"regression": [
-0.36968576709796674,
0.30959752321981426
],
"allowance": 0.3708281829419036,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 38362.0,
"head_median": 39779.0,
"change": [
0.956764272117955,
-2.0434667373442883,
6.3519620811287485
],
"regression": [
-2.0434667373442883,
6.3519620811287485
],
"allowance": 6.073896863370547,
"verdict": "INCONCLUSIVE"
},
{
"metric": "peak RSS MiB",
"base_median": 73.64453125,
"head_median": 73.603515625,
"change": [
-0.06267554273962415,
-0.31027162375483097,
0.06187799721549012
],
"regression": [
-0.31027162375483097,
0.06187799721549012
],
"allowance": 0.23923444976076555,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 54.341796875,
"head_median": 54.75390625,
"change": [
0.18931634779456225,
-0.01415227851684121,
1.5011132658191482
],
"regression": [
-0.01415227851684121,
1.5011132658191482
],
"allowance": 0.3385765676800451,
"verdict": "INCONCLUSIVE"
}
]
} |
Forced-termination evidence preservation fixed in draft PR #609Pushed harness commit
Allocator acceptance remains failed on profiler p99. The byte-threshold source step remains INCONCLUSIVE at 120/arm on EPYC 7763 and does not explain or waive the full-candidate EPYC 9V45 failure. The two earlier source steps are now isolated with the fixed harness, exact original profiler row, 60-to-120 samples/arm and unchanged null/sanity gates:
These are source attribution, not retries of the failed fixed-baseline candidate for acceptance. No cross-host pooling or addition of per-step deltas is planned. The old three-control group will not be restarted; its controls remain unverified until future exact individual-row measurements after the profiler failure is addressed. Random/8 remains INCONCLUSIVE in both independent runs. The existing short-generation follow-up 39526576845 remains live. No memory policy was changed. |
Local-Gate: v1 tree=05076c3b29c48822ea0e6acd6f881407cbe6276d secs=32 lanes=lint:run
Ci-Attestation: {"at":1792549262,"gate":"python/all/lint","host":"linux-x86_64","key":"63d62f4417f1c2377d1ff20ec0a0df7c310ec4da1e1527e337923837406dc40e","lane":"lint","parents":["0331d770f2db9d5815682205fc16ca3173d79b39"],"secs":32,"stamp":"fcf15aa913aa910ac08f28673454ecc5","tree":"05076c3b29c48822ea0e6acd6f881407cbe6276d","v":1,"via":"run"}
New word-hash CPU prototype: correctness checked, performance pendingNew candidate The recorded hypothesis comes from post-processing six existing local cycle recordings, not new local timing: 39.1–55.7% of stack-intern samples lie in the byte-FNV loop under The prototype uses mimalloc's existing stateless Verification so far:
First performance comparisons use the exact original profiler row, 60-to-120 samples per arm, balanced order, null/sanity enabled and unchanged thresholds:
No PASS is claimed yet. The first comparison must pass before broadening performance validation. Source-step results on different hosts will not be pooled or added. Older exact-fit/prefix profiler comparisons 39531382092/39531384408 continue on their original refs; the old chart-short follow-up 39526576845 has now finished INCONCLUSIVE on faults at 240/arm (other five metrics PASS). This is evidence for old 15, not new-hash acceptance. The lost-controls timeout and both random/8 INCONCLUSIVE results remain recorded. No new memory policy is accepted, and #585 remains incomplete. |
Previous-candidate chart-short follow-up: faults still INCONCLUSIVE at maximum budgetRun 39526576845, frozen base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All four actual timing/latency executable hashes verify and are byte-identical to the original chart batch. Row parameters, sampling and arm-order protocol match. The earlier EPYC 7763 short-generation result at 120/arm remains INCONCLUSIVE on faults/peak/drain; different hosts/models are not pooled or reclassified. Eight workers on four logical CPUs is 2:1 oversubscription. This is evidence for old 15bfcba, not the newly pushed word-hash prototype 6f7be38. No further acceptance retry of old 15 is planned: its profiler p99 FAIL and unresolved controls remain visible. New prototype acceptance begins with exact profiler run39532706232; no PASS is yet claimed for it. No trace, sampling, threshold or memory-policy change. {
"base/chart/perf_ab": "306d5c8a089d29a7ff276df621226fe2d08ffe7013669627f38dc5c0020d33f7",
"base/chart/perf_ab_latency": "2f883b21f4d48e9e1b404c3110ca058902c68cba23e99a82edd58026c7495090",
"head/chart/perf_ab": "51b9dde3f2ce5e446eb8443af7ab56136af1b146bbdb61acad221d2a652a0bff",
"head/chart/perf_ab_latency": "41d91d58314a4ca202ec6b0e4a8a3d62995a39a3aac891edab560f22c0500240"
}Full-precision cells{
"actual_counts": {
"base": 240,
"head": 240,
"null": 240
},
"saved_percent": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 6833744.35,
"head_median": 6919946.55,
"change": [
1.035337311654582,
0.6967528846682861,
1.443178132619091
],
"regression": [
-1.443178132619091,
-0.6967528846682861
],
"allowance": 0.49396029343884473,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 0.4704,
"head_median": 0.46665,
"change": [
-0.8615666670377912,
-1.1967247533067358,
-0.42698548249359564
],
"regression": [
-1.1967247533067358,
-0.42698548249359564
],
"allowance": 0.7380073800738038,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 251.0,
"head_median": 252.0,
"change": [
0.0,
-0.39370078740157477,
0.7905138339920948
],
"regression": [
-0.39370078740157477,
0.7905138339920948
],
"allowance": 0.4016064257028112,
"verdict": "INCONCLUSIVE"
},
{
"metric": "p99 operation ns",
"base_median": 1943.0,
"head_median": 1933.0,
"change": [
-0.5146816721671118,
-1.0240655401945724,
0.0
],
"regression": [
-1.0240655401945724,
0.0
],
"allowance": 0.5282620179609087,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 72.73046875,
"head_median": 74.7265625,
"change": [
0.0,
0.0,
0.005371146202599634
],
"regression": [
0.0,
0.005371146202599634
],
"allowance": 2.5404386226059343,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 72.73046875,
"head_median": 74.7265625,
"change": [
0.0,
0.0,
0.005371146202599634
],
"regression": [
0.0,
0.005371146202599634
],
"allowance": 2.5404386226059343,
"verdict": "PASS"
}
]
} |
New-head correctness gate failure: musl dead-thread page reuseHead This is distinct from the earlier GCC ASan idle-release failure. In this run the idle/reserved/slack checks all passed at222/205/224ms. No retry, timeout relaxation, assertion weakening or RUN_SERIAL change is being made. Cause is not yet established; profiling/hash code being the latest change does not prove causation. The existing four independent performance comparisons continue unchanged. Next diagnostic: use the actual downloaded musl CI test executable in Alpine3.20, matching the bundle environment, on a declared maximum10-run local plan per source and stop on the first failure. Compare prior15bfcba0's actual CI binary, then frozen baseline source if needed. Local probes are reproduction/attribution only, never gate acceptance. Actual failed bundle executable SHA256: {
"name": "test-purge-holes",
"argv": [
"${BUNDLE}/mimalloc-test-purge-holes"
],
"env": {},
"cwd": "${BUNDLE}",
"timeout": 600.0,
"expect_nonzero": false,
"expect_text": null,
"forbid_text": null,
"labels": [],
"serial": false
}Full failed testcase output retained from JUnit |
#585: new word-hash fixed-baseline profiler row PASS; correctness remains unresolvedRun 39532706232 completed. Exact source pair: base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All four actual timing/latency executables were downloaded and SHA256-verified against the artifact manifest: {
"base/pprof/perf_ab": "c6784c2c2cb2619d31ce4bc679f840ad7d9a2cf04aba6ce4dda8c374d5c8ba37",
"base/pprof/perf_ab_latency": "f4e223e4c1489619ef2cd9831bb88a204931fa207d7ec18a17411e2899eef12f",
"head/pprof/perf_ab": "4dda45628f9468f02b2f3696824e8f9f02bfc9f1407ef242586af6ff8ea88dc7",
"head/pprof/perf_ab_latency": "4ca3b9a3cff79cfee4e24f5ed292ae8ecd2357166339ec62d701a374c9730c31"
}Zero memory credit. CPU and p99 intervals contain zero: this is a nonregression PASS for this row, not proof of a CPU or RSS improvement. Previous Full-precision ledger cells: {
"workload": "large-class/8 (profiler on)",
"counts": {
"base": 60,
"head": 60,
"null": 60
},
"saved": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 3456017.8,
"head_median": 3472172.75,
"change": [
1.0977398707608812,
-1.0723401554954406,
2.4589613632465848
],
"regression": [
-2.4589613632465848,
1.0723401554954406
],
"allowance": 2.4038033110313246,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 3.60135,
"head_median": 3.5604,
"change": [
-0.9128394244298823,
-2.7399764920987386,
1.3912565301072366
],
"regression": [
-2.7399764920987386,
1.3912565301072366
],
"allowance": 2.736381241054142,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 1630.5,
"head_median": 1632.5,
"change": [
0.12302036306259127,
-0.43023970497848807,
0.4904966278356836
],
"regression": [
-0.43023970497848807,
0.4904966278356836
],
"allowance": 0.8064516129032258,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 39587.5,
"head_median": 38576.5,
"change": [
-1.3605905491732997,
-3.7010567155035994,
0.6678667397145913
],
"regression": [
-3.7010567155035994,
0.6678667397145913
],
"allowance": 2.6945892072336033,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 71.103515625,
"head_median": 71.4921875,
"change": [
-0.4040898292187779,
-2.548122000992775,
1.3524703284570798
],
"regression": [
-2.548122000992775,
1.3524703284570798
],
"allowance": 1.5512006624344465,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 68.9453125,
"head_median": 68.685546875,
"change": [
-0.2635640745575398,
-1.3862386350881133,
0.6390318384889442
],
"regression": [
-1.3862386350881133,
0.6390318384889442
],
"allowance": 0.8086405790545126,
"verdict": "PASS"
}
]
} |
Local-Gate: v1 tree=a6aa7b0bbd4650301c3f628fa62d8a4c0083dfc1 secs=35 lanes=lint:run
Ci-Attestation: {"at":1792551246,"gate":"python/all/lint","host":"linux-x86_64","key":"0a9efb8ac5d221f37cd2f7026a58bcc3655fc7ea641c9d5b78be4a75369ef3a1","lane":"lint","parents":["6f7be38f43be01f7017a1c7432421e97cd3f30b5"],"secs":35,"stamp":"e282c045455acae91bba9af2ddbdb288","tree":"a6aa7b0bbd4650301c3f628fa62d8a4c0083dfc1","v":1,"via":"run"}
#585: deterministic reserved-page scanner interleaving; serial test fixture separatedThe failed musl Controlled mechanism evidence: scratch copies of frozen baseline EE and candidate 6f both have byte-identical Both EE and 6f, built via A Fixture correction: #493 explicitly requests a deterministic test with serial threads. The existing Verification: native and musl fixture runs each cover five cases: scavenger-enabled invocation must reject (exit 1), serial reuse ON/OFF pass, original background ON/OFF pass. Actual CMake/CTest registration uses the explicit arguments and startup environment; all four CTest entries passed (7.24 s). All compiler invocations, including CMake's compiler, route through soldr. Same sole primary reviewer clean. Required local gate 45/45 PASS in 35 s. Attested test-only C-core commit Reproducer (diagnostic scratch only, not shipped)Hook inserted immediately before the original comment extern void reserve_probe_claimed(mi_page_t* page);
reserve_probe_claimed(page);#include <assert.h>
#include <pthread.h>
#include <stdbool.h>
#include <stdio.h>
#include <string.h>
#include "mimalloc.h"
#include "mimalloc-stats.h"
#include "mimalloc/internal.h"
#define PROBE_BLOCK_BYTES (400 * 1024)
#define PROBE_BLOCK_COUNT 3
#define PROBE_RESERVE_DELAY_MS 1000
static mi_page_t* page_a;
static mi_page_t* page_b;
static size_t stats_bin;
static bool inject;
static bool hook_fired;
static void* thread_a(void* arg) {
(void)arg;
void* blocks[PROBE_BLOCK_COUNT];
for (size_t i=0;i<PROBE_BLOCK_COUNT;i++) {
blocks[i]=mi_malloc(PROBE_BLOCK_BYTES);assert(blocks[i]!=NULL);
memset(blocks[i],1,PROBE_BLOCK_BYTES);
}
page_a=_mi_ptr_page(blocks[PROBE_BLOCK_COUNT-1]);
stats_bin=_mi_page_stats_bin(page_a);
assert(!mi_page_is_singleton(page_a));
for (size_t i=0;i<PROBE_BLOCK_COUNT;i++) mi_free(blocks[i]);
return NULL;
}
static void* thread_b(void* arg) {
(void)arg;
void* p=mi_malloc(PROBE_BLOCK_BYTES);assert(p!=NULL);
page_b=_mi_ptr_page(p);memset(p,2,PROBE_BLOCK_BYTES);mi_free(p);
return NULL;
}
static void run_thread(void*(*fn)(void*)) {
pthread_t t;assert(pthread_create(&t,NULL,fn,NULL)==0);
assert(pthread_join(t,NULL)==0);
}
void reserve_probe_claimed(mi_page_t* page) {
if (!inject || hook_fired || page!=page_a) return;
assert(mi_page_is_owned(page));
assert(page->retired_at!=0);
assert(mi_page_all_free(page));
hook_fired=true;
run_thread(thread_b);
}
static int64_t created(void) {
mi_stats_t_decl(st);assert(mi_stats_get(&st));
return st.page_bins[stats_bin].total;
}
int main(int argc,char** argv) {
(void)argv;
mi_option_set(mi_option_scavenger,0);
mi_collect(true);
mi_option_set(mi_option_purge_delay,PROBE_RESERVE_DELAY_MS);
mi_option_set(mi_option_page_reserve,1);
inject=(argc>1);
run_thread(thread_a);
const int64_t before=created();
if (inject) {
(void)_mi_arenas_release_reserved(mi_heap_main(),false);
assert(hook_fired);
} else run_thread(thread_b);
const int64_t delta=created()-before;
printf("mode=%s hook=%d same_page=%d new_pages=%lld\n",inject?"scanner-owns-page":"control",hook_fired,page_a==page_b,(long long)delta);
const bool observed=inject?(page_a!=page_b && delta==1):(page_a==page_b && delta==0);
mi_collect(true);
return observed?0:1;
}Actual fixture results: [
{
"name": "enabled-reject",
"env": {
"MIMALLOC_SCAVENGER": "1"
},
"args": [
"--serial-reserve-reuse"
],
"actual": 1,
"expected": 1
},
{
"name": "reuse-on",
"env": {
"MIMALLOC_SCAVENGER": "0",
"MIMALLOC_PURGE_HOLES": "1"
},
"args": [
"--serial-reserve-reuse"
],
"actual": 0,
"expected": 0
},
{
"name": "reuse-off",
"env": {
"MIMALLOC_SCAVENGER": "0",
"MIMALLOC_PURGE_HOLES": "0"
},
"args": [
"--serial-reserve-reuse"
],
"actual": 0,
"expected": 0
},
{
"name": "background",
"env": {},
"args": [],
"actual": 0,
"expected": 0
},
{
"name": "background-holes-off",
"env": {
"MIMALLOC_PURGE_HOLES": "0"
},
"args": [],
"actual": 0,
"expected": 0
}
]CTest: SHA256 evidence manifest: {
"probe.c": "e4ed4a5f2249f4ec0cc13d741c5f666f4d4c0cc653366820023b8eb2e3d958d0",
"baseline-probe": "cecf91e9315efba1cbb42840aff44ea2ce9dff55097bcb858d1d2e9eb0e8e303",
"candidate-probe": "c8f2914534111545f1f5f4823a8db15216a2b4d2e65486388dffa134340a9561",
"baseline-musl-probe": "e90f7a1a1a38b1798ca66d40b60be527f13cd802dde08d7c64cf3f8192df15d8",
"purge-holes-fixture": "34113d7473acd3e0c8e66a56a38356c56a0e76b0fec632f400ac9751395d873f",
"purge-holes-fixture-musl": "8b0f5e3523827b586dfc88a58bd278887af4d6b9cf4ab35c8ddc45bf39a2e7dd",
"fixture-results.json": "9c1ed4b68275175a38635d86e63e43d545183158590f8018f25b66720fa2786c",
"ctest.log": "99148153a95fcc66f21bed28d02315f1106f427dcd21c14ff7df075dfc4b882c",
"musl-fixture.log": "25f7a5fd694b5a41dcb4edd2aee29c0455953da09379f32a4f047f8bf5b4be23"
} |
#585: earlier profiler source comparisons completed (attribution only)These isolate earlier source steps; neither is acceptance of the current full candidate. Original exact profiler trace, null/sanity gates, balanced arm ordering, unchanged 60-to-120 plan. The display header's 60 is the requested starting count; actual counts below come from raw sample-array lengths. Different hosts/source pairs are not pooled and their deltas are not added. Prior full-pair p99 failures are not waived or reclassified. Frozen EE → exact-fit-only a9d7d2eRun 39531382092, AMD EPYC 7763 64-Core Processor. Actual samples per arm: {'base': 95, 'head': 95, 'null': 95}. All four actual timing/latency executable hashes verified. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
Verified executable manifest: {
"base/pprof/perf_ab": "c6784c2c2cb2619d31ce4bc679f840ad7d9a2cf04aba6ce4dda8c374d5c8ba37",
"base/pprof/perf_ab_latency": "f4e223e4c1489619ef2cd9831bb88a204931fa207d7ec18a17411e2899eef12f",
"head/pprof/perf_ab": "63884c20d9d2a222c1983534b9bc18d1d0d22ec844c843aac44fe49009564ca9",
"head/pprof/perf_ab_latency": "da62ab296e1914727452b2e820282472f716f22c69a61635791b04c3e46c8fd5"
}Full-precision cells: [
{
"workload": "large-class/8 (profiler on)",
"counts": {
"base": 95,
"head": 95,
"null": 95
},
"saved": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 4202652.8,
"head_median": 4220480.1,
"change": [
0.4905421269055018,
0.14327076722488194,
0.708496763044393
],
"regression": [
-0.708496763044393,
-0.14327076722488194
],
"allowance": 0.36209909439788956,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 2.9412,
"head_median": 2.9252,
"change": [
-0.6035251354502379,
-0.7802428930049573,
-0.3786457444994143
],
"regression": [
-0.7802428930049573,
-0.3786457444994143
],
"allowance": 0.3764370739647942,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 1625.0,
"head_median": 1624.0,
"change": [
-0.06172839506172839,
-0.3703703703703704,
0.24645717806531117
],
"regression": [
-0.3703703703703704,
0.24645717806531117
],
"allowance": 0.3094059405940594,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 38031.0,
"head_median": 39160.0,
"change": [
0.8800679776644816,
-2.1027695012943877,
4.457564988667208
],
"regression": [
-2.1027695012943877,
4.457564988667208
],
"allowance": 5.077547572678852,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 75.48046875,
"head_median": 73.66015625,
"change": [
-0.02652942112803099,
-2.387789028260541,
0.1750941794450045
],
"regression": [
-2.387789028260541,
0.1750941794450045
],
"allowance": 0.3979413169204648,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 54.41796875,
"head_median": 54.37109375,
"change": [
-0.4606636435615058,
-1.6683161317686979,
0.19435646415202995
],
"regression": [
-1.6683161317686979,
0.19435646415202995
],
"allowance": 0.4507994646756357,
"verdict": "PASS"
}
]
}
]Exact-fit-only a9d7d2e → prefix-reuse 9b67e41Run 39531384408, AMD EPYC 9V45 96-Core Processor. Actual samples per arm: {'base': 120, 'head': 120, 'null': 120}. All four actual timing/latency executable hashes verified. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
Verified executable manifest: {
"base/pprof/perf_ab": "63884c20d9d2a222c1983534b9bc18d1d0d22ec844c843aac44fe49009564ca9",
"base/pprof/perf_ab_latency": "da62ab296e1914727452b2e820282472f716f22c69a61635791b04c3e46c8fd5",
"head/pprof/perf_ab": "9249fe01572f8989a024d0f6ee72b6ded65ea6feff2343e2274ae800b23a5cc3",
"head/pprof/perf_ab_latency": "43e680d121436d9e4d8f61d66e11a0d162c166a8d7bccaa69209c264355f2022"
}Full-precision cells: [
{
"workload": "large-class/8 (profiler on)",
"counts": {
"base": 120,
"head": 120,
"null": 120
},
"saved": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 4851513.949999999,
"head_median": 4863665.4,
"change": [
0.35597414933043087,
-0.15125286464079588,
0.6580069460234763
],
"regression": [
-0.6580069460234763,
0.15125286464079588
],
"allowance": 0.7196587330291655,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 2.5869,
"head_median": 2.57885,
"change": [
-0.30566662877407647,
-0.631283610007014,
0.05429302722408633
],
"regression": [
-0.631283610007014,
0.05429302722408633
],
"allowance": 0.45851441330089876,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 1630.5,
"head_median": 1634.0,
"change": [
0.24517501874690995,
0.0,
0.5528255528255528
],
"regression": [
0.0,
0.5528255528255528
],
"allowance": 0.4914004914004914,
"verdict": "INCONCLUSIVE"
},
{
"metric": "p99 operation ns",
"base_median": 24206.0,
"head_median": 23535.5,
"change": [
-2.2627565700302705,
-8.938805338666873,
2.8635559787783973
],
"regression": [
-8.938805338666873,
2.8635559787783973
],
"allowance": 8.434593145059917,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 73.9453125,
"head_median": 75.625,
"change": [
0.15954175054025288,
-0.06872488898287164,
0.7231176860234046
],
"regression": [
-0.06872488898287164,
0.7231176860234046
],
"allowance": 0.7875316664830928,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 72.041015625,
"head_median": 71.94140625,
"change": [
-0.4801196767773953,
-1.0019932123040456,
-0.027114967462039046
],
"regression": [
-1.0019932123040456,
-0.027114967462039046
],
"allowance": 0.4405286343612335,
"verdict": "PASS"
}
]
}
]The exact-fit profiler step PASSes all six metrics at 95/arm; prefix-reuse remains INCONCLUSIVE on minor faults at 120/arm, with the other five PASS. Both have zero memory credit. Neither establishes the cause of the old full-candidate p99 failures. The new word-hash EE→6f profiler row separately PASSed at 60/arm; fresh full platform CI on test-only da1fdb9 is pending. |
#585: capacity-based fit does not rescue tight geometry; rejected before RSS/timingNew diagnostic idea: previous tight geometry kept whole-block capacity but failed the existing large-bin fungibility test (four distinct pages where at most three are allowed). Instead of requiring the complete nominal span in the retired-page search, require its whole-block capacity: This is distinct from repeating the rejected geometry unchanged, but the structural result is still negative. Scratch-only code based on allocator candidate 6f7be38, compiled via soldr with debug invariants, diagnostics and statistics enabled. Both the tight-only control and tight+capacity-fit variant fail the unchanged rotation check: four distinct pages for 160 requests over ten bins (limit three). The capacity-preservation checks pass, but fungibility does not. Therefore rejected, with no RSS screen or timed CI dispatch, and no production source or option changed. There is no claimed memory saving or performance acceptance. Actual executions (all outcomes retained): [
{
"variant": "tight-only",
"returncode": 1,
"binary_sha256": "f7d311999c3a04ef19d5de9c0977f1b5456797f3d534a8e4e012346aef59ba12",
"stdout": "",
"stderr": "(a) largest span of a busy 200 KiB bin: 15 slices (cap 16, full 64)\n(b) 4 distinct pages for 160 allocations rotating over 10 bins\nFAILED: (b) rotating over 10 bins with one live block kept 4 pages (want at most 3)\n1 check(s) failed\n"
},
{
"variant": "tight-capacity-fit",
"returncode": 1,
"binary_sha256": "1c46c015a0d21a29e340300d2bae6160182df2d3ab53b9acf7f846311a0869e3",
"stdout": "",
"stderr": "(a) largest span of a busy 200 KiB bin: 15 slices (cap 16, full 64)\n(b) 4 distinct pages for 160 allocations rotating over 10 bins\nFAILED: (b) rotating over 10 bins with one live block kept 4 pages (want at most 3)\n1 check(s) failed\n"
}
]Diagnostic-only patch (not in the PR): --- a/include/mimalloc/types.h
+++ b/include/mimalloc/types.h
@@ -386,6 +386,14 @@
// (the #532 behaviour); `-DMI_LARGE_SPAN_MAX_KIB=0` builds that in.
#ifndef MI_LARGE_SPAN_MAX_KIB
#define MI_LARGE_SPAN_MAX_KIB (1024)
+#endif
+// #585 experiment: return whole slices that cannot hold another block without
+// reducing block capacity. Disabled until paired CPU/fault/RSS/latency checks pass.
+#ifndef MI_LARGE_SPAN_TIGHT_GEOMETRY
+#define MI_LARGE_SPAN_TIGHT_GEOMETRY (0)
+#endif
+#ifndef MI_LARGE_REPURPOSE_CAPACITY_FIT
+#define MI_LARGE_REPURPOSE_CAPACITY_FIT (0)
#endif
// each demand step multiplies the span by 2^MI_LARGE_SPAN_GROW_SHIFT (1: 1 -> 2 -> 4 MiB)
#ifndef MI_LARGE_SPAN_GROW_SHIFT
--- a/src/large-span.c
+++ b/src/large-span.c
@@ -139,6 +139,18 @@
const size_t min_slices = mi_slice_count_of_size(MI_LARGE_SPAN_MIN_BLOCKS * block_size + overhead);
if (slices < min_slices) { slices = min_slices; }
if (slices > full) { slices = full; }
+ #if MI_LARGE_SPAN_TIGHT_GEOMETRY
+ // A capped span need not retain whole slices past its last possible block.
+ // Keep the same block capacity and demand level; no resident byte is discarded.
+ // The uncapped opt-out retains its original exact power-of-two geometry.
+ if (cap < full) {
+ const size_t bytes = mi_size_of_slices(slices);
+ const size_t blocks = (bytes > overhead ? (bytes - overhead) / block_size : 0);
+ if (blocks >= MI_LARGE_SPAN_MIN_BLOCKS) {
+ slices = mi_slice_count_of_size(blocks * block_size + overhead);
+ }
+ }
+ #endif
return slices;
}
--- a/src/page.c
+++ b/src/page.c
@@ -587,6 +587,10 @@
MI_EVENT(MI_EVENT_LARGE_REPURPOSE_NONE);
return NULL;
}
+ #if MI_LARGE_REPURPOSE_CAPACITY_FIT
+ // A nominal span promises whole blocks, not unusable bytes after its last block.
+ want = (want / block_size) * block_size;
+ #endif
if (want < min_bytes) { want = min_bytes; }
mi_page_queue_t* best_pq = NULL;
size_t best_size = SIZE_MAX;
--- a/test/test-large-fungible.c
+++ b/test/test-large-fungible.c
@@ -136,6 +136,31 @@
#endif
}
+#if defined(MI_LARGE_SPAN_TIGHT_GEOMETRY) && MI_LARGE_SPAN_TIGHT_GEOMETRY
+static THREAD_RET tight_main(void* arg) {
+ (void)arg;
+ void* const probe = mi_malloc(16);
+ assert(probe != NULL);
+ mi_theap_t* const theap = _mi_ptr_page(probe)->theap;
+ const size_t nominal = mi_size_of_slices(MI_LARGE_SPAN_COMPACT_SLICES);
+ for (size_t i = 0; i < NBINS; i++) {
+ const size_t bsize = mi_good_size(BIN_SIZES[i] * KIB);
+ if (bsize <= MI_MEDIUM_MAX_OBJ_SIZE || bsize > MI_LARGE_MAX_OBJ_SIZE) continue;
+ const size_t blocks = nominal / bsize;
+ const size_t slices = _mi_large_span_peek_slices(theap, bsize, 0);
+ const size_t expected = mi_slice_count_of_size(blocks * bsize);
+ CHECK(slices == expected, "(c) %zu KiB: %zu slices instead of tightly packed %zu", bsize / KIB, slices, expected);
+ CHECK(mi_size_of_slices(slices) / bsize == blocks, "(c) trimming lost block capacity");
+ }
+ mi_free(probe);
+ return THREAD_OK;
+}
+
+static void case_tight(void) {
+ thread_t t; thread_start(&t, &tight_main, NULL); thread_join(t);
+}
+#endif
+
int main(void) {
#if !MI_ENABLE_LARGE_PAGES
fprintf(stderr, "skipped: no large pages in this build\n");
@@ -154,6 +179,9 @@
mi_option_set(mi_option_page_reserve, 0);
case_cap();
case_rotate();
+ #if defined(MI_LARGE_SPAN_TIGHT_GEOMETRY) && MI_LARGE_SPAN_TIGHT_GEOMETRY
+ case_tight();
+ #endif
#endif
if (failures > 0) { fprintf(stderr, "%d check(s) failed\n", failures); return 1; }
fprintf(stderr, "ok\n"); |
#585: current plain executables are byte-identical to the earlier measured candidateBuild-only run 39536056572 freshly built frozen EE and current test-only Downloaded actual files: verified all six report SHA256 values (two timing executables, two separate latency executables, two static libraries). Then compared the whole four linked executables byte for byte with the actual timed EE/15 binaries from run39520753532, not merely symbol addresses or text sizes. All four are identical. The static libraries are verified against this report; no assertion of old-library identity is made. This carries forward measured evidence for those identical GCC13 plain executables; it is not a new measurement on da1fdb9, does not relabel original JSON/source refs, and does not pool different hosts. The original counts, CPU models, traces, null/sanity checks, protocol and verdicts remain attached to their original runs. There are six distinct already-PASS plain cells: main, plain ephemeral, sparse, bursty, Larson/1 and random/1. The earlier seventh ordinary PASS was standard chart ephemeral, whose changed executable must be measured again. The default-aging trial duplicates main, not a seventh distinct cell. Plain small/8 and Larson/8 remain unverified after their original job timeout; random/8 remains INCONCLUSIVE. Binary identity does not resolve those outcomes. Chart enables PPROF, so the word-hash change can affect its linked executable even with profiling unarmed. All three chart cells (Larson/8, standard ephemeral, short generations) require current-candidate validation. Current fixed-baseline profiler row separately PASSed at 60/arm. Broader dispatch remains gated on fresh platform CI for the serial fixture correction. No memory win or whole-candidate acceptance is claimed. Same sole primary reviewer checked this evidence scope; its PASS-cell-count correction is reflected above. Verified build metadata: {
"complete": false,
"diagnostic": "build-only; no measurements or regression verdict",
"harness_sha": "da1fdb9819fdaa42a57e96d1d6cd741cf2711d03",
"compiler": "cc (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0\nCopyright (C) 2023 Free Software Foundation, Inc.\nThis is free software; see the source for copying conditions. There is NO\nwarranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n\n",
"linker": "GNU ld (GNU Binutils for Ubuntu) 2.42\nCopyright (C) 2024 Free Software Foundation, Inc.\nThis program is free software; you may redistribute it under the terms of\nthe GNU General Public License version 3 or (at your option) a later version.\nThis program has absolutely no warranty.\n",
"cmake": "cmake version 3.31.6\n\nCMake suite maintained and supported by Kitware (kitware.com/cmake).\n",
"runner": {
"ImageOS": "ubuntu24",
"ImageVersion": "20261018.339.1",
"RUNNER_OS": "Linux",
"RUNNER_ARCH": "X64"
},
"flags": [
"-DCMAKE_BUILD_TYPE=Release",
"-DMI_BUILD_SHARED=OFF",
"-DMI_BUILD_OBJECT=OFF",
"-DMI_BUILD_TESTS=OFF",
"-DMI_OVERRIDE=OFF",
"-DMI_PPROF=OFF",
"-DMI_MEMEVT=OFF",
"-DMI_DIAGNOSTICS=OFF",
"-DMI_DHAT=OFF",
"-DMI_OWNER_GATE=OFF"
],
"arms": {
"base": {
"ref": "ee3623f6b7a485754a0a0aa3462d8cba43390e8b",
"sha": "ee3623f6b7a485754a0a0aa3462d8cba43390e8b",
"kind": "plain",
"files": {
"perf_ab": {
"sha256": "7f0c70ce2450dfa09126c1674bddab53187614f12d5bd648b4fda33ccbd3c712",
"bytes": 336296
},
"perf_ab_latency": {
"sha256": "ef2d5d273681a473e253d1dd2080307f1ae9f09d34977bda2ad4cab9cbfe7823",
"bytes": 336776
},
"libmimalloc.a": {
"sha256": "0bb97ca951ed5404f754091fa4ec4907949a1973ea6d2f6d559889c66c028d47",
"bytes": 513378
}
}
},
"head": {
"ref": "da1fdb9819fdaa42a57e96d1d6cd741cf2711d03",
"sha": "da1fdb9819fdaa42a57e96d1d6cd741cf2711d03",
"kind": "plain",
"files": {
"perf_ab": {
"sha256": "9ee3617e371e93195de1374ef403455e7ae660848d689ce4bf0ae0ce2f8565b1",
"bytes": 336296
},
"perf_ab_latency": {
"sha256": "c14766d194fd7f3c07652b148d1de18e3e4db90b4414bd563398f5f44ee9b982",
"bytes": 336776
},
"libmimalloc.a": {
"sha256": "7f5a84746d6ad6520117d81e67f5ae37506d08f51f29ed78b91a27e470f1ddc2",
"bytes": 513418
}
}
}
}
} |
#585: isolated word-hash step improves CPU/p99; drain RSS remains INCONCLUSIVERun 39532708955 compares This isolates the word-hash source step, not frozen-baseline acceptance. It is INCONCLUSIVE overall on after-drain RSS. CPU -1.84% [-2.11%,-1.55%], throughput +1.91% [1.60%,2.33%], p99 -2.33% [-4.57%,-1.04%]; those three intervals exclude zero in the favorable direction on this host. Peak/faults PASS. Drain +5.87% [1.40%,27.29%] against 1.91% null noise remains INCONCLUSIVE, not PASS. Zero memory credit. No retry or relaxed gate is planned merely to obtain a PASS. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All four actual executables verified against the SHA256 manifest. The baseline timing/latency files are byte-identical to the original old15 profiler-p99 FAIL run; head files are byte-identical to the new EE→6f profiler PASS run. Those runs used different CPU models, so they are not pooled and their deltas are not added. The old full-pair p99 FAIL is not erased; this step result does not prove its cause. New fixed-baseline acceptance remains the separate Xeon8573C 60/arm all-six PASS; full candidate acceptance additionally requires pending correctness and remaining cells. Verified manifest: {
"base/pprof/perf_ab": "ef1f0135f09bec19ff3cba3c26a211359243cbcb40825818f9bddb3c512771ba",
"base/pprof/perf_ab_latency": "1f432e1f3a4c857bde03dd44859eac39b92a510806cceb4841087d16e98d7a8e",
"head/pprof/perf_ab": "4dda45628f9468f02b2f3696824e8f9f02bfc9f1407ef242586af6ff8ea88dc7",
"head/pprof/perf_ab_latency": "4ca3b9a3cff79cfee4e24f5ed292ae8ecd2357166339ec62d701a374c9730c31"
}Full-precision cells: [
{
"workload": "large-class/8 (profiler on)",
"counts": {
"base": 120,
"head": 120,
"null": 120
},
"saved": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 4564683.800000001,
"head_median": 4663804.05,
"change": [
1.9106592534741764,
1.5997172951026735,
2.3261056701665552
],
"regression": [
-2.3261056701665552,
-1.5997172951026735
],
"allowance": 0.3972306579034035,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 2.7489,
"head_median": 2.69435,
"change": [
-1.843709296422775,
-2.1118825197193862,
-1.5540988436326064
],
"regression": [
-2.1118825197193862,
-1.5540988436326064
],
"allowance": 0.5431068589720495,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 1624.0,
"head_median": 1625.0,
"change": [
0.03076923076923077,
-0.24706609017912293,
0.12376237623762376
],
"regression": [
-0.24706609017912293,
0.12376237623762376
],
"allowance": 0.5572755417956656,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 38774.0,
"head_median": 37862.5,
"change": [
-2.331617592969682,
-4.571640958929566,
-1.0423082389728526
],
"regression": [
-4.571640958929566,
-1.0423082389728526
],
"allowance": 3.044764977204255,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 73.630859375,
"head_median": 73.603515625,
"change": [
0.06809540390977623,
-0.10833677259595542,
0.3701083111086921
],
"regression": [
-0.10833677259595542,
0.3701083111086921
],
"allowance": 0.4559673400137851,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 63.412109375,
"head_median": 71.775390625,
"change": [
5.866936206493472,
1.4017381553125876,
27.28913571676501
],
"regression": [
1.4017381553125876,
27.28913571676501
],
"allowance": 1.9091229306093693,
"verdict": "INCONCLUSIVE"
}
]
}
] |
#585: six individual remaining measurements dispatched, original numeric plan unchangedFixed baseline
Scheduling adjustment, not a numeric-plan change: the earlier assistant-imposed plan said to wait for every full-CI check before dispatching. Full Linux parallel/serial, musl including both unskipped original same-page assertions, eight memory-gate runs, leak-positive control, diagnostics, native MSVC, both native Macs, Rust, cross and ASan have now passed. The independent Windows bundle runner is still alive, currently in the MSVC debug-full bundle. These Linux measurements can run on separate VMs while Windows finishes. Full candidate acceptance still requires every correctness gate, including that Windows job. No CI job was cancelled or restarted. The old three-control job timed out at 90 minutes without artifacts; that group is not restarted. Its remaining rows now have individual jobs under the checkpoint-capable harness. The six prior distinct plain PASS cells are carried forward only through verified whole-executable identity, retaining original refs, hosts, actual counts, outcomes and scope. Original failed and INCONCLUSIVE runs remain visible; no cross-CPU pooling, JSON relabeling, changed traces, new discard policy, or loosened thresholds. |
#585: completed reuse counts and observed free-capacity time answer different questionsOffline analysis of the existing four local arena claim/free traces (old allocator policy For each observed free-slice interval of duration
Cutoff limit: the table stops at the last recorded arena claim, which is only a provisional diagnostic cutoff. It is NOT an authoritative work-end or peak-RSS marker; allocations can continue using existing pages afterward. Full-trace results are also retained below and include the roughly 2.6-second drain/shutdown window. That full window is 93–97% dominated by censored free-interval time and must not be interpreted as work-phase resident pressure. Inference: roughly 99% of completed reuse intervals being short does not show that all free capacity is hot. A minority of longer intervals and the still-free tail make a substantial contribution to observed virtual free-slice time. This does not prove those bytes remain resident, quantify a memory win, or justify selecting a discard threshold. The claim/free trace has no purge/residency events or authoritative phase markers. Before a new cold-release design, collect those missing facts together; do not retry the rejected hot-body discard or infer that a count percentile supplies the cold resident bucket. No new memory setting or policy is retained. Same sole primary reviewer checked the arithmetic and inference scope clean. Full results, limits and raw trace SHA256 identities: {
"purpose": "Descriptive existing local traces; virtual arena-free slice-time, not residency or acceptance",
"limits": [
"Callbacks instrumented the old8b policy; wall time affected by instrumentation.",
"Arena claim/free only; no purge/residency events and no page-retirement events.",
"Full trace includes drain/shutdown. Last recorded arena allocation is only a second diagnostic cutoff, NOT an authoritative work/peak marker.",
"Intervals still free at cutoff are right-censored; observed time is a lower bound on lifetime."
],
"traces": {
"sparse": {
"sha256": "19bd45eec192010986375b3762ba8e27046a73e4689daa8f65c6e722e713dae7",
"full_trace": {
"valid": true,
"events": 120750,
"window_ms": 2783.352852,
"global_timestamp_inversions": 175,
"completed_reuse_slices": 2249519,
"right_censored_free_slices": 2640,
"completed_within_1ms_fraction": 0.9901925700560875,
"known_free_MiB_at_cutoff": 165.0,
"time_average_known_free_MiB": 133.03359631626307,
"right_censored_share_of_observed_free_slice_time": 0.9699571379033658,
"share_of_free_slice_time_with_age_above_ms": {
"1": 0.9783965215334672,
"10": 0.9684443110796214,
"25": 0.9616512222774563,
"50": 0.9516937710078794,
"100": 0.9328180829413952
},
"known_free_MiB_with_age_at_least_ms_at_cutoff": {
"1": 137.5,
"10": 137.5,
"25": 137.5,
"50": 137.5,
"100": 137.5
},
"free_slice_time_conservation_exact": true
},
"last_recorded_arena_claim_cutoff": {
"valid": true,
"events": 120717,
"window_ms": 182.330442,
"global_timestamp_inversions": 175,
"completed_reuse_slices": 2249519,
"right_censored_free_slices": 2055,
"completed_within_1ms_fraction": 0.9901925700560875,
"known_free_MiB_at_cutoff": 128.4375,
"time_average_known_free_MiB": 69.30854652654492,
"right_censored_share_of_observed_free_slice_time": 0.11971170719839819,
"share_of_free_slice_time_with_age_above_ms": {
"1": 0.36923228362605637,
"10": 0.1413967486349996,
"25": 0.08394538990507293,
"50": 0.05211795702653927,
"100": 0.023897529158928045
},
"known_free_MiB_with_age_at_least_ms_at_cutoff": {
"1": 97.0,
"10": 19.4375,
"25": 15.5,
"50": 5.1875,
"100": 4.5625
},
"free_slice_time_conservation_exact": true
}
},
"sparse-2": {
"sha256": "67267e1bafdf3f8c829e69502e15b5b68cb8fa8aadbf52c2a725da5bb5f855e1",
"full_trace": {
"valid": true,
"events": 120758,
"window_ms": 2829.861223,
"global_timestamp_inversions": 127,
"completed_reuse_slices": 2249524,
"right_censored_free_slices": 2679,
"completed_within_1ms_fraction": 0.9878818807890025,
"known_free_MiB_at_cutoff": 167.4375,
"time_average_known_free_MiB": 134.48932033892532,
"right_censored_share_of_observed_free_slice_time": 0.9673337360956915,
"share_of_free_slice_time_with_age_above_ms": {
"1": 0.9768765425616812,
"10": 0.9656415940471132,
"25": 0.9586010737574893,
"50": 0.9489569464646093,
"100": 0.9305597977462671
},
"known_free_MiB_with_age_at_least_ms_at_cutoff": {
"1": 140.4375,
"10": 139.9375,
"25": 139.9375,
"50": 139.9375,
"100": 139.9375
},
"free_slice_time_conservation_exact": true
},
"last_recorded_arena_claim_cutoff": {
"valid": true,
"events": 120727,
"window_ms": 228.535267,
"global_timestamp_inversions": 127,
"completed_reuse_slices": 2249524,
"right_censored_free_slices": 2229,
"completed_within_1ms_fraction": 0.9878818807890025,
"known_free_MiB_at_cutoff": 139.3125,
"time_average_known_free_MiB": 72.46655080482611,
"right_censored_share_of_observed_free_slice_time": 0.24930835243074984,
"share_of_free_slice_time_with_age_above_ms": {
"1": 0.4705149117375412,
"10": 0.2414720186460886,
"25": 0.1538933950421305,
"50": 0.08156383267192319,
"100": 0.010497782491972196
},
"known_free_MiB_with_age_at_least_ms_at_cutoff": {
"1": 105.25,
"10": 69.6875,
"25": 44.0625,
"50": 36.875,
"100": 11.3125
},
"free_slice_time_conservation_exact": true
}
},
"sparse-3": {
"sha256": "cd075a518ab104daec05cb9526ab427a4419850919b69af2b99e9fcca25a6dac",
"full_trace": {
"valid": true,
"events": 120748,
"window_ms": 2813.780453,
"global_timestamp_inversions": 174,
"completed_reuse_slices": 2249443,
"right_censored_free_slices": 2707,
"completed_within_1ms_fraction": 0.9883348900149949,
"known_free_MiB_at_cutoff": 169.1875,
"time_average_known_free_MiB": 136.6107164314669,
"right_censored_share_of_observed_free_slice_time": 0.9659005265574301,
"share_of_free_slice_time_with_age_above_ms": {
"1": 0.9770499461387435,
"10": 0.9654998327687769,
"25": 0.9578835784971109,
"50": 0.9476435125332179,
"100": 0.9290353914360091
},
"known_free_MiB_with_age_at_least_ms_at_cutoff": {
"1": 142.1875,
"10": 141.6875,
"25": 141.6875,
"50": 141.6875,
"100": 141.6875
},
"free_slice_time_conservation_exact": true
},
"last_recorded_arena_claim_cutoff": {
"valid": true,
"events": 120714,
"window_ms": 212.169042,
"global_timestamp_inversions": 174,
"completed_reuse_slices": 2249443,
"right_censored_free_slices": 2100,
"completed_within_1ms_fraction": 0.9883348900149949,
"known_free_MiB_at_cutoff": 131.25,
"time_average_known_free_MiB": 74.35060220048503,
"right_censored_share_of_observed_free_slice_time": 0.16908587043965168,
"share_of_free_slice_time_with_age_above_ms": {
"1": 0.4425570470824442,
"10": 0.18053715036602946,
"25": 0.08718664262036008,
"50": 0.03304739246110129,
"100": 0.006376495999629009
},
"known_free_MiB_with_age_at_least_ms_at_cutoff": {
"1": 111.5625,
"10": 101.6875,
"25": 18.5,
"50": 18.25,
"100": 2.5
},
"free_slice_time_conservation_exact": true
}
},
"persistent": {
"sha256": "abbda7837a018cae3b229912e083e0bd762598012ccb0783ffd9e2d47b4b4ff6",
"full_trace": {
"valid": true,
"events": 85442,
"window_ms": 2912.37183,
"global_timestamp_inversions": 35,
"completed_reuse_slices": 687745,
"right_censored_free_slices": 1043,
"completed_within_1ms_fraction": 0.9899105046201717,
"known_free_MiB_at_cutoff": 65.1875,
"time_average_known_free_MiB": 19.456344345911525,
"right_censored_share_of_observed_free_slice_time": 0.9348786434633712,
"share_of_free_slice_time_with_age_above_ms": {
"1": 0.9595934305494658,
"10": 0.9409516759899553,
"25": 0.9304576425330084,
"50": 0.9185971665333778,
"100": 0.8998645878380296
},
"known_free_MiB_with_age_at_least_ms_at_cutoff": {
"1": 20.0,
"10": 20.0,
"25": 20.0,
"50": 20.0,
"100": 20.0
},
"free_slice_time_conservation_exact": true
},
"last_recorded_arena_claim_cutoff": {
"valid": true,
"events": 85396,
"window_ms": 310.716953,
"global_timestamp_inversions": 35,
"completed_reuse_slices": 687745,
"right_censored_free_slices": 304,
"completed_within_1ms_fraction": 0.9899105046201717,
"known_free_MiB_at_cutoff": 19.0,
"time_average_known_free_MiB": 14.897483660152911,
"right_censored_share_of_observed_free_slice_time": 0.20282506100520625,
"share_of_free_slice_time_with_age_above_ms": {
"1": 0.5062660823817091,
"10": 0.2858446232932152,
"25": 0.18049351279615292,
"50": 0.08844231107336567,
"100": 0.003910237776696162
},
"known_free_MiB_with_age_at_least_ms_at_cutoff": {
"1": 17.9375,
"10": 13.875,
"25": 11.8125,
"50": 8.9375,
"100": 0.0
},
"free_slice_time_conservation_exact": true
}
}
}
}Reproducer (scratch-only standard-library analysis): """Offline arena-free slice-time; not RSS, retired-page lifetimes or a discard-policy trial."""
import hashlib,json,struct
from pathlib import Path
ROOT=Path('/tmp/mimalloc-585')
SLICE_BYTES=64*1024
MIB=1024*1024
AGE_THRESHOLDS_NS={str(ms):ms*1_000_000 for ms in (1,10,25,50,100)}
NAMES=('sparse','sparse-2','sparse-3','persistent')
def analyze(events,cutoff):
selected=events[:cutoff+1]
start=min(x[1] for x in selected);end=max(x[1] for x in selected)
live=set();free={};completed=[];changes=[];negative=[]
for addr,ns,count,freed in selected:
assert addr%SLICE_BYTES==0
for index in range(count):
key=addr//SLICE_BYTES+index
if freed:
assert key in live,('free without claim',key)
live.remove(key);assert key not in free
free[key]=ns;changes.append((ns,1))
else:
assert key not in live,('duplicate claim',key)
live.add(key)
if key in free:
duration=ns-free.pop(key)
if duration<0:negative.append(duration)
completed.append(duration);changes.append((ns,-1))
if negative:
return {'valid':False,'negative_reuse_intervals':len(negative),'minimum_ns':min(negative)}
censored=[end-ns for ns in free.values()]
completed_area=sum(completed);censored_area=sum(censored);area=completed_area+censored_area
# Independent conservation: integrate the inventory of known free slices over time.
inventory=0;integral=0;previous=start
for ns,delta in sorted(changes):
integral+=inventory*(ns-previous);inventory+=delta;previous=ns
assert inventory>=0
integral+=inventory*(end-previous)
assert integral==area,(integral,area)
assert inventory==len(free)
durations=completed+censored
return {'valid':True,'events':len(selected),'window_ms':(end-start)/1e6,
'global_timestamp_inversions':sum(a[1]>b[1] for a,b in zip(selected,selected[1:])),
'completed_reuse_slices':len(completed),'right_censored_free_slices':len(censored),
'completed_within_1ms_fraction':sum(d<1_000_000 for d in completed)/len(completed),
'known_free_MiB_at_cutoff':len(free)*SLICE_BYTES/MIB,
'time_average_known_free_MiB':area/(end-start)*SLICE_BYTES/MIB,
'right_censored_share_of_observed_free_slice_time':censored_area/area,
'share_of_free_slice_time_with_age_above_ms':{label:sum(max(0,d-age) for d in durations)/area for label,age in AGE_THRESHOLDS_NS.items()},
'known_free_MiB_with_age_at_least_ms_at_cutoff':{label:sum(d>=age for d in censored)*SLICE_BYTES/MIB for label,age in AGE_THRESHOLDS_NS.items()},
'free_slice_time_conservation_exact':True}
out={'purpose':'Descriptive existing local traces; virtual arena-free slice-time, not residency or acceptance',
'limits':['Callbacks instrumented the old8b policy; wall time affected by instrumentation.',
'Arena claim/free only; no purge/residency events and no page-retirement events.',
'Full trace includes drain/shutdown. Last recorded arena allocation is only a second diagnostic cutoff, NOT an authoritative work/peak marker.',
'Intervals still free at cutoff are right-censored; observed time is a lower bound on lifetime.'],
'traces':{}}
old=json.loads((ROOT/'reuse-trace-reports.json').read_text())
for name in NAMES:
path=ROOT/f'reuse-{name}.bin';raw=path.read_bytes();events=list(struct.iter_unpack('<QQII',raw))
last_claim=max(i for i,e in enumerate(events) if e[3]==0)
full=analyze(events,len(events)-1);proxy=analyze(events,last_claim)
assert full['valid'] and proxy['valid'],(name,full,proxy)
assert full['completed_reuse_slices']==old[name]['reused_slices']
assert full['completed_within_1ms_fraction']==old[name]['fraction_reused_under_1ms']
out['traces'][name]={'sha256':hashlib.sha256(raw).hexdigest(),'full_trace':full,'last_recorded_arena_claim_cutoff':proxy}
print(name,'lastclaim: avgfreeMiB',round(proxy['time_average_known_free_MiB'],2),'censoredshare',round(proxy['right_censored_share_of_observed_free_slice_time'],3),'areaage>1ms',round(proxy['share_of_free_slice_time_with_age_above_ms']['1'],3),'areaage>10ms',round(proxy['share_of_free_slice_time_with_age_above_ms']['10'],3),flush=True)
(ROOT/'reuse-time-weighted.json').write_text(json.dumps(out,indent=2)+'\n') |
#585: current platform CI complete — 104 SUCCESS, five intentional SKIPs, zero failuresCurrent head Passed coverage includes:
The serial reuse fixture is not skipped: actual JUnit has 36 passing purge/reuse entries across nine Linux bundles; musl's two new reuse entries both ran Intentional skipped checks:
Recovery execution is manual/selective; native Macs passed in this full lane. The minimal-only memory-gate check skips because the full Linux lane executed its real baseline gate and injected-leak positive control. Windows bootstrap warnings are a separate limitation inside successful steps, as stated above. The automatic perf-ab/sentinel triggers skip independently of the six explicit live diagnostic/acceptance dispatches. None is represented as a performance PASS. Exact check names, states and links: [
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119021576073",
"name": "PR test gate (windows-bundles) / verify",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119016149541",
"name": "PR test gate (macos-bundles) / verify",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119015779521",
"name": "PR test gate (c-unit) / verify",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119015711503",
"name": "coverage",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628377/job/119013565588",
"name": "PR test gate (rust-native) / verify",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628345/job/119013443953",
"name": "PR test gate (asan) / verify",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628332/job/119012771747",
"name": "PR test gate (fuzz) / verify",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628381/job/119012753140",
"name": "PR test gate (purge-teardown-regression) / verify",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119012598915",
"name": "ctest (windows-latest)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119012445195",
"name": "run-windows",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119012258846",
"name": "PR test gate (cross) / verify",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119012190771",
"name": "run-macos-native-full (arm64)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119012190686",
"name": "run-macos-native-full (x64)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119012192604",
"name": "run-macos-x64-selective (tests labelled macos, in Recovery)",
"state": "SKIPPED"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119012192003",
"name": "run-macos-x64 (macOS Recovery on Linux)",
"state": "SKIPPED"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011940711",
"name": "test (aarch64-unknown-linux-gnu)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011940670",
"name": "test (x86_64-pc-windows-msvc)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011940665",
"name": "test (x86_64-unknown-linux-musl)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011940646",
"name": "test (x86_64-pc-windows-gnu)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011940726",
"name": "test (x86_64-unknown-linux-gnu)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011886702",
"name": "run-linux",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011886669",
"name": "run-linux-serial",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011839571",
"name": "build-windows-gnu (windows-gnu-x64-leak)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011839584",
"name": "build-windows-msvc (windows-msvc-x64-leak)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011839442",
"name": "build-rust (x86_64-pc-windows-gnu)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011839377",
"name": "build-rust (x86_64-pc-windows-msvc)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011839511",
"name": "build-windows-msvc (windows-msvc-x64-shared)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011839494",
"name": "build-windows-gnu (windows-gnu-x64-gated)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011839499",
"name": "build-windows-gnu (windows-gnu-x64-release)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011839436",
"name": "build-windows-gnu (windows-gnu-x64-debug-full)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011839526",
"name": "build-windows-msvc (windows-msvc-x64-release)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011839508",
"name": "build-windows-msvc (windows-msvc-x64-gated)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011839487",
"name": "build-windows-msvc (windows-msvc-x64-debug-full)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011839444",
"name": "build-windows-gnu (windows-gnu-x64-shared)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011787833",
"name": "pr-ci-mode / decide",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628340/job/119011787615",
"name": "resolve-candidate (exact SHA )",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628381/job/119011745436",
"name": "reproduce (ON, true)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628381/job/119011745341",
"name": "reproduce (OFF, false)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628381/job/119011745370",
"name": "reproduce (ON, false)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545971",
"name": "build (memory-gate-leak)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119011570767",
"name": "build-macos (macos-x64-leak)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545931",
"name": "build (memevt-only)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545912",
"name": "build (gated)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545899",
"name": "build (diag-pprof-on)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545896",
"name": "build (debug-full)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545932",
"name": "build (guarded)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545738",
"name": "fastpath-identity",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119011570795",
"name": "build-macos (macos-arm64-release)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545857",
"name": "build (shared)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119011570703",
"name": "build-macos (macos-arm64-debug-full)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119011570764",
"name": "build-macos (macos-arm64-leak)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545805",
"name": "build-bun-objects (glibc)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119011570729",
"name": "build-macos (macos-x64-release)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119011570733",
"name": "build-macos (macos-x64-debug-full)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011593800",
"name": "memory-gate (minimal lane, linux pprof)",
"state": "SKIPPED"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545863",
"name": "build (release)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545625",
"name": "isa-baseline (ubuntu-24.04-arm)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011531383",
"name": "build (x86_64-pc-windows-msvc)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545889",
"name": "build (debug3-extra)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545856",
"name": "build (dhat-on)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545858",
"name": "build (musl-pprof-off)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545824",
"name": "build (isa-portable)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545868",
"name": "build (diag-pprof-off)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545675",
"name": "build-bun-objects (musl)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628289/job/119011544119",
"name": "lint",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545830",
"name": "build (musl)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545849",
"name": "build (pprof-off)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545961",
"name": "build (isa-arch)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545655",
"name": "decide (memory gate lane)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011545662",
"name": "build-windows-native",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011531274",
"name": "build (x86_64-pc-windows-gnu)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628377/job/119011538138",
"name": "test-win-gnu",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628377/job/119011538126",
"name": "test (windows-latest)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628377/job/119011538104",
"name": "test-no-pprof",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628377/job/119011538011",
"name": "test (ubuntu-latest)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011531252",
"name": "build (aarch64-pc-windows-msvc)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628332/job/119011536645",
"name": "fuzz",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119011529772",
"name": "build-rust (x86_64-apple-darwin)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119011529711",
"name": "build-rust (aarch64-apple-darwin)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011531262",
"name": "build (x86_64-unknown-linux-gnu)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011531354",
"name": "build (aarch64-unknown-linux-gnu)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011531333",
"name": "build (aarch64-apple-darwin)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119011529748",
"name": "decide (selective macOS lane)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011531300",
"name": "build (x86_64-apple-darwin)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011531330",
"name": "build (x86_64-unknown-linux-musl)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628345/job/119011527076",
"name": "address sanitizer (clang, Debug, owner-gate)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628345/job/119011527114",
"name": "address sanitizer (clang, Debug)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628345/job/119011527687",
"name": "address sanitizer (gcc, RelWithDebInfo)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628345/job/119011526983",
"name": "multi-threaded stress (debug-full, tsan)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628345/job/119011527112",
"name": "address sanitizer (clang, RelWithDebInfo)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628381/job/119011527229",
"name": "build (ON, false)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628381/job/119011527314",
"name": "build (OFF, false)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628381/job/119011527415",
"name": "build (ON, true)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628332/job/119011489896",
"name": "pr-ci-mode / decide",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119011490396",
"name": "resolve-candidate (exact SHA)",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628302/job/119011489875",
"name": "vendored C amalgamation matches src/",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011490911",
"name": "pr-ci-mode / decide",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011490091",
"name": "resolve-candidate (exact SHA )",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628377/job/119011490234",
"name": "resolve-candidate (exact SHA )",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628309/job/119011490106",
"name": "check",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628337/job/119011490533",
"name": "pr-ci-mode / decide",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628381/job/119011490224",
"name": "pr-ci-mode / decide",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628289/job/119011489608",
"name": "ci-mode",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628377/job/119011490486",
"name": "pr-ci-mode / decide",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628336/job/119011490428",
"name": "pr-ci-mode / decide",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628345/job/119011489885",
"name": "pr-ci-mode / decide",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628470/job/119011490518",
"name": "resolve-candidate (exact SHA )",
"state": "SUCCESS"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628292/job/119011490601",
"name": "benchmark-sentinel (${{ matrix.os }})",
"state": "SKIPPED"
},
{
"link": "https://github.com/zackees/mimalloc-pprof/actions/runs/39535628288/job/119011490559",
"name": "ab",
"state": "SKIPPED"
}
] |
Local-Gate: v1 tree=f7471f8b9f0d0d334508f5208f17de19f724b238 secs=35 lanes=lint:run
Ci-Attestation: {"at":1792553893,"gate":"python/all/lint","host":"linux-x86_64","key":"59477bdcb03af0225aea620651d157f51ae10560f6977c0d0152eee4315bc4f6","lane":"lint","parents":["da1fdb9819fdaa42a57e96d1d6cd741cf2711d03"],"secs":35,"stamp":"1992829c7fbbf4ecd3222dd9111a8f0f","tree":"f7471f8b9f0d0d334508f5208f17de19f724b238","v":1,"via":"run"}
#585: current short-generation chart row PASS at 60 actual samples per armRun39537470652, frozen merged base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All four actual timing/latency files SHA256-verified. Both baseline chart files match the earlier actual chart baseline byte for byte. Both head chart files differ from old15, as expected with the word-hash source change compiled into chart. The old15 120/240-sample INCONCLUSIVE outcomes remain unchanged and are not pooled into this different-source result. Chart has profiling compiled in but runtime-unarmed; its CPU result must not be attributed to executing the new hash loop. The fixed-baseline ledger reports the whole candidate. A new config/docs-only commit activates Windows memory baselines from historical merged-policy artifacts; it changes no allocator, header, benchmark or Rust source. This measured allocator remains da1fdb9; no JSON source ref is relabeled. The other five original individual runs remain live and are not cancelled or restarted. Actual executable manifest: {
"base/chart/perf_ab": "306d5c8a089d29a7ff276df621226fe2d08ffe7013669627f38dc5c0020d33f7",
"base/chart/perf_ab_latency": "2f883b21f4d48e9e1b404c3110ca058902c68cba23e99a82edd58026c7495090",
"head/chart/perf_ab": "83f28e34da86483f7c43947f64060944f344b460577d03849e2ac952915f8fbc",
"head/chart/perf_ab_latency": "1f77d5ba8f0e3da542412443fb9569b27fd898dd116f584cee79f4185a9acdad"
}Full-precision ledger: [
{
"workload": "large-class-ephemeral/8 short generations (chart build)",
"counts": {
"base": 60,
"head": 60,
"null": 60
},
"saved": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 7625855.5,
"head_median": 7677377.199999999,
"change": [
1.017136857954018,
0.5234892416622049,
1.7481919919695432
],
"regression": [
-1.7481919919695432,
-0.5234892416622049
],
"allowance": 0.7417747291280052,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 0.428,
"head_median": 0.42315,
"change": [
-1.214893414000124,
-1.6053978594695248,
-0.6511627906976672
],
"regression": [
-1.6053978594695248,
-0.6511627906976672
],
"allowance": 0.6052141527001842,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 250.0,
"head_median": 250.0,
"change": [
0.39526309211348576,
-0.3952569169960474,
0.8032128514056224
],
"regression": [
-0.3952569169960474,
0.8032128514056224
],
"allowance": 1.2096774193548387,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 1904.0,
"head_median": 1883.5,
"change": [
-1.0676485119880375,
-1.9969278033794162,
-0.5252100840336135
],
"regression": [
-1.9969278033794162,
-0.5252100840336135
],
"allowance": 1.0638297872340425,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 74.59765625,
"head_median": 74.6015625,
"change": [
0.0,
-2.6158788435061955,
2.6325430470508193
],
"regression": [
-2.6158788435061955,
2.6325430470508193
],
"allowance": 2.738620467018186,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 74.59765625,
"head_median": 74.6015625,
"change": [
0.0,
-2.6158788435061955,
2.6755327504057806
],
"regression": [
-2.6158788435061955,
2.6755327504057806
],
"allowance": 2.7546134395007265,
"verdict": "PASS"
}
]
}
] |
#585: activate missing Windows memory comparisons from historical merged-policy dataCurrent green platform run39535628340 had collected Windows results but skipped both memory baseline comparisons and leak-positive controls because the compiler-specific references were absent. This limitation is retained in the full-CI report. Green steps alone were not evidence that those comparisons ran. Historical reference, not candidate data: full run36812458448 executed source
Using these references, each current eight-run normal set PASSes the existing minimum-of-eight comparator, and each eight-run injected-leak set FAILs it as intended (GNU minimum125.7MB; MSVC131.6MB vs75.2MB allowance). Both Pushed attested config/docs-only Comparator outputSource path identity and all raw input SHA256 values{
"source_path_identity_verified": [
"src",
"include",
"test",
"CMakeLists.txt",
"cmake",
".github/workflows/windows-bundles.yml",
"ci/bundle_tests.py",
"ci/run_test_bundle.py",
"ci/memory_gate.py"
],
"checks": {
"gnu": {
"frozen_source": "80a513dfa6f84a16a07b8612c4041acb7cee4091",
"frozen_run": "36812458448",
"current_source": "da1fdb9819fdaa42a57e96d1d6cd741cf2711d03",
"current_run": "39535628340",
"compiler": "soldr-mingw-gcc-15",
"normal_old": 8,
"normal_current": 8,
"leak_current": 8,
"baseline_from_old_rc": 0,
"current_check_rc": 0,
"current_control_rc": 0
},
"msvc": {
"frozen_source": "80a513dfa6f84a16a07b8612c4041acb7cee4091",
"frozen_run": "36812458448",
"current_source": "da1fdb9819fdaa42a57e96d1d6cd741cf2711d03",
"current_run": "39535628340",
"compiler": "soldr-clang-cl-21",
"normal_old": 8,
"normal_current": 8,
"leak_current": 8,
"baseline_from_old_rc": 0,
"current_check_rc": 0,
"current_control_rc": 0
}
},
"artifact_sha256": {
"old-gnu/result-1.json": "20ce9f2f9ea5bd887e952e73bf4c7dc6f015df3dd7479cc6603f48f0db9226c5",
"old-gnu/result-2.json": "1cd1bbd6ed97a75cf47c7a86a0dcf798c2b372d64f119dc581b3a954a7e9407c",
"old-gnu/result-3.json": "26b65dad2175b5f6bcccd86af633cb10c93ad5fdcd30721d702327ec266d7abf",
"old-gnu/result-4.json": "72aa3ad19fb0addd8248e163f34252e1cef0dbe6bcf380df79733e1b7077e13b",
"old-gnu/result-5.json": "b77896aa12495379fa3882d3168288b22e560699008754f927870c36e39f1559",
"old-gnu/result-6.json": "7ca80dec1fdeecc8b551aa867708f0b6b35a16433f7b58140025127971d3ceb5",
"old-gnu/result-7.json": "57d530b510ae17f29ce163219a63ff3bd99552e5dbafb692f9f9c18d7a5f4614",
"old-gnu/result-8.json": "cb33fc68953d920be2dedac68bd460a11588578ddc4cfe49fec5691ebb50c347",
"current-gnu/result-1.json": "8994d207d1e8915fa800912d5d17201494b16a2ecf1ce9b54a0742418325c240",
"current-gnu/result-2.json": "bb21c5bd929630826ff05991f8b8829c3745741702c98885046b0286ca2abc80",
"current-gnu/result-3.json": "70ac449238c532e06d897141812871ca4cd7c2a1c736f5b194fa7ed83f233150",
"current-gnu/result-4.json": "72aa3ad19fb0addd8248e163f34252e1cef0dbe6bcf380df79733e1b7077e13b",
"current-gnu/result-5.json": "9ff362f50571e6b6c4676ebcd64814051ff154d3c65ca37ecc883d08cd4f7b8f",
"current-gnu/result-6.json": "6550b65fdf64b3d394f592a401ea658b46cf6adbab063a94b01a216c06a3197c",
"current-gnu/result-7.json": "9e35c2ea0cfe862fd88b2c1b5c6dee0db90e981efd4b089f4ef72f4ae5b32893",
"current-gnu/result-8.json": "d876eebc4df097f2ec67464aaa08fcfa83b2142d9cd1fd86f99a93d5c5225198",
"current-gnu/leak-1.json": "726df29832b82e3051a387c68671560bafb20ea7524ad609b7e4697711ebed23",
"current-gnu/leak-2.json": "2f9ea995cfadb1d67ecbe257a12d4e63f9a4cb2e71ad6dde7be18cb9b40acebe",
"current-gnu/leak-3.json": "49134c62c527d6783d89a416d951c241c6ba3370c38e530b03ab85b61db11383",
"current-gnu/leak-4.json": "f710ce8a3acae4aba7d020e3570c2eaa7401be5bd415c7e5d596170ff31a7352",
"current-gnu/leak-5.json": "cdab56f95f8c39fce18d46c88a12b9eb89de9e5bbcad42b6ebe9e70eb6e8a1af",
"current-gnu/leak-6.json": "8d94d01971f58878eeef72439ecd64e4ab00d2ad894b502eb85f172a89ac3686",
"current-gnu/leak-7.json": "c1b9a75c922b27fe57c02d7418e4853d397b13594bf18277d77a240f7fbc8e13",
"current-gnu/leak-8.json": "1db77d9a7218621cc455b311eb04aff1bea2578421d20ac55b2295fe4c2dbb3b",
"old-msvc/msvc-result-1.json": "0b6f467ea1176fae063c5b165b31de7570be6dcb0401512a9899ce5c0d8135f7",
"old-msvc/msvc-result-2.json": "52e64e84746407d291ae2b9ca62c990cee230698c91e0e8db206cd0c410aa8f4",
"old-msvc/msvc-result-3.json": "0d9706ae80a7148985a9c7e1ba5ac4376d3a55c1ddfdb8d2cdc65ca0338d0bbf",
"old-msvc/msvc-result-4.json": "7fa2daf1b2f91b627d457b735b7c41354741c14f610e84cc4353a7f378a4d99b",
"old-msvc/msvc-result-5.json": "99a124e148b9be1c1a3e65555f80e76f8d11b937119ff9fd30e816035d39a4e1",
"old-msvc/msvc-result-6.json": "b80e9b8798addf85705baaee1f453b1faa84faf8a940ba00aed968f0efa91ee8",
"old-msvc/msvc-result-7.json": "1e0805e67bc9d8dd6fd8e3687f6cd48f84d87329b7c06c2ce6062ba7e51ab39b",
"old-msvc/msvc-result-8.json": "1f112edd79aa7ee6cb3b6ecf978e6ac276a1a07c02749c52108d7b7a102bff39",
"current-msvc/msvc-result-1.json": "72aa3ad19fb0addd8248e163f34252e1cef0dbe6bcf380df79733e1b7077e13b",
"current-msvc/msvc-result-2.json": "26b65dad2175b5f6bcccd86af633cb10c93ad5fdcd30721d702327ec266d7abf",
"current-msvc/msvc-result-3.json": "af88bd4dc8e79b323b30efd4122da3d25b8d9a321be54d76c3afbed75e82fdd9",
"current-msvc/msvc-result-4.json": "bbd2154229bd97979bffd7731f3a6a7a855ef36b36752b882b84400e77f1f2ab",
"current-msvc/msvc-result-5.json": "3c691487494f64035c81292276c0774392c020e95b2036f103aa07709e19b081",
"current-msvc/msvc-result-6.json": "af88bd4dc8e79b323b30efd4122da3d25b8d9a321be54d76c3afbed75e82fdd9",
"current-msvc/msvc-result-7.json": "9ff362f50571e6b6c4676ebcd64814051ff154d3c65ca37ecc883d08cd4f7b8f",
"current-msvc/msvc-result-8.json": "b80e9b8798addf85705baaee1f453b1faa84faf8a940ba00aed968f0efa91ee8",
"current-msvc/msvc-leak-1.json": "4b7b0257df3bca46d32801269ae4778fea34d40b7dcce4a5ba1d8ce1d468c4a9",
"current-msvc/msvc-leak-2.json": "c13ecb965c06f951d8ba06dd9d96319a82c2ff0c3d170ae35b7bb7da6038751b",
"current-msvc/msvc-leak-3.json": "a43263983405932788ef84e228fb6f0433dce41c8259871b0ba54e0ef8cc5967",
"current-msvc/msvc-leak-4.json": "ed99d3810925099bfb9890401ddedbff2abf861ab44c86b943cdca20ccaafbdb",
"current-msvc/msvc-leak-5.json": "4f6f401390615ae60e206c18685f58ee411d757ec3ab260e430868e5cc54b1bc",
"current-msvc/msvc-leak-6.json": "d81f05b0d7feda227dbe35e50047363a82afca361b67cddb14ad64a40cee6b48",
"current-msvc/msvc-leak-7.json": "fab9fdadf833cad412045c8c40e5153dd9c2681ed564a2541cf7c766f2a70917",
"current-msvc/msvc-leak-8.json": "016cbde5c0de98b9db32d3611b2842b6c89efdf416c4f2b63365df345a1bf3c5"
}
}Authoritative ancestry{"ahead_by":6,"behind_by":0,"merge_base_commit":"80a513dfa6f84a16a07b8612c4041acb7cee4091","status":"ahead"} |
#585: random/8 and standard chart ephemeral PASS at 80 actual samples per armBoth runs retain frozen EE → allocator candidate da1fdb9, original exact workload parameters, balanced arm order, separate latency replay, same-run null arm and unchanged sanity/ledger/latency gates. 80 actual base/head/null samples per row, not the header's starting count60. Eight workers on four logical CPUs remain2:1 oversubscribed. No pooling across CPU models or source refs; no original failed/INCONCLUSIVE outcome is erased. 39537464950 — AMD EPYC 9V45 96-Core Processorbase Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All six metrics PASS, zero memory credit. CPU +0.244% [-0.874%,+0.723%] fits0.805%nullnoise; this is not a CPU improvement claim. After-drain RSS interval [-0.0189825%,0%] is not identically zero. All four actual files are byte-identical to the earlier old15 plain pair; this is additional independently reported evidence, not a source-change claim or relabeling of its two prior INCONCLUSIVE runs. Four actual executable SHA256 entries verified: {
"base/plain/perf_ab": "7f0c70ce2450dfa09126c1674bddab53187614f12d5bd648b4fda33ccbd3c712",
"base/plain/perf_ab_latency": "ef2d5d273681a473e253d1dd2080307f1ae9f09d34977bda2ad4cab9cbfe7823",
"head/plain/perf_ab": "9ee3617e371e93195de1374ef403455e7ae660848d689ce4bf0ae0ce2f8565b1",
"head/plain/perf_ab_latency": "c14766d194fd7f3c07652b148d1de18e3e4db90b4414bd563398f5f44ee9b982"
}Full-precision cells: [
{
"workload": "random-large/8",
"counts": {
"base": 80,
"head": 80,
"null": 80
},
"saved": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 987053.05,
"head_median": 987009.8999999999,
"change": [
0.13973691200725502,
-1.0739905599530075,
0.8145600147609683
],
"regression": [
-0.8145600147609683,
1.0739905599530075
],
"allowance": 1.506251182961233,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 1.28225,
"head_median": 1.2824499999999999,
"change": [
0.2437701956490842,
-0.8738481093104464,
0.7234073964715458
],
"regression": [
-0.8738481093104464,
0.7234073964715458
],
"allowance": 0.8046593608705613,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 264.0,
"head_median": 264.0,
"change": [
0.37736386420596946,
-0.7299270072992701,
0.7782101167315175
],
"regression": [
-0.7299270072992701,
0.7782101167315175
],
"allowance": 1.532567049808429,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 15768.5,
"head_median": 15768.5,
"change": [
0.4694225007753404,
-2.647684081685241,
2.576347869800176
],
"regression": [
-2.647684081685241,
2.576347869800176
],
"allowance": 3.326234026559759,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 182.375,
"head_median": 182.3828125,
"change": [
-0.007548307659872912,
-1.113746082832218,
0.06639537374170058
],
"regression": [
-1.113746082832218,
0.06639537374170058
],
"allowance": 1.1474368800227734,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 20.578125,
"head_median": 20.578125,
"change": [
0.0,
-0.018982536066818528,
0.0
],
"regression": [
-0.018982536066818528,
0.0
],
"allowance": 0.018982536066818528,
"verdict": "PASS"
}
]
}
]39537468554 — AMD EPYC 7763 64-Core Processorbase Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All six metrics PASS, zero memory credit. CPU -1.060% [-1.258%,-0.855%], p99 -1.606% [-2.176%,-1.115%]. The minor-fault upper bound1.185770751% is below1.190476190%nullnoise; the drain upper bound10.591350397% equals its original null-noise allowance exactly (PASS under the unchanged <= rule). These comparisons use unrounded values. No RSS reduction is established. Both actual chart executables match the new short-generation chart files; chart has profiling compiled in but runtime-unarmed, so the CPU result is not evidence of executing a faster hash loop. Four actual executable SHA256 entries verified: {
"base/chart/perf_ab": "306d5c8a089d29a7ff276df621226fe2d08ffe7013669627f38dc5c0020d33f7",
"base/chart/perf_ab_latency": "2f883b21f4d48e9e1b404c3110ca058902c68cba23e99a82edd58026c7495090",
"head/chart/perf_ab": "83f28e34da86483f7c43947f64060944f344b460577d03849e2ac952915f8fbc",
"head/chart/perf_ab_latency": "1f77d5ba8f0e3da542412443fb9569b27fd898dd116f584cee79f4185a9acdad"
}Full-precision cells: [
{
"workload": "large-class-ephemeral/8 (chart build)",
"counts": {
"base": 80,
"head": 80,
"null": 80
},
"saved": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 8196132.699999999,
"head_median": 8311705.65,
"change": [
1.603993854553773,
1.0971011154024806,
1.8347869001337975
],
"regression": [
-1.8347869001337975,
-1.0971011154024806
],
"allowance": 0.43192585753215545,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 1.5466,
"head_median": 1.5307,
"change": [
-1.0597716959638057,
-1.2584327970939337,
-0.8548669127647238
],
"regression": [
-1.2584327970939337,
-0.8548669127647238
],
"allowance": 0.394362554952159,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 252.0,
"head_median": 252.0,
"change": [
0.3929288250733364,
-0.7843137254901961,
1.185770750988142
],
"regression": [
-0.7843137254901961,
1.185770750988142
],
"allowance": 1.1904761904761905,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 1853.0,
"head_median": 1823.0,
"change": [
-1.6060072235935183,
-2.176220806794055,
-1.1154489682097044
],
"regression": [
-2.176220806794055,
-1.1154489682097044
],
"allowance": 1.1031439602868174,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 74.638671875,
"head_median": 76.482421875,
"change": [
-0.010473124962651806,
-0.06959634122091866,
0.08169100377820893
],
"regression": [
-0.06959634122091866,
0.08169100377820893
],
"allowance": 2.5929806181246726,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 8.7890625,
"head_median": 8.818359375,
"change": [
0.02206531332744925,
-0.044111160123511246,
10.59135039717564
],
"regression": [
-0.044111160123511246,
10.59135039717564
],
"allowance": 10.59135039717564,
"verdict": "PASS"
}
]
}
]The three remaining original control jobs (small/8, Larson/8 plain/chart) are still live. New head3db601e1 changes only historical Windows memory references and documentation; allocator/header/test/CMake/Rust/benchmark/harness paths are byte-identical to da1fdb9. Fresh platform CI must exercise the newly enabled Windows checks; the prior104PASS5intentionalSKIP result belongs to da. No whole-candidate completion is claimed. |
#585: final controls — two PASS, chart Larson remains INCONCLUSIVEOriginal frozen EE→da1fdb98 candidate source, exact workload params, separate latency replay, balanced arm positions, unchanged null/sanity/ledger gates and requested60-to-120 plan. Actual counts come from raw sample-array lengths below. Four actual timing/latency hashes per row were verified. These results establish12of13 original cells with supported PASS evidence; chart Larson remains unresolved, so whole-candidate acceptance is incomplete. 39537461436 — AMD EPYC 9V45 96-Core ProcessorActual counts: {'base': 110, 'head': 110, 'null': 110}. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All six PASS at110/arm, zero credit. P99, peak and drain percent-change intervals are exactly[0,0]; that statement concerns the paired sign-test intervals, not every individual observation. CPU interval crosses0; no CPU improvement claim. Verified actual executable manifest: {
"base/plain/perf_ab": "7f0c70ce2450dfa09126c1674bddab53187614f12d5bd648b4fda33ccbd3c712",
"base/plain/perf_ab_latency": "ef2d5d273681a473e253d1dd2080307f1ae9f09d34977bda2ad4cab9cbfe7823",
"head/plain/perf_ab": "9ee3617e371e93195de1374ef403455e7ae660848d689ce4bf0ae0ce2f8565b1",
"head/plain/perf_ab_latency": "c14766d194fd7f3c07652b148d1de18e3e4db90b4414bd563398f5f44ee9b982"
}Full-precision cells: [
{
"workload": "small/8 (control)",
"counts": {
"base": 110,
"head": 110,
"null": 110
},
"saved": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 137009922.8,
"head_median": 136653328.95,
"change": [
0.5727066285388089,
-0.2537313947741482,
1.0784954237927962
],
"regression": [
-1.0784954237927962,
0.2537313947741482
],
"allowance": 1.1835247411663612,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 1.1146500000000001,
"head_median": 1.108,
"change": [
-0.5910932884950753,
-1.2378026743765853,
1.3266536784488336
],
"regression": [
-1.2378026743765853,
1.3266536784488336
],
"allowance": 1.535068372298179,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 160.0,
"head_median": 160.0,
"change": [
0.0,
-0.6172839506172839,
0.0
],
"regression": [
-0.6172839506172839,
0.0
],
"allowance": 0.6172839506172839,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 90.0,
"head_median": 90.0,
"change": [
0.0,
0.0,
0.0
],
"regression": [
0.0,
0.0
],
"allowance": 0.0,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 16.484375,
"head_median": 16.484375,
"change": [
0.0,
0.0,
0.0
],
"regression": [
0.0,
0.0
],
"allowance": 0.0,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 16.484375,
"head_median": 16.484375,
"change": [
0.0,
0.0,
0.0
],
"regression": [
0.0,
0.0
],
"allowance": 0.0,
"verdict": "PASS"
}
]
}
]39537463373 — AMD EPYC 7763 64-Core ProcessorActual counts: {'base': 115, 'head': 115, 'null': 115}. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
All six PASS at115/arm, zero credit. P99 upper bound0.43478260869565216% equals the original null-noise allowance exactly, so it passes the unchanged <= comparison. CPU interval crosses0. RSS intervals are not identically zero. Verified actual executable manifest: {
"base/plain/perf_ab": "7f0c70ce2450dfa09126c1674bddab53187614f12d5bd648b4fda33ccbd3c712",
"base/plain/perf_ab_latency": "ef2d5d273681a473e253d1dd2080307f1ae9f09d34977bda2ad4cab9cbfe7823",
"head/plain/perf_ab": "9ee3617e371e93195de1374ef403455e7ae660848d689ce4bf0ae0ce2f8565b1",
"head/plain/perf_ab_latency": "c14766d194fd7f3c07652b148d1de18e3e4db90b4414bd563398f5f44ee9b982"
}Full-precision cells: [
{
"workload": "larson/8",
"counts": {
"base": 115,
"head": 115,
"null": 115
},
"saved": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 69396089.1,
"head_median": 69419002.6,
"change": [
-0.5763927626097618,
-1.4295460178576465,
0.6754231657754093
],
"regression": [
-0.6754231657754093,
1.4295460178576465
],
"allowance": 1.6922745482674078,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 1.0922,
"head_median": 1.0964,
"change": [
0.02773925104023939,
-0.7326355851569972,
1.0236781199928846
],
"regression": [
-0.7326355851569972,
1.0236781199928846
],
"allowance": 1.1007250808071936,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 623.0,
"head_median": 607.0,
"change": [
-1.611047180667434,
-9.824561403508772,
6.772334293948126
],
"regression": [
-9.824561403508772,
6.772334293948126
],
"allowance": 10.208816705336426,
"verdict": "PASS"
},
{
"metric": "p99 operation ns",
"base_median": 230.0,
"head_median": 230.0,
"change": [
0.0,
0.0,
0.43478260869565216
],
"regression": [
0.0,
0.43478260869565216
],
"allowance": 0.43478260869565216,
"verdict": "PASS"
},
{
"metric": "peak RSS MiB",
"base_median": 289.56640625,
"head_median": 289.7890625,
"change": [
-0.026319945254513873,
-0.5409219190968956,
1.1169807353675523
],
"regression": [
-0.5409219190968956,
1.1169807353675523
],
"allowance": 2.463345528567106,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 209.44921875,
"head_median": 209.7578125,
"change": [
0.028017482909335426,
-0.38910505836575876,
0.2346456106372677
],
"regression": [
-0.38910505836575876,
0.2346456106372677
],
"allowance": 0.3609570944814498,
"verdict": "PASS"
}
]
}
]39537466521 — AMD EPYC 7763 64-Core ProcessorActual counts: {'base': 120, 'head': 120, 'null': 120}. base Rule-12 ledger (CLAUDE.md; ci/perf_ab_ledger.py):
At120/arm, INCONCLUSIVE on minor faults, p99 and drain RSS. CPU, throughput and peak PASS, zero credit. P99 upper0.4166666666666667% exceeds0.398406374501992%nullnoise; rounding both to0.4% cannot turn this into PASS. The absolute head/base p99 delta interval is[0,1]ns and null/base is[-1,0]ns, but the declared percentage-based rule remains authoritative. Fault upper14.4068% exceeds12.4218%noise; drain upper2.3268% exceeds1.4882%noise. There is no FAIL verdict, but this is not accepted. Verified actual executable manifest: {
"base/chart/perf_ab": "306d5c8a089d29a7ff276df621226fe2d08ffe7013669627f38dc5c0020d33f7",
"base/chart/perf_ab_latency": "2f883b21f4d48e9e1b404c3110ca058902c68cba23e99a82edd58026c7495090",
"head/chart/perf_ab": "83f28e34da86483f7c43947f64060944f344b460577d03849e2ac952915f8fbc",
"head/chart/perf_ab_latency": "1f77d5ba8f0e3da542412443fb9569b27fd898dd116f584cee79f4185a9acdad"
}Full-precision cells: [
{
"workload": "larson/8 (chart build)",
"counts": {
"base": 120,
"head": 120,
"null": 120
},
"saved": 0.0,
"cells": [
{
"metric": "ops/s",
"base_median": 64240218.6,
"head_median": 64660126.400000006,
"change": [
0.6639348247785972,
-0.26962345445616737,
1.3098932733762891
],
"regression": [
-1.3098932733762891,
0.26962345445616737
],
"allowance": 0.9956393865420201,
"verdict": "PASS"
},
{
"metric": "cpu s",
"base_median": 1.1873,
"head_median": 1.18275,
"change": [
-0.5590210131837745,
-1.2940575515069046,
0.5143591941705962
],
"regression": [
-1.2940575515069046,
0.5143591941705962
],
"allowance": 1.4303329223181143,
"verdict": "PASS"
},
{
"metric": "minor faults",
"base_median": 663.5,
"head_median": 652.5,
"change": [
-1.9171514139261647,
-13.478260869565217,
14.40677966101695
],
"regression": [
-13.478260869565217,
14.40677966101695
],
"allowance": 12.421805183199284,
"verdict": "INCONCLUSIVE"
},
{
"metric": "p99 operation ns",
"base_median": 250.0,
"head_median": 250.0,
"change": [
0.0,
0.0,
0.4166666666666667
],
"regression": [
0.0,
0.4166666666666667
],
"allowance": 0.398406374501992,
"verdict": "INCONCLUSIVE"
},
{
"metric": "peak RSS MiB",
"base_median": 285.2578125,
"head_median": 285.533203125,
"change": [
0.7498337345892028,
-0.7914719235044394,
2.3084175480981703
],
"regression": [
-0.7914719235044394,
2.3084175480981703
],
"allowance": 3.2188553991113014,
"verdict": "PASS"
},
{
"metric": "RSS 0.5 s after drain MiB",
"base_median": 46.3515625,
"head_median": 46.748046875,
"change": [
0.7405989186355686,
-0.4105831016515383,
2.3267750213858
],
"regression": [
-0.4105831016515383,
2.3267750213858
],
"allowance": 1.4881698854610819,
"verdict": "INCONCLUSIVE"
}
]
}
]The new3db601e1 commit changes only Windows historical reference data and documentation; no measured program source changed. Fresh correctness CI is still running to exercise the newly activated Windows memory comparisons and controls. No source refs are relabeled, no different-CPU pooling, no failed/INCONCLUSIVE results erased, and no acceptance exception is made. A focused larger-budget chart-Larson follow-up is being evaluated; it will retain the original trace, sampling and thresholds. |
Current 3db Windows memory comparisons and injected-leak controls executedRun 39539769768 has completed all four memory/comparator/control steps with SUCCESS on current config-only Both artifacts are already uploaded. Downloading and rejudging their actual eight normal and eight leak records per compiler with the unchanged committed comparator independently reproduces normal PASS and positive-control PASS. A positive-control PASS means the leak set correctly triggers a regression FAIL.
These are peak commit, not peak RSS or a new memory saving. The existing gate uses the minimum of eight runs; this table does not assert each individual run is below the allowance. References still contain only historical merged-source normal samples documented in the provenance report; no current candidate record initializes a reference, and no tolerance changed. The prior da run's missing-baseline bootstrap remains disclosed. Terminal job logs will be inspected once available. Artifacts: SHA256 of the 32 actual CI input records |
Current full-platform verification completed (3db601e)Current PR head Passing gates include native Microsoft Terminal logs now prove the Windows memory comparators and injected-leak controls actually executed against the two intended compiler-specific references. Both normal minima are 71.6 MB peak commit against 75.2 MB allowed; GNU leak minimum131.6 MB and clang-cl leak minimum126.8 MB trigger expected FAIL and The actual Linux log likewise executes its memory gate: normal58.1 MB peak RSS against baseline58.0/allowed60.9 MB, and injected-leak154.1 MB correctly rejected with Intentional skipped checks:
Performance acceptance remains separate and incomplete: twelve of thirteen original cells have supported PASS evidence; chart Larson is INCONCLUSIVE120/arm, with the single predeclared 120-to-180 follow-up39541307803 still running. No memory policy or RSS saving is newly accepted, and earlier failures are preserved. |
This branch has not been deployed
1 participant
#585: large-page CPU costs and measurement gaps
Draft; 12 of 13 original performance cells have PASS support. Chart Larson remains INCONCLUSIVE; one bounded follow-up is running; full platform CI has completed. Current head is config/docs-only
3db601e13aace89ca44b2be9dbd59b2ee8f5f034, activating historical Windows memory references. Allocator/header/test/CMake/Rust/benchmark/harness source is unchanged from measuredda1fdb98; that source includes word-hash C0331d770, separate Rust regeneration6f7be38f, and the serial test fixture correction. Fixed baseline isee3623f6b7a485754a0a0aa3462d8cba43390e8b. Previous15bfcba0profiler p99 FAIL remains recorded.Earlier failed and incomplete runs remain recorded below. The new prototype changes only profiler stack hashing. Harness checkpoints now survive forced termination; previous candidate failures remain visible. A rebase onto unchanged origin/main completed; final revalidation and scratch cleanup remain pending.
Allocator candidate
Frozen baseline:
ee3623f6b7a485754a0a0aa3462d8cba43390e8b. Old timed candidate:9b67e411fc8c8899e07e05d528a4c5b37ab20417. Its results below do not validate the new byte-threshold prototype in C commitfc1ee95e. The prototype precomputes the minimum capacity in bytes with overflow handling, removing a per-candidate division while preserving eligibility and first-best selection; Rust regeneration is a separate commit.No new memory policy has been retained. Geometry trimming, singleton compaction and resident-retry probes were rejected for failed fungibility, worse RSS or no useful benefit. A new capacity-based fit variant also failed the unchanged fungibility check (four pages vs limit three) and was rejected before RSS/timing: diagnostic patch and outcomes.
New word-hash prototype validation
Current
6f7be38freplaces byte-wise FNV rounds with the existing stateless word mixer instack_hash; full depth/PC equality remains the collision check. No layout, capture, sampling/PRNG, arena or record behavior changes. Six saved local cycle annotations place 39.1–55.7% of intern samples in the old hash loop under the global profiler lock; this is a CPU target, not an established CI failure cause.Debug profiler tests: 10/10 PASS; soldr regeneration/check PASS; same sole primary review clean; 45/45 local gate PASS in 32s. Platform CI failed the musl page-reuse assertion; no retry or relaxed assertion is being used to waive it. Fixed-baseline exact-profiler acceptance 39532706232 compares EE → 6f7be38; hash attribution 39532708955 compares 15bfcba → 6f7be38. Both retain 60-to-120/null/sanity/original trace. Fixed-baseline run 39532706232 completed with 60 actual samples per arm, all six metrics PASS on Xeon 8573C: CPU -0.9% [-2.7%,+1.4%], p99 -1.4% [-3.7%,+0.7%], zero memory credit. Four actual executable hashes verified. This one-row result does not establish an RSS improvement or resolve the musl correctness failure. Hash-step attribution completed at 120 actual samples/arm on EPYC9V74: CPU -1.84% [-2.11%,-1.55%], p99 -2.33% [-4.57%,-1.04%]; after-drain RSS remains INCONCLUSIVE, so the row is not an overall PASS. Full ledger. The six individual Linux measurements have completed: five PASS, chart Larson INCONCLUSIVE. Acceptance still requires all full CI gates and resolution of that cell. Controlled baseline interleaving and fixture verification: EE and 6f both create one new page if the scanner owns A’s reserved page during B’s allocation; both reuse A with no scanner. This proves a possible mechanism, not the historical CI cause. The original same-page and zero-created assertions run in fresh processes with scavenging disabled and verified; background release cases remain enabled. Native/musl verification and four actual CTest entries passed; 45/45 local gate passed in 35s; same reviewer clean.
Current plain executable identity and remaining validation
Build-only 39536056572 freshly built EE and da1fdb9. All four actual linked plain timing/latency files are byte-for-byte identical to the earlier timed EE/15 pair; six report file hashes verified. This carries forward six distinct plain PASS rows (main, plain ephemeral, sparse, bursty, Larson/1, random/1) with their original refs/hosts/counts/verdicts; no new timing, relabeling or pooling. It does not resolve plain small/8, Larson/8 or random/8. Chart enables PPROF and its binaries changed, so Larson/8 plus both ephemeral chart rows need new measurements. All six individual measurements completed under the original 60-to-120/null/sanity plans: five PASS and chart Larson INCONCLUSIVE. Linux parallel/serial, both native Macs, native MSVC, Rust/cross/ASan passed on current 3db; Windows bundles subsequently completed SUCCESS. Removing the assistant-imposed scheduling dependency changes no final acceptance gate. Six live run handles and scheduling scope. Identity evidence and scope.
Windows memory baseline verification gap fixed
The completed da1fdb9 CI collected Windows data but skipped both compiler-specific baseline checks and leak-positive controls. The new references use only eight historical normal runs per compiler from merged-source80a513/run36812458448, whose allocator/test/build paths are byte-identical to frozen EE. No candidate data or changed tolerance initializes a reference. Current minimum-of-eight sets PASS at71.6MB peak commit, and GNU/MSVC leak sets fail at125.7/131.6MB. Current3db CI has now executed all four memory checks and controls: normal minimum71.6MB for each compiler, GNU leak131.6MB and MSVC leak126.8MB correctly rejected. All32 actual input hashes verified, unchanged comparator independently reproduced, same sole reviewer clean; fresh artifact results. Overall Windows job subsequently completed SUCCESS; current full CI is104SUCCESS/fiveintentionalSKIPs. New3db601e1 activates those previously skipped CI paths; full provenance, output and input hashes. This is historical Windows memory-gate validation, not a same-host performance ledger or RSS saving.
Current remaining-row outcomes
Every completed new row's four actual timing/latency file hashes are verified. Old15 INCONCLUSIVE outcomes stay attached to their original source/hosts/counts; they are not relabeled or pooled. Chart executables changed with word-hash compiled in but profiling runtime-unarmed, so chart CPU deltas are not attributed to executing that loop. Current3db config changes no measured program source.
Previous byte-threshold candidate validation
15bfcba0against frozenee3623f6: exact plain ephemeral/8 PASS, 65 actual samples/arm, EPYC 7763, all four actual executable hashes verified. Baseline executables match the prior failed run byte for byte; separate VMs/source pairs are not pooled. The old candidate's failure remains below.The following original-cell validations retain 60-to-120 samples/arm, original traces and unchanged gates:
The isolated aging-setting trial 39521589663 finished at 30 samples/arm on Xeon 8573C. Default 4 PASSed all six metrics, with CPU -1.9%; its executables match the current candidate byte for byte. Setting 2 is rejected: peak RSS -3.9% [-5.3%, -2.6%], but p99 +9.8% [7.6%, 10.8%] versus 1.4% noise FAIL. It also costs 2.15% CPU relative to the default on the same VM. Both ledgers retain the frozen baseline; no setting was changed or retried after failure. The separate ordinary main/profiler run passed plain but failed profiler p99; the four-arm default-trial PASS does not waive that failure. All local screen pairs and scope limits.
Corrected old-candidate measurements
All rows use the exact source pair above, a same-run null arm, balanced arm positions, original work traces and unchanged sanity/ledger thresholds. Actual timing and separate latency executables are preserved with verified SHA256 manifests. Each result is a same-host paired comparison; CPU models are listed because these runs are not pooled across models.
The original EPYC 9V74 small/8 and Larson/1 rows remain INCONCLUSIVE; independent 7763 passes do not reclassify those rows. Every control now has a conclusive PASS in a corrected run. The focused control binaries match the original control binaries byte for byte. All old-candidate control/random follow-ups are terminal; different CPU models are not pooled.
The target failure is preserved with all twelve actual timing/latency executable hashes verified: full target ledger and binary evidence. The candidate remains unaccepted. Source attribution is terminal: baseline → exact-fit-only (39514174840) passed all six metrics with 60 samples/arm on EPYC 9V74. Exact-fit-only → mapped-prefix reuse (39514578328) remains INCONCLUSIVE with 60/arm on EPYC 9V45. An earlier prefix diagnostic (39514177655) failed its first sanity screen, including the unchanged-baseline null executable: all raw data is retained as incomplete, with no verdict or pooling. None of these comparisons erases the combined-candidate p99 failure. Full attribution and limits.
Full ledgers and scope limits: initial controls and untimed residency attribution, focused controls, random/sparse and follow-ups.
Earlier failures remain in the record
Previous head
c54e2a26had 102 successes, five skips and two failed checks from the GCC ASan idle-release deadline assertion and verifier. Its cause remains unexplained; the current source subsequently passed that lane and all 104 platform checks. Full current platform evidence.The previous perf-ab arm schedule only reversed endpoints, leaving head in the middle every repetition. Those runs remain diagnostic; their order protocol cannot be pooled with corrected runs.
Failure evidence, corrected exact-fit diagnostic and baseline-only drain observations.
Tooling delivered
No new workflow file was added. C-core and Rust paths are separate in every commit.
Historical CPU attribution
Corrected pre/post-merge runs preserve exact binaries and null arms. Separate CPU models and VMs prevent adding these deltas into a decomposition of the whole fork/upstream gap.
Full corrected historical ledgers and limits. #544 is the issue for refining demand-sized spans; #569 and #570 are two merged implementations within that issue. Its broader acceptance checklist is not independently certified by these narrow comparisons. Historical results do not substitute for current-candidate acceptance.
Remaining memory evidence
Same-run competitor scaling used 15 balanced blocks, 1,800 samples and zero paired trace/checksum mismatches. Allocator C matches the old timed candidate; its temporary harness also carried the tested benchmark-builder dependency from #608. No chart was published.
The prior sparse gap was not reproduced. Persistent and power-of-two gaps remain. Scaling CPU/op, faults and p99 were not measured; this is diagnostic evidence, not a fixed-baseline acceptance ledger. Persistent fork/upstream throughput remains 0.874 [0.860, 0.905]. Scaling report.
Independent prefix attribution locates about 27 MiB of resident unformed capacity in persistent/8, already present early. These are medians of three separate truncated processes per prefix, not continuous-run snapshots. Completed reuse intervals cluster near 14–16 microseconds, with censoring limits documented; no discard policy was retained. Attribution and rejected-probe context.
Verification and remaining work
Local debug C suite: 89/89. Default fast-path identity passed against pinned upstream and frozen baseline; structural RED/GREEN, same-build layout checks and Rust regeneration are covered. Previous-head full platform checks include Linux ON/OFF, native MSVC
cl, win-gnu/MSVC bundles, native macOS ARM64/x64, Rust cross targets, ASan, TSan and fuzz. The selector head subsequently failed the GCC ASan release-deadline assertion. Matching GCC 13.3 local probes passed 10 unchanged-baseline and 10 unchanged-old-candidate runs; 20 additional diagnostic runs per arm also passed with used=1, capacity=1 and reserved=3 before free. Those local observations do not explain the prior CI failure. The new candidate subsequently passed that GCC ASan lane; all 104 platform checks subsequently passed (five intentional skips). Failure evidence.Remaining: resolve chart Larson performance uncertainty; archive the now-verified actual Windows/Linux memory comparisons and leak controls; record decisions for remaining memory gaps; finish the requirement audit; fetch and rebase onto origin/main again; archive final evidence and remove owned scratch artifacts. Temporary remote branches and extra worktrees are already removed. This PR does not claim that #585 is complete.
Free-capacity time attribution
Existing trace reanalysis reproduces completed reuse counts and independently conserves observed free-slice time. About 99% of completed intervals are under 1ms, but at the provisional last-arena-claim cutoff, 37–51% of virtual free-slice time has age over 1ms. This is not resident-byte-time, a peak/work-phase marker, or evidence for a new discard threshold. Phase and residency events are required before a cold-release policy. All raw trace hashes, observed censored intervals, clock-order inversions, full-window results and analysis code are recorded; no new memory policy was retained.