ci: move benchmark runners to c7gd.metal (Graviton3)

.github/actions/setup-duckdb/action.yml

@@ -10,8 +10,13 @@ runs:
     - name: Download DuckDB
       shell: bash
       run: |
+        case "$(uname -m)" in
+          x86_64) arch=amd64 ;;
+          aarch64) arch=arm64 ;;
+          *) echo "unsupported architecture: $(uname -m)" >&2; exit 1 ;;
+        esac
         wget -qO- \
-          "https://github.com/duckdb/duckdb/releases/download/v${{ inputs.duckdb_version }}/duckdb_cli-linux-amd64.zip" \
+          "https://github.com/duckdb/duckdb/releases/download/v${{ inputs.duckdb_version }}/duckdb_cli-linux-${arch}.zip" \
           | funzip > duckdb
         chmod +x duckdb
         echo "$PWD" >> "$GITHUB_PATH"

.github/workflows/develop-bench.yml

@@ -31,11 +31,14 @@ jobs:
           bash scripts/commit-json.sh > new-commit.json
           bash scripts/cat-s3.sh vortex-ci-benchmark-results commits.json new-commit.json
 
+  # Benchmarks run on Graviton3 (c7gd.metal). The binary is built on the same CPU family so
+  # `-C target-cpu=native` matches, and both jobs pin the arm64 image because the
+  # `bench-dedicated` runner defaults to x64.
   build:
     timeout-minutes: 60
     runs-on: >-
       ${{ github.repository == 'vortex-data/vortex'
-          && format('runs-on={0}/runner=bench-dedicated/family=c6id.8xlarge/extras=s3-cache/tag=build-{1}', github.run_id, matrix.benchmark.id)
+          && format('runs-on={0}/runner=bench-dedicated/family=c7gd.8xlarge/image=ubuntu24-full-arm64-pre-v2/extras=s3-cache/tag=build-{1}', github.run_id, matrix.benchmark.id)
           || 'ubuntu-latest' }}
     strategy:
       fail-fast: false
@@ -59,8 +62,12 @@ jobs:
           enable-sccache: ${{ github.repository == 'vortex-data/vortex' && 'true' || 'false' }}
       - name: Build binary
         shell: bash
+        # `-C target-feature=-sve,-sve2`: on Graviton3, `target-cpu=native` (neoverse-v1) makes rustc
+        # 1.98 / LLVM 22 miscompile widening `u8 -> usize` sums at 256-bit SVE, which undersizes
+        # the OnPair and FSST decode buffers for every Vortex file with `u8` string lengths. See
+        # `benchmarks/string-bench/examples/sve_widening_sum.rs`.
         env:
-          RUSTFLAGS: "-C target-cpu=native -C force-frame-pointers=yes"
+          RUSTFLAGS: "-C target-cpu=native -C target-feature=-sve,-sve2 -C force-frame-pointers=yes"
         run: |
           cargo build --bin ${{ matrix.benchmark.id }} --profile release_debug ${{ matrix.benchmark.build_args }} --features unstable_encodings
       - name: Pre-upload benchmark debuginfo to Polar Signals
@@ -83,7 +90,7 @@ jobs:
     timeout-minutes: 120
     runs-on: >-
       ${{ github.repository == 'vortex-data/vortex'
-          && format('runs-on={0}/runner=bench-dedicated/family=c6id.metal/extras=s3-cache/tag={1}', github.run_id, matrix.benchmark.id)
+          && format('runs-on={0}/runner=bench-dedicated/family=c7gd.metal/image=ubuntu24-full-arm64-pre-v2/extras=s3-cache/tag={1}', github.run_id, matrix.benchmark.id)
           || 'ubuntu-latest' }}
     strategy:
       fail-fast: false

.github/workflows/nightly-bench.yml

@@ -23,11 +23,15 @@ jobs:
     with:
       mode: "develop"
       machine_type: ${{ matrix.machine_type.instance_name }}
+      build_machine_type: ${{ matrix.machine_type.build_instance_name }}
+      machine_image: ${{ matrix.machine_type.image }}
       matrix_preset: "nightly"
     strategy:
       # A single run not should kill the others
       fail-fast: false
       matrix:
         machine_type:
-          - id: x86
-            instance_name: c6id.metal
+          - id: arm64
+            instance_name: c7gd.metal
+            build_instance_name: c7gd.8xlarge
+            image: ubuntu24-full-arm64-pre-v2

.github/workflows/pr-bench-runner.yml

@@ -28,11 +28,14 @@ permissions:
   id-token: write  # enables AWS-GitHub OIDC
 
 jobs:
+  # Benchmarks run on Graviton3 (c7gd.metal). The binary is built on the same CPU family so
+  # `-C target-cpu=native` matches, and both jobs pin the arm64 image because the
+  # `bench-dedicated` runner defaults to x64.
   build:
     timeout-minutes: 60
     runs-on: >-
       ${{ github.repository == 'vortex-data/vortex'
-          && format('runs-on={0}/runner=bench-dedicated/family=c6id.8xlarge/tag=build-{1}{2}', github.run_id, inputs.benchmark_id, github.event.pull_request.head.repo.fork == false && '/extras=s3-cache' || '')
+          && format('runs-on={0}/runner=bench-dedicated/family=c7gd.8xlarge/image=ubuntu24-full-arm64-pre-v2/tag=build-{1}{2}', github.run_id, inputs.benchmark_id, github.event.pull_request.head.repo.fork == false && '/extras=s3-cache' || '')
           || 'ubuntu-latest' }}
     steps:
       - uses: runs-on/action@v2
@@ -49,8 +52,12 @@ jobs:
           enable-sccache: ${{ github.event.pull_request.head.repo.fork == false && 'true' || 'false' }}
       - name: Build binary
         shell: bash
+        # `-C target-feature=-sve,-sve2`: on Graviton3, `target-cpu=native` (neoverse-v1) makes rustc
+        # 1.98 / LLVM 22 miscompile widening `u8 -> usize` sums at 256-bit SVE, which undersizes
+        # the OnPair and FSST decode buffers for every Vortex file with `u8` string lengths. See
+        # `benchmarks/string-bench/examples/sve_widening_sum.rs`.
         env:
-          RUSTFLAGS: "-C target-cpu=native -C force-frame-pointers=yes"
+          RUSTFLAGS: "-C target-cpu=native -C target-feature=-sve,-sve2 -C force-frame-pointers=yes"
         run: |
           cargo build --package ${{ inputs.benchmark_id }} --profile release_debug \
             --features ${{ inputs.with_lance && 'lance,' || '' }}unstable_encodings
@@ -75,7 +82,7 @@ jobs:
     timeout-minutes: 120
     runs-on: >-
       ${{ github.repository == 'vortex-data/vortex'
-          && format('runs-on={0}/runner=bench-dedicated/family=c6id.metal/tag={1}{2}', github.run_id, inputs.benchmark_id, github.event.pull_request.head.repo.fork == false && '/extras=s3-cache' || '')
+          && format('runs-on={0}/runner=bench-dedicated/family=c7gd.metal/image=ubuntu24-full-arm64-pre-v2/tag={1}{2}', github.run_id, inputs.benchmark_id, github.event.pull_request.head.repo.fork == false && '/extras=s3-cache' || '')
           || 'ubuntu-latest' }}
     steps:
       # `extras=s3-cache` points ACTIONS_RESULTS_URL at the RunsOn artifact proxy, but only

.github/workflows/sql-bench-matrix.yml

@@ -10,10 +10,21 @@ on:
         required: true
         type: string
         description: "Named benchmark matrix to run"
+      # The three machine inputs must agree on CPU architecture: the build job's binaries
+      # run on `machine_type`, and `bench-dedicated` defaults to an x64 image, so an arm64
+      # family needs the arm64 image spelled out.
       machine_type:
         required: false
         type: string
-        default: c6id.metal
+        default: c7gd.metal
+      build_machine_type:
+        required: false
+        type: string
+        default: c7gd.8xlarge
+      machine_image:
+        required: false
+        type: string
+        default: ubuntu24-full-arm64-pre-v2
 
 jobs:
   resolve-matrix:
@@ -50,7 +61,7 @@ jobs:
       FLAT_LAYOUT_INLINE_ARRAY_NODE: "1"
     runs-on: >-
       ${{ github.repository == 'vortex-data/vortex'
-          && format('runs-on={0}/runner=bench-dedicated/family=c6id.8xlarge/tag=build{1}', github.run_id, (inputs.mode != 'pr' || github.event.pull_request.head.repo.fork == false) && '/extras=s3-cache' || '')
+          && format('runs-on={0}/runner=bench-dedicated/family={1}/image={2}/tag=build{3}', github.run_id, inputs.build_machine_type, inputs.machine_image, (inputs.mode != 'pr' || github.event.pull_request.head.repo.fork == false) && '/extras=s3-cache' || '')
           || 'ubuntu-latest' }}
     steps:
       - uses: runs-on/action@v2
@@ -66,8 +77,12 @@ jobs:
           enable-sccache: ${{ (inputs.mode != 'pr' || github.event.pull_request.head.repo.fork == false) && 'true' || 'false' }}
       - name: Build binaries
         shell: bash
+        # `-C target-feature=-sve,-sve2`: on Graviton3, `target-cpu=native` (neoverse-v1) makes rustc
+        # 1.98 / LLVM 22 miscompile widening `u8 -> usize` sums at 256-bit SVE, which undersizes
+        # the OnPair and FSST decode buffers for every Vortex file with `u8` string lengths. See
+        # `benchmarks/string-bench/examples/sve_widening_sum.rs`.
         env:
-          RUSTFLAGS: "-C target-cpu=native -C force-frame-pointers=yes"
+          RUSTFLAGS: "-C target-cpu=native -C target-feature=-sve,-sve2 -C force-frame-pointers=yes"
         run: |
           packages=(--bin data-gen --bin datafusion-bench --bin duckdb-bench)
           if [ "${{ inputs.mode }}" != "pr" ]; then
@@ -112,7 +127,7 @@ jobs:
 
     runs-on: >-
       ${{ github.repository == 'vortex-data/vortex'
-          && format('runs-on={0}/runner=bench-dedicated/family={1}/tag={2}{3}', github.run_id, inputs.machine_type, matrix.id, (inputs.mode != 'pr' || github.event.pull_request.head.repo.fork == false) && '/extras=s3-cache' || '')
+          && format('runs-on={0}/runner=bench-dedicated/family={1}/image={2}/tag={3}{4}', github.run_id, inputs.machine_type, inputs.machine_image, matrix.id, (inputs.mode != 'pr' || github.event.pull_request.head.repo.fork == false) && '/extras=s3-cache' || '')
           || 'ubuntu-latest' }}
     steps:
       - uses: runs-on/action@v2

benchmarks/string-bench/Cargo.toml

@@ -20,6 +20,7 @@ bytes = { workspace = true }
 clap = { workspace = true, features = ["derive"] }
 futures = { workspace = true }
 indicatif = { workspace = true }
+num-traits = { workspace = true }
 parquet = { workspace = true }
 regex = { workspace = true }
 tabled = { workspace = true, features = ["std"] }
@@ -43,5 +44,9 @@ unstable_encodings = [
 name = "string-bench"
 required-features = ["unstable_encodings"]
 
+[[example]]
+name = "arm64_repro"
+required-features = ["unstable_encodings"]
+
 [lints]
 workspace = true