Task Manager

Task Manager is a Fabric client profiler for modded Minecraft that helps you identify frame drops, high MSPT, hot mods, rendering spikes, block-entity pressure, and JVM/runtime issues while the game is still running.

Open the profiler in-game with F12. Toggle the HUD with F10 and export Sessions using F9.

Key Features

• Live per-mod CPU attribution with sampled stack ownership and event invokes shown separately
• Estimated per-mod GPU attribution using render-thread sampling plus measured GPU timer-query time
• Per-mod memory estimates, shared JVM/runtime buckets, and top live class families
• Loader-aware startup timing measured from Fabric entrypoint execution and shown explicitly in ms
• Timeline graphs for frametime, FPS, network throughput, and disk throughput
• World diagnostics with Lag Map and Block Entities mini-tabs
• JVM thread diagnostics, waits/locks, and sensor/runtime health in the System tab
• Session recording and export with summarized worst frame, MSPT spikes, top CPU/GPU/memory mods, hot chunk, and block-entity hotspots
• Optional HUD overlay with configurable sections, layout, position, and trigger mode

Tabs

• Tasks: per-mod CPU share, sampled threads, sample counts, invokes, filtering, sorting, and row detail
• GPU: estimated per-mod GPU share, render samples, thread counts, filtering, sorting, and row detail
• Render: render-phase CPU/GPU timings and call counts
• Startup: Fabric startup/entrypoint timing by mod in explicit milliseconds
• Memory: per-mod heap estimates, shared JVM buckets, top classes, filtering, sorting, and shared-family drilldown
• Flame: sampled stack/flame-style hotspot view
• Timeline: frametime, FPS, jitter, variance, frame percentiles, GPU frame span (a GL timestamp span of the render call, not GPU busy time), and MSPT overview
• Network: inbound/outbound throughput graphs plus packet/category breakdowns
• Disk: disk read/write throughput graphs
• World: lag map, hot chunks, entity hotspots, and block-entity drilldown
• System: CPU/GPU load, temperatures, thread load, waits/locks, sensors, and runtime diagnostics
• Settings: session options, HUD options, and configurable table column visibility

Named Iris GPU passes

In Render, select Passes (V) and turn Collect ON (C). Collection is independent of the normal capture mode and continues after closing TaskManager; turn it off explicitly when finished. S reverses frame-cost sorting. Phases retains the existing broad CPU/GPU view.

Named composite, shadow-composite (including compute-only), and final shader intervals use nonblocking GPU timestamps. Pass names/stages are Iris runtime metadata, not inferred source filenames. These measured rows are separate from estimated per-mod GPU attribution. Pass time is inclusive; never add it to its parent pipeline interval or to future nested compute breakdowns.

MS/FRAME sums repeated invocations in each fully completed executing frame, then averages those frame sums. MS/CALL divides by completed GPU invocations, not CPU calls. Pending/failed frames do not become zero-cost frames. Sample counts and tooltips distinguish nonexecution in an observed stage from unavailable measurement; unobserved hooks are unsupported/unconfirmed, not zero milliseconds.

The normal Export action (F9) also writes taskmanager-shader-passes-<timestamp>.csv and .metadata.json in taskmanager-sessions. The CSV contains raw completed inclusive intervals with session,frame,dimension,stage,pass,invocation,gpu_ns,status. Its sidecar records versions, shader pack/resolved settings, framebuffer resolution, render distance, GPU, frame coverage, status counts and retention limits. A capture retains at most 600 recent frames / 32,768 pass invocations; oldest whole frames are evicted. The timestamp pool admits at most 2,048 outstanding intervals and counts drops. Toggling capture, reloading/replacing a pipeline, or changing resolution/render distance starts a new session and cancels old pending queries.

For comparisons, use the same stationary scene, camera, pack settings, resolution and distance. First measure profiling OFF versus ON to establish overhead; these hooks alone do not establish timing accuracy or an optimization speedup.

Controls

• F12: open Task Manager
• F9: start/stop session recording
• F10: toggle HUD overlay

All keybinds can also be changed in Minecraft controls.

Searching, Sorting, and Columns

Tasks, GPU, and Memory support:

• typed search/filter boxes
• clickable sortable columns
• configurable column visibility from the Settings tab

Sessions and Export

Session exports include a polished summary with:

• worst frame
• worst MSPT spike
• top CPU mods
• top GPU mods
• top memory mods
• hot chunk summary
• block entity classes
• repeated conflict edges and pairwise lock contention summaries
• rule findings and sensor diagnostics

For MANUAL_DEEP, press F9 to start recording, play normally while reproducing the issue, then press F9 again to stop. The session will continue recording even if the Task Manager screen is closed.

Notes on Accuracy

Task Manager aims to be honest about what is measured versus estimated:

• CPU attribution is based on sampled stack ownership
• GPU attribution is estimated from render-thread samples weighted by GPU timer-query time
• Memory attribution is based on JVM/class-histogram style ownership and shared runtime buckets
• Per-mod heap histograms are on demand: use Refresh Heap in Memory. This can pause the game; ordinary JVM memory metrics still update automatically. Per-mod rows and exported attribution retain the last requested heap sample until refreshed.
• Session summary points follow the configured metrics interval (500 ms by default); frame/spike tracking and raw GPU pass samples keep their own cadence.
• Windows telemetry helpers stop when live collection is inactive. Session recording, background capture modes, or an enabled HUD still intentionally collect.
• Temperatures and some system metrics depend on what the platform or an external sensor provider exposes
• For CPU temperature on Windows, running Core Temp is the primary recommended setup; LibreHardwareMonitor/OpenHardwareMonitor and HWiNFO can also work

That means the profiler is very useful for finding hotspots and spikes, but some values, especially GPU-per-mod and shared JVM memory, should be interpreted as guidance rather than perfect ground truth.