Skip to content

Latest commit

 

History

History
108 lines (70 loc) · 3.97 KB

profiling.md

File metadata and controls

108 lines (70 loc) · 3.97 KB

Profiling Guide

This guide covers how to capture, analyze, and interpret profiling data from AORTA benchmark runs.

Profiling Outputs

Each rank writes artifacts/rank_<rank>_metrics.jsonl containing iteration-level telemetry:

  • Per-stream durations (ms) for compute, all-reduce, reduce-scatter, and auxiliary streams
  • Overlap segments with concurrency statistics and utilisation ratios
  • ROCm diagnostic output when enabled (rank-local)
  • Loss, learning rate, gradient norms, and global step counters

Events are captured using torch.cuda.Event(enable_timing=True) for microsecond fidelity. Distributed collectives are monkey-patched at runtime so that all-reduce and reduce-scatter operations execute on dedicated streams and contribute to overlap calculation.

Torch Profiler Traces

Enable PyTorch's profiler by toggling the profiling block in your config or via CLI override:

torchrun --nproc_per_node 4 train.py \
  --config config/default.yaml \
  --override profiling.enabled=true \
  --override profiling.wait=1 \
  --override profiling.warmup=1 \
  --override profiling.active=2

Output Locations

  • TensorBoard traces write to artifacts/torch_profiler/rank*/ by default
  • Launch tensorboard --logdir artifacts/torch_profiler and use the Profile tab for stream timelines

Chrome Traces

  • Enable via profiling.chrome_trace=true
  • Not recommended on ROCm - the toolkit disables them automatically to avoid known Kineto crashes
  • Adjust wait, warmup, active, and repeat to control capture cadence
  • Shapes and memory statistics are recorded by default

ROCm rocprofv3 Capture

Use the wrapper script to profile an entire ROCm run:

bash scripts/rocprof_capture.sh config/default.yaml --override training.max_steps=50

Output Location

Outputs land under rocprof_traces/run_<timestamp>/.

Environment Variables

Override location or extra flags with environment variables:

  • ROCPROF_OUTPUT_DIR=/path/to/out
  • ROCPROF_ARGS="--att --kernel-trace --kernel-symbols"

The script mirrors launch_rocm.sh but executes through rocprofv3, so you can merge traces with the JSONL metrics using the shared iteration timestamps.

Generating Reports

Run the analyser to build summaries and plots from one or more log directories:

python analysis/overlap_report.py \
  --log-dir artifacts_rocm --label rocm \
  --log-dir artifacts_cuda --label cuda \
  --output reports/2024-roc-vs-cuda \
  --reference cuda --candidate rocm

Report Outputs

  • summary.json - Aggregate metrics per dataset plus comparative ratios
  • {label}_timeline.png - Overlays showing compute and overlap durations per global step

Use these artefacts to pinpoint scheduling or synchronisation regressions between hardware backends.

Diagnostic Insights

Overlap Breakdown

Overlap Ratio

Key Metrics

Metric Interpretation
Overlap Ratio (overlap_ratio) Values close to 1 indicate strong overlap; values near 0 imply communications block compute
Compute All-Reduce (compute_allreduce_ms) Time spent in all-reduce operations
Compute Reduce-Scatter (compute_reducescatter_ms) Time spent in reduce-scatter operations

Analysis Tips

  • Compare compute_allreduce_ms vs compute_reducescatter_ms to determine which collective dominates stall time
  • Inspect active_segments in the JSONL logs to align iteration windows with external profilers (e.g., ROCm tracer)
  • Cross-reference rocm_smi_output against overlap dips to correlate DVFS throttling or memory pressure with scheduling gaps

Advanced Profiling

For deeper inspection, combine these scripts with nsys, rocprof, or PyTorch profiler traces using the iteration timestamps documented in the JSON traces.

Next Steps