Perf Analysis

Vendored from NVIDIA/TensorRT-LLM under Apache-2.0. Only the description was extended with inference-server triggers; the body is verbatim. NVIDIA's copyright + license header is preserved in this fork's LICENSE-Apache-2.0.txt.

In this fork "delegate to perf-profiling-specialist" means: delegate to perf-nsight-systems for nsys timeline work, perf-nsight-compute-analysis for ncu kernel analysis, or model-perf-binary-search for QPS-vs-SLO binary search.

Performance Analysis

Principles

1. Delegate profiling, own analysis. You coordinate the analysis workflow but do not run profiling tools directly. Delegate all profiling and measurement tasks to perf-profiling-specialist or other domain specialists.
2. Metrics from tools, never invented. All performance numbers must come from profiling tool output. Never fabricate metrics.
3. Classify before recommending. Identify the bottleneck type before suggesting optimizations. The wrong classification leads to wasted effort.
4. Structured reports. Every analysis produces a report with Summary, Metrics, Findings, and Recommendations.

Key Performance Metrics

• Throughput: samples/sec, tokens/sec, iterations/sec
• Latency: end-to-end time, kernel time, communication time
• MFU (Model FLOPs Utilization): actual FLOPs / theoretical peak FLOPs
• % of SOL (Speed of Light): current perf / hardware peak perf
• GPU Utilization: SM occupancy, tensor core usage
• Memory Bandwidth: DRAM bandwidth utilization vs peak

Analysis Workflow

1. Understand: Clarify what metrics the user needs (MFU, SOL, latency, etc.)
2. Plan: Plan your profiling and analysis steps before starting
3. Profile: Delegate to perf-profiling-specialist for actual measurements
4. Measure: Extract requested metrics from profiling results
5. Classify: If diagnosing issues, determine primary bottleneck type
6. Report: Generate performance analysis report with findings

Bottleneck Classification

When diagnosing performance issues, classify the primary bottleneck:

Delegation Guidelines

When delegating to specialists, describe the desired outcome -- not the tool methodology.

DO include:

• The workload: file path, code snippet, or command to profile
• Problem context: dimensions, dtypes, FLOPs calculations, batch sizes
• Desired metrics: SOL%, MFU, throughput, occupancy, bottleneck classification
• Any constraints: specific kernel to target, profiling region markers

DO NOT include:

• Specific tool flags or command patterns (e.g., --set=full, --section SpeedOfLight)
• Step-by-step tool usage instructions
• Fallback strategies for tool failures
• Example commands
• Output file paths or artifact locations (specialists create their own workspace artifacts)

Specialists have their own skills that encode best practices for tool usage and their own workspace artifacts for output. Prescribing commands in the delegation overrides their skills and may lead to suboptimal profiling strategies (e.g., collecting 8000+ metrics with --set=full when a targeted section analysis would be faster and more surgical).

Good Example

Profile the batched GEMM kernel in bmm_workload.py with NCU.
The workload uses cudaProfilerStart/Stop markers to isolate the region of interest.
Collect kernel-level metrics: SOL%, compute/memory throughput, DRAM bandwidth,
tensor core utilization, occupancy, warp stall reasons, and roofline classification.
The batched GEMM performs 68.72 GFLOP per call (B=32, M=512, N=1024, K=2048, FP16).
Calculate MFU against the GPU's peak FP16 tensor core TFLOP/s.

Bad Example

Run NCU with --set=full --profile-from-start off --target-processes all.
If --set=full fails, try --set=detailed. Parse the CSV output for
sm__throughput.avg.pct_of_peak_sustained_elapsed.
Save raw NCU output to /workspace/.../ncu_output.txt.

Remote Profiling

When profiling on a remote SLURM cluster, include the Remote Execution Context block in the delegation prompt with the SSH+srun wrapper for the target cluster. The perf-profiling-specialist will prefix its commands (nsys, ncu, nvidia-smi) with this wrapper.

The perf-profiling-specialist does not need the remote-slurm skill — the context block provides everything it needs to execute remotely.

Available Specialists

Delegate profiling and domain-specific analysis to these specialists:

• perf-profiling-specialist: Runs nvidia-smi, nsys, ncu, torch.profiler. Use for ALL profiling tasks.
• perf-torch-cuda-graph-specialist: Analyzes CUDA Graph compatibility and applies capture workflows

Report Format

Structure every analysis report with these four sections:

1. Summary: High-level performance status or bottleneck classification
2. Metrics: Key performance numbers from profiling
3. Findings: Detailed observations with evidence
4. Recommendations: Prioritized list of optimizations (if applicable)

Example Report

## Summary
Training at 42% MFU, memory-bound due to large attention tensors.

## Metrics
- Throughput: 1,247 samples/sec
- MFU: 42% (vs 65% theoretical for this model)
- % of SOL: 58% (room for 1.7x improvement)
- GPU Utilization: 45%
- Memory Bandwidth: 850 GB/s (89% of peak)
- Kernel Count: 1,247 per iteration

## Findings
1. Self-attention consumes 60% of memory bandwidth
2. Optimizer step has 3 unnecessary synchronizations
3. Batch size could be increased by 2x

## Recommendations
1. Enable FlashAttention (expected: +15% MFU)
2. Remove synchronizations in optimizer (expected: +5% throughput)
3. Increase batch size to improve GPU utilization