Perf Nsight Systems

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

In this fork there is no perf-profiling-specialist agent (NVIDIA ships one alongside this skill in the TRT-LLM repo). The workflows below describe running nsys directly. For end-to-end inference QPS binary-search work, this skill composes with model-perf-binary-search (see that skill's "Profiling integration" section for how to hand off between the two).

Nsight Systems Profiling

NVIDIA Nsight Systems (nsys) is a system-level performance analysis tool that captures CPU/GPU activity timelines, API traces, and OS-level events. Unlike Nsight Compute (kernel-level), nsys shows the big picture — how kernels, memory transfers, communication, and CPU work overlap in time.

When to Use

Reach for this skill when you encounter:

• Triggers: User wants to profile a training script end-to-end, analyze GPU utilization, find pipeline bottlenecks, check communication/compute overlap, or interpret .nsys-rep reports
• Symptoms: Training slower than expected, GPU idle between iterations, need to understand where time is spent across CPU and GPU, poor scaling in distributed training
• Keywords: "nsys", "nsight systems", "GPU timeline", "GPU utilization", "kernel launch overhead", "training profiling", "NCCL overlap", "nsys-rep", "cuda trace", "GPU idle", "pipeline stall", "data loading bottleneck"

Do NOT use this skill for:

• Kernel-level optimization (use Nsight Compute / ncu instead)
• GPU hardware metrics like SM throughput, cache hit rates (use ncu)
• GPU monitoring without profiling (use nvidia-smi)
• Non-CLI usage (GUI workflows, IDE integration) — consult official docs

Requirements

Permissions: nsys may require sudo or CAP_SYS_ADMIN for system-wide tracing and GPU metrics. In containers, use --privileged or --cap-add=SYS_ADMIN.

Reporting Principles

Every number must have an authoritative source. When presenting timing data, kernel counts, API call durations, or any quantitative metric, always cite the source: nsys stats report output, nsys analyze rule output, exported SQLite query result, recipe CSV, or raw command output. Show the actual command and its output before interpreting. Never synthesize, estimate, or extrapolate numbers that did not come from a tool output.

Use nsys stats for structured analysis, not raw trace data. Always extract metrics via targeted nsys stats -r <report> commands rather than trying to read or interpret .nsys-rep files directly. Stats reports produce compact, tabular summaries; raw trace data can be enormous (especially with backtraces or verbose API tracing). Run the smallest set of reports needed for the task, then request additional reports only if the initial results raise questions.

Workflows

Workflow 1: Profile a DL Training Script

Goal: Capture a clean, focused profile of steady-state training iterations.

Step 1 — Add profiler markers to your training script to skip warmup:

# In training script
for i, batch in enumerate(train_loader):
    if i == warmup_iters:
        torch.cuda.cudart().cudaProfilerStart()
    train_step(model, batch)
    if i == warmup_iters + profile_iters:
        torch.cuda.cudart().cudaProfilerStop()
        break

Step 2 — Profile with cudaProfilerApi capture range:

nsys profile -c cudaProfilerApi \
    -t cuda,nvtx,cudnn,cublas \
    --pytorch=autograd-nvtx \
    -o train_profile -- python train.py

This captures only steady-state iterations — no warmup, no initialization noise.

Note: -t cuda,nvtx,cudnn,cublas enables API-specific tracing. By default, -t cuda only traces the CUDA runtime/driver layer — you see kernel names and launch times but cannot attribute them to higher-level libraries. Adding cudnn and cublas traces the library-level API calls, letting you distinguish convolution time (cuDNN) from GEMM time (cuBLAS) and measure library overhead separately from raw kernel execution.

Step 3 — Quick summary:

nsys stats -r cuda_gpu_kern_sum,cuda_api_sum,cuda_gpu_mem_time_sum \
    train_profile.nsys-rep

When you traced library APIs (cudnn, cublas in -t), also run the library-specific reports to see API-level overhead (workspace allocation, algorithm selection) separately from raw kernel execution:

nsys stats -r cudnn_api_sum,cublas_api_sum train_profile.nsys-rep

Step 4 — Detect anti-patterns:

nsys analyze -r all train_profile.nsys-rep

Step 5 — Dig deeper based on findings. See Tier 2 references.

Workflow 2: Diagnose GPU Idle Time

Goal: Find why the GPU is idle between training iterations.

Step 1 — Profile with OS runtime tracing:

nsys profile -t cuda,nvtx,osrt \
    --pytorch=autograd-nvtx \
    -o idle_debug -- python train.py

Step 2 — Check GPU gaps and utilization:

nsys analyze -r gpu_gaps,gpu_time_util idle_debug.nsys-rep

Step 3 — Check kernel launch phases:

nsys stats -r cuda_kern_exec_sum idle_debug.nsys-rep

High queue time = GPU was busy (not the issue). Near-zero queue time for all kernels = GPU was starved (host not submitting work fast enough).

Step 4 — Common causes and fixes:

Workflow 3: Profile Distributed Training

Goal: Profile multi-GPU/multi-node training with communication analysis.

Step 1 — Collect per-rank profiles:

nsys profile -t cuda,nvtx,mpi,ucx \
    --pytorch=autograd-nvtx \
    -o profile_%q{RANK} \
    -- torchrun --nproc_per_node=8 train.py

Step 2 — Analyze NCCL communication/compute overlap:

nsys recipe nccl_gpu_overlap_trace -- profile_*.nsys-rep
nsys recipe nccl_gpu_time_util_map -- profile_*.nsys-rep

Step 3 — Check per-rank utilization:

nsys recipe cuda_gpu_time_util_map -- profile_*.nsys-rep

Step 4 — Check for stragglers:

Compare cuda_gpu_kern_sum across ranks. If one rank is slower, check its network and data loading patterns.

Workflow 4: Analyze Iteration Time Consistency

Goal: Check whether training iterations are stable or have outliers.

# Profile with NVTX iteration markers
nsys profile --pytorch=autograd-nvtx -t cuda,nvtx \
    -o iter_check -- python train.py

# Check iteration timing distribution
nsys stats -r nvtx_pushpop_sum iter_check.nsys-rep

# Check GPU projection per NVTX range
nsys stats -r nvtx_gpu_proj_sum iter_check.nsys-rep

# Visual pace analysis
nsys recipe nvtx_pace -- iter_check.nsys-rep

High StdDev in iteration duration indicates inconsistency — investigate outlier iterations on the timeline.

Workflow 5: Attribute Kernels to Source Code via Stack Traces

Goal: Identify which Python function or code path triggers expensive GPU kernels.

Step 1 — Profile with backtrace collection:

nsys profile -t cuda,nvtx \
    --backtrace=cuda \
    --python-backtrace=lbr \
    --pytorch=autograd-nvtx \
    -o stacktrace_profile -- python train.py

• --backtrace=cuda: Captures CUDA API call stacks (C/C++ frames) so each cudaLaunchKernel shows the host-side call chain that triggered it.
• --python-backtrace=lbr: Captures Python-level call stacks, correlating GPU work back to specific Python functions (e.g., compute_attention vs compute_ffn).

Step 2 — Get kernel summary and NVTX attribution:

Use targeted stats reports to identify top kernels and their NVTX context:

# Top kernels by total GPU time
nsys stats -r cuda_gpu_kern_sum stacktrace_profile.nsys-rep

# Kernels attributed to NVTX ranges (maps kernels to annotated code regions)
nsys stats -r nvtx_kern_sum stacktrace_profile.nsys-rep

The nvtx_kern_sum report (requires --pytorch=autograd-nvtx or manual NVTX annotations) maps each kernel to its enclosing NVTX range, directly showing which Python function or autograd op launched it. This is more efficient than manually cross-referencing raw backtrace data.

Step 3 — For PyTorch models, --pytorch=autograd-nvtx automatically wraps each autograd op in an NVTX range. Combined with backtrace, this maps: GPU kernel → CUDA API call → Python function → PyTorch autograd op.

When to use: Workloads with multiple code paths launching similar kernels (e.g., attention vs FFN both calling GEMM). Stack traces disambiguate which caller is responsible for the dominant kernel time.

Output Formats

Report files (.nsys-rep): Binary format, viewable in GUI or processed with nsys stats, nsys analyze, nsys export, nsys recipe.

Stats output formats: column (terminal), csv, json, table, tsv, hdoc, htable.

Export formats: sqlite (SQL queries), arrow/parquetdir (Pandas/Dask), hdf, jsonlines, text.

Recipe output: Directory with CSV/Parquet data + Plotly HTML visualizations

• .nsys-analysis (Jupyter notebook).

Key stats report columns:

Examples

Example 1: Quick DL Profile and Summary

# Profile
nsys profile -t cuda,nvtx,cudnn,cublas \
    --pytorch=autograd-nvtx --stats=true \
    -o quick_profile -- python train.py

# Auto-generates stats at the end of profiling

Example 2: Detect Sync Memcpy in DataLoader

nsys profile -t cuda,nvtx -o dataloader_check -- python train.py
nsys analyze -r cuda_memcpy_sync,cuda_memcpy_async dataloader_check.nsys-rep

If flagged, fix with:

loader = DataLoader(dataset, pin_memory=True, num_workers=4)
tensor_gpu = tensor_cpu.to(device, non_blocking=True)

Example 3: Multi-Node NCCL Analysis

# Collect
nsys profile -t cuda,nvtx,mpi -o rank_%q{RANK} \
    -- torchrun --nproc_per_node=8 train.py

# Analyze overlap
nsys recipe nccl_gpu_overlap_trace -- rank_*.nsys-rep

# Visualize
nsys recipe nccl_gpu_time_util_map -- rank_*.nsys-rep

Example 4: API-Level Breakdown (cuDNN vs cuBLAS)

# Profile with library-level tracing
nsys profile -t cuda,nvtx,cudnn,cublas \
    -o api_breakdown -- python model.py

# cuDNN API summary (convolution calls)
nsys stats -r cudnn_api_sum api_breakdown.nsys-rep

# cuBLAS API summary (GEMM calls)
nsys stats -r cublas_api_sum api_breakdown.nsys-rep

# Compare with kernel-level view
nsys stats -r cuda_gpu_kern_sum api_breakdown.nsys-rep

The API-level reports (cudnn_api_sum, cublas_api_sum) show time spent in library calls including overhead (workspace allocation, algorithm selection), while cuda_gpu_kern_sum shows only raw GPU kernel execution. The difference reveals library-side overhead.

Error Handling

Finding More Information

Tier 1: This File (SKILL.md)

You are reading it now. The workflows and error table above cover the most common DL profiling tasks. Search this file first.

Tier 2: references/ Directory

Grep for keywords across references/ — headers are grep-friendly:

• references/cli-profiling.md — Complete nsys profile flags for DL
• references/cli-post-collection.md — nsys stats, analyze, export, recipe commands
• references/app-preparation.md — Focused profiling, NVTX markers, PyTorch patterns
• references/stats-reports.md — CUDA statistical report columns and meanings
• references/expert-systems.md — Expert system rules, anti-pattern detection
• references/recipes-dl.md — DL-relevant advanced recipes with examples
• references/nvtx-analysis.md — NVTX statistical reports for annotated code

How to search:

1. Grep for your keyword across references/
2. Read only the file that Grep points to

Tier 3: Official Documentation

If Tiers 1-2 don't answer:

• User Guide — Full CLI reference, all tracing options
• Analysis Guide — Stats reports, expert systems, recipes

WebFetch or WebSearch these URLs for the latest content. Consider distilling new findings back into references/.