--- name: perf-nsight-systems-dfb23913b3ec description: >- Nsight Systems (nsys) CLI for system-level timeline profiling. Use when the user wants to run nsys profile, analyze .nsys-rep reports, use nsys stats/analyze/recipe commands, diagnose GPU idle time from timeline traces, or profile distributed training with NCCL overlap analysis. NOT for kernel-level metrics like SOL%, occupancy, or roofline (use perf-nsight-compute-analysis for ncu). NOT for writing or generating kernels. NOT for applying optimizations like CUDA Graphs. license: Apache-2.0 metadata: author: NVIDIA Corporation --- # 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 | Dependency | Version | Notes | |------------|---------|-------| | CUDA Toolkit | >=11.0 | Includes `nsys` | | `nsys` binary | Match CUDA version | Verify with `nsys -v` | | NVIDIA GPU | Any supported | | 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 ` 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: ```python # 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:** ```bash 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:** ```bash 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: ```bash nsys stats -r cudnn_api_sum,cublas_api_sum train_profile.nsys-rep ``` **Step 4 — Detect anti-patterns:** ```bash 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:** ```bash nsys profile -t cuda,nvtx,osrt \ --pytorch=autograd-nvtx \ -o idle_debug -- python train.py ``` **Step 2 — Check GPU gaps and utilization:** ```bash nsys analyze -r gpu_gaps,gpu_time_util idle_debug.nsys-rep ``` **Step 3 — Check kernel launch phases:** ```bash 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:** | GPU idle cause | Evidence | Fix | |---------------|----------|-----| | Slow data loading | CPU busy in DataLoader during gaps | Increase `num_workers`, use `pin_memory=True` | | Synchronous memcpy | `cuda_memcpy_sync` rule fires | Use `non_blocking=True` transfers | | Over-synchronization | Frequent `cudaDeviceSynchronize` in trace | Remove unnecessary sync calls | | Host-side computation | CPU sampling shows compute during gaps | Move to GPU or overlap with async ops | | Python GIL contention | GIL trace shows contention | Use multiprocessing, reduce Python overhead | ### Workflow 3: Profile Distributed Training Goal: Profile multi-GPU/multi-node training with communication analysis. **Step 1 — Collect per-rank profiles:** ```bash 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:** ```bash 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:** ```bash 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. ```bash # 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:** ```bash 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: ```bash # 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:** | Report | Key columns | |--------|------------| | `cuda_gpu_kern_sum` | Time%, Total Time, Instances, Kernel Name | | `cuda_api_sum` | Time%, Total Time, Num Calls, API Name | | `cuda_kern_exec_sum` | API Time, Queue Time, Kernel Time | | `cuda_gpu_mem_time_sum` | Time%, Total Time, Operations, Direction | | `nvtx_gpu_proj_sum` | Projected Duration, Original Duration, GPU Op Count | ## Examples ### Example 1: Quick DL Profile and Summary ```bash # 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 ```bash 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: ```python loader = DataLoader(dataset, pin_memory=True, num_workers=4) tensor_gpu = tensor_cpu.to(device, non_blocking=True) ``` ### Example 3: Multi-Node NCCL Analysis ```bash # 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) ```bash # 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 | Error | Cause | Fix | |-------|-------|-----| | `nsys: command not found` | Not in PATH | `export PATH=$PATH:/usr/local/cuda/bin` | | `Permission denied` or `requires root` | Needs elevated privileges | `sudo nsys ...` or `--cap-add=SYS_ADMIN` in containers | | No CUDA activity captured | App didn't use GPU during collection window | Adjust `--delay`/`--duration`, or use `cudaProfilerApi` capture range | | Report file very large | Long profile with many APIs traced | Use focused capture (`-c cudaProfilerApi`), reduce `--duration` | | `--pytorch` has no effect | Wrong nsys version or Python env | Verify nsys version supports `--pytorch`; check Python is in PATH | | `nsys stats` shows empty reports | No matching activity in report | Check `--trace` flags included the right APIs | | MPI rank profiles out of sync | Clock skew between nodes | Use NTP sync; analyze per-rank independently | | `cudaProfilerStart` not captured | Missing `-c cudaProfilerApi` flag | Add `--capture-range=cudaProfilerApi` | | Recipe fails with import error | Missing Python dependencies | Install recipe dependencies: `pip install pandas plotly` | ## 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](https://docs.nvidia.com/nsight-systems/UserGuide/index.html) — Full CLI reference, all tracing options - [Analysis Guide](https://docs.nvidia.com/nsight-systems/AnalysisGuide/index.html) — Stats reports, expert systems, recipes WebFetch or WebSearch these URLs for the latest content. Consider distilling new findings back into `references/`. --- Source attribution: NVIDIA Corporation Original project: https://github.com/yo-steven/skills-exploration-20260522/blob/6e1127ffb7690aa8038b172d0ae7cc720754870d/skills/TensorRT-LLM/perf-nsight-systems/SKILL.md License: Apache-2.0 License information: https://www.apache.org/licenses/LICENSE-2.0 Original upstream instructions; downloaded from Mega AI Hub. ## Upstream license notice SPDX-License-Identifier: Apache-2.0 AND CC-BY-4.0 Copyright (c) 2026 NVIDIA Corporation This project is licensed under a dual license: - Apache License 2.0 (for code and configuration) - Creative Commons Attribution 4.0 International (CC BY 4.0) (for documentation) You may use the contents of this repository under the terms of either license, or both, as applicable to the type of content. ================================================================================ Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 1. Definitions. "License" shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document. "Licensor" shall mean the copyright owner or entity authorized by the copyright owner that is granting the License. "Legal Entity" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, "control" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity. "You" (or "Your") shall mean an individual or Legal Entity exercising permissions granted by this License. "Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files. "Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types. "Work" shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below). "Derivative Works" shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof. "Contribution" shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to the Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution." "Contributor" shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by the Licensor and subsequently incorporated within the Work. 2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form. 3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed. 4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions: (a) You must give any other recipients of the Work or Derivative Works a copy of this License; and (b) You must cause any modified files to carry prominent notices stating that You changed the files; and (c) You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and (d) If the Work includes a "NOTICE" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding any notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License. You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License. 5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions. 6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file. 7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License. 8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages. 9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability. END OF TERMS AND CONDITIONS ================================================================================ Creative Commons Attribution 4.0 International License (CC BY 4.0) https://creativecommons.org/licenses/by/4.0/ You are free to: Share — copy and redistribute the material in any medium or format for any purpose, even commercially. Adapt — remix, transform, and build upon the material for any purpose, even commercially. The licensor cannot revoke these freedoms as long as you follow the license terms. Under the following terms: Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use. No additional restrictions — You may not apply legal terms or technological measures that legally restrict others from doing anything the license permits. Notices: You do not have to comply with the license for elements of the material in the public domain or where your use is permitted by an applicable exception or limitation. No warranties are given. The license may not give you all of the permissions necessary for your intended use. For example, other rights such as publicity, privacy, or moral rights may limit how you use the material. For the full legal text, see: https://creativecommons.org/licenses/by/4.0/legalcode # Mega AI Hub resource metadata Creator / publisher: NVIDIA Corporation Original project: https://github.com/yo-steven/skills-exploration-20260522/blob/6e1127ffb7690aa8038b172d0ae7cc720754870d/skills/TensorRT-LLM/perf-nsight-systems/SKILL.md Attribution basis: Source repository record Fields of work: Development & Engineering, Data & Analytics Roles: ML / AI Engineer, Data Scientist, Backend Developer Specialties: Backend Tools: Nvidia Nsight Systems Country applicability: Country not specified Classification basis: Source category, role and specialization labels Instruction SHA-256: 852b54a24d8918f0cf8ee8a6804027aac084d2b7b6182e891eb24775232d92c0