HolisticTraceAnalysis

What it is and what it does

Holistic Trace Analysis is a performance profiling tool built to analyze PyTorch Profiler traces from distributed training workloads. It ingests trace files collected via PyTorch's Kineto profiler and provides structured breakdowns of GPU activity—computation time, communication time, memory events, and idle periods—across all training ranks. The package runs as a Python library, typically within Jupyter notebooks, and exposes an API for querying temporal breakdowns, kernel statistics, communication-computation overlap, memory bandwidth utilization, and queue lengths.

The tool helps identify where training time is spent and why GPUs are idle, which is essential for optimizing large-scale distributed training. It includes trace comparison capabilities to visualize differences between runs, CUDA kernel pattern analysis to find frequently launched operations, and experimental GPU performance counter analysis. Dependencies include numpy, pandas, networkx for data processing and plotly for visualization.

Use it for:

• Identify GPU idle time causes and communication bottlenecks in multi-GPU or multi-node training runs.
• Compare performance profiles across different training configurations or hardware setups to validate optimization changes.
• Analyze kernel launch overhead and memory bandwidth utilization to guide kernel-level optimization efforts.
• Investigate temporal patterns of computation and communication overlap to improve distributed training efficiency.
• Diagnose performance regressions by comparing trace files from different training runs side-by-side.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Analyzes PyTorch Profiler traces to identify performance bottlenecks in distributed training workloads by breaking down GPU time, kernel execution, communication overlap, and idle periods.

Yes. The package is actively maintained, has no known vulnerabilities, installs with low friction, and solves a specific and important problem for anyone optimizing distributed PyTorch training. The MIT license imposes no restrictions. Install it if you profile distributed training workloads and need structured analysis of PyTorch Profiler traces.

Install

holistictraceanalysis on PyPI

pip install holistictraceanalysis

uv add holistictraceanalysis

poetry add holistictraceanalysis

Installing HolisticTraceAnalysis

Before you install

Low friction installation as a pure Python wheel. Active maintenance with recent commits; last release 442 days ago. Requires Python 3.8 or later and depends on common data science libraries (numpy, pandas, networkx, plotly) plus jupyterlab and pytest.

License in practice

MIT license is permissive; you can use, modify, and distribute this package freely with minimal restrictions, making it suitable for both open and proprietary projects.

Quickstart

pip install HolisticTraceAnalysis

from holistictraceanalysis.trace_analysis import TraceAnalysis
analyzer = TraceAnalysis(trace_dir="/path/to/traces")
temporal_breakdown = analyzer.get_temporal_breakdown()
kernel_breakdown = analyzer.get_gpu_kernel_breakdown()

Requires PyTorch Profiler traces collected from a distributed training job; traces must reside in a single folder.

Verify before relying

• Whether the package works with PyTorch versions beyond those explicitly tested in CI
• Performance characteristics when analyzing very large trace files or high-rank distributed jobs
• Compatibility with GPU types beyond NVIDIA CUDA (e.g., AMD ROCm, Intel Arc)
• Exact import path structure for the TraceAnalysis API

Package facts

Evidence: holistictraceanalysis-0.5.0-py3-none-any.whl

Tags

More Artificial Intelligence packages

Further reading