analyze-zymtrace-workload

To find GPU or CPU performance bottlenecks using flamegraphs, you must cross-check the opposite side. Pulling synced CPU and GPU flamegraphs for the same workload slice ensures your bottleneck comparison is valid and identifies whether the issue is sync or transfer bound.

vLLM inference workloads showing GPU underutilization often have hidden CPU overhead. By cross-checking the CPU flamegraph against the GPU view using the same resolved filter, you can identify if CPU overhead or hot kernels are stalling the GPU pipeline.

Analyzing a zymtrace workload over a specific time range requires using the zymtrace MCP to identify the workload and fetch flamegraphs. You apply the resolved filter to pull matching CPU and GPU views, then name hot stacks to surface bottleneck patterns.

Yes, diagnosing hot kernels during GPU training jobs requires cross-view analysis. By pulling the opposite-side CPU or GPU flamegraph with the same resolved filter, you can verify if the training bottleneck is computation, memory transfer, or synchronization bound.

The best way to get actionable fixes for workload bottlenecks is generating a recap from cross-view flamegraph data. Naming hot stacks and matching CPU and GPU views grounds the analysis in actual profiling data, surfacing concrete fixes for the top issues.

Workload analysis is unreliable without pulling both CPU and GPU flamegraphs. Looking at only one side misses the real bottleneck, because the performance constraint may be on the opposite side, making enforced cross-view matching essential for valid comparisons.