Skip to content

Tracing's overhead may be too high for profiling CPU-bound operations #4892

@james7132

Description

@james7132

Bevy version

0.7 and current head (cc4062e)

What you did

Attempted to profile many_foxes with trace_tracy enabled. Commented out spans for system_overhead and system_commands.

What went wrong

Profiled measurements for each stage show significant changes. Here's a comparison of the Update stage measurements:

image

There's only 10-20 systems in that stage, and the overhead scales with the number of systems. This overhead should be much lower, and it makes it difficult to make perf decisions on a macro level.

Metadata

Metadata

Assignees

No one assigned

    Labels

    A-DiagnosticsLogging, crash handling, error reporting and performance analysisC-BugAn unexpected or incorrect behavior

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions