Skip to content

Add a performance-diagnosis layer for better user usage #50

Description

@minghangli-uni

The Trace Explorer added in #48 makes individual PET timelines easier to inspect but users still have to interpret those timelines manually. Currently looking at representative ranks such as PET0 and the 1st ocean PET is useful for forming a hypothesis but it cannot establish whole component balance or identify the rank that actually finishes fast.

We now need a performance-diagnosis layer that consumes all parsed PET spans and reduces them to compact, defensible component and synchronization metrics.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions