Skip to content

[BUG] MTL: multiple-pause-resume.sh takes 10 times longer with _NOCODEC #8174

Closed

Description

@marc-hb

MTLP_RVP_NOCODEC takes TEN times longer (!) to run multiple-pause-resume-50.sh than MTLP_RVP_SDW or MTLP_RVP_HDA

All tests seem slower on _NOCODEC but multiple-pause-resume-50.sh is especially slower.

This is "bottlenecking" and slowing down the entire validation.

Examples:

https://sof-ci.01.org/linuxpr/PR4577/build6527/devicetest/index.html
multiple-pause-resume-50.sh
MTLP_RVP_NOCODEC: 30 minutes
MTLP_RVP_SDW, MTLP_RVP_HDA: 3 minutes

Same in https://sof-ci.01.org/sofpr/PR8161/build12679/devicetest/index.html
and in every mtl_pr test run I've looked at in the last few days.

Activity

  1. added
    bugSomething isn't working as expected
    MTLApplies to Meteor Lake platform
    on Sep 7, 2023
  2. singalsu commented on Sep 7, 2023

    @singalsu
    Collaborator

    So average 36s vs. 3.6s per suspend-resume cycle, that's a big difference!

  3. marc-hb commented on Sep 7, 2023

    @marc-hb
    CollaboratorAuthor

    Thanks @plbossart :

    The answer is in the topology: there are just more PCM devices in the
    nocodec topology and the script blindly tests them all. In theory the
    test duration should be a linear factor of the number of PCM devices,
    but that could be impacted by firmware transitions since some PCM
    devices go into mixers and pipelines are activated with a refcount.

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

Metadata

Metadata

Assignees

No one assigned

    Labels

    MTLApplies to Meteor Lake platformbugSomething isn't working as expected

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions