In debugging recent cyclictest issues on OCP 4.16 (5.14.0-427.22.1.el9_4.x86_64+rt kernel), we have discovered that the "psi=1" kernel cmdline argument, which is now added by default due to cgroupsv2 being enabled, is causing latency issues (both cyclictest and timerlat are failing to meet the latency KPIs we commit to for Telco RAN DU deployments). See RHEL-42737 for reference.

OCP 4.16

Cyclictest and timerlat consistently fail on long duration runs (e.g. 12 hours).

    1. Install OCP 4.16 and configure with the Telco RAN DU reference configuration.
    2. Run a long duration cyclictest or timerlat test    

Maximum latencies are detected above 20us.

All latencies are below 20us.

See RHEL-42737 for test results and debugging information. This was originally suspected to be an RHEL issue, but it turns out that PSI is being enabled by OpenShift code (which adds psi=1 to the kernel cmdline).