Normal Description of problem:
hardware advertises NNPA and Triton uses Telum zDNN but the PAI PMU is not present on the host is missing /sys/bus/event_source/devices/pai_ext so perf does not register nnpa events and fails with bad event name
Version-Release number of selected component (if applicable):
ocp 4.19.4 Worker kernel: 5.14.0-570.26.1.el9_6.s390x OS image: RHCOS 9 (node shows 9.6; debug pod base is RHEL 9.4) CRI: cri-o 1.32.6 z/OS zCX perf: 5.14.0-427.76.1.el9_4.s390x cpu flags include nnpa
How reproducible:
always
Steps to Reproduce:
1. Deploy Triton Inference Server using an ONNX-MLIR/zDNN model on s390x workers and send inference 2. oc debug node/<worker>; nsenter -a -t 1 3. run ls -ld /sys/bus/event_source/devices/pai_ext
perf list | grep -E '(^| )pai_ext/|^ *NNPA_'
perf stat -a -e NNPA_ALL sleep 10
Actual results:
- /sys/bus/event_source/devices/pai_ext is missing - perf list shows no NNPA events - perf ... -e NNPA_* → Bad event name
Expected results:
- pai_ext device exists on host - perf list shows NNPA_* events (e.g., NNPA_ALL, NNPA_MATMUL_OP, NNPA_ADD) - perf -e NNPA_* returns nonzero counts under NNPA load
Additional info:
