Issues with ocperf and toplev about pmu-tools HOT 3 CLOSED

The graphs are not required to add up to 100%, there may be no 100%
bottleneck sum.

Typical reasons for not counted are:

• the event truly does not happen in your workload (that is not
  uncommon)
• due to multiplexing errors. The CPU can normally only measure 4
  counters at a time. Multiplexing assumes the workload
  is "steady state" (does mostly the same) and runs reasonably long.
  In general you get less error when you with lower levels, as
  those need less groups to measure (-l1 does not multiplex at all)
• You can play around with the setting in
  /sys/devices/cpu/perf_event_mux_interval_ms
  (note set it to 0 again when you're finished)
• Older kernels had various scheduling counter scheduling bugs.
• If your workload is very reproducible it's possible to disable
  multiplexing

In your graph it's a bit odd that the top level does not show the
lower levels. Do you see the same with just -l1?

I think your kernel may not correctly schedule one of the L1 events
(unless Ubuntu backported the full Haswell PMU code)
In this case a kernel update would help.

from pmu-tools.

• Also if the workload is thread intensive it is currently better to disable hyper threading in the BIOS.
  The full SMT support for Haswell is still WIP.

from pmu-tools.

There were a lot of fixes recently on issues that may have caused non counted.
Also one kernel bug was found on Haswell (see https://github.com/andikleen/pmu-tools/wiki/toplev-kernel-support)

I would expect the issues to be fixed with all of these.

from pmu-tools.