I run perf record -a -- sleep 3 on the host machine for sampling. There may be binaries with the same path but different buildids in different containers. For example, container A has /tmp/a.out, and container B also has /tmp/a.out. The final sampled perf.data can only output one of the buildids through perf buildid-list.

I run perf record -a -- sleep 3 on the host machine for sampling. There may be binaries with the same path but different buildids in different containers. For example, container A has /tmp/a.out, and container B also has /tmp/a.out. The final sampled perf.data can only output one of the buildids through perf buildid-list.

This is an excerpt of a perf report:

This is an excerpt of a perf report:

This is an excerpt of a perf report:

I want to run the perf record, but filter few pids. Reading the perf record doc I am little confused as to how to mention the pid filter. This does not seems to work:

My goal is to use perf record on the host machine to collect the hot symbols of all processes, including those in the container.
But I found that the symbol column of the process in the container only has the address. I tried to copy the binary in the container to the host machine, keeping the directory consistent, so that the symbols can be seen. So the problem seems to be that the perf tool cannot correctly identify the binary path in the container.
Is there a solution to this problem? Because it seems unrealistic to copy all the process binaries in the container to the host machine.

I’m benchmarking a program, and observing that cycle_activity.stalls_any is significantly higher than cycle_activity.stalls_l1_miss, which to me indicates that the cpu is spending time stalled on resources other than data access. What other resources might those be, and what perf counters correspond to them? In other words, I would expect that stalls_any = stalls_l1_miss + X + Y + .... What are X and Y?