In comparing operf to perf, I am not sure the "minimum counts" that operf is correct. Take the case where you've isolated a process/thread to a single cpu, and isolated the cpu. This thread may be low frequency, but needs to execute fast when woke - essentially attempting to use operf/perf to monitor latency inducing cache misses in the thread.
When I analyze using perf, I can capture the full events triggered by the thread, but with operf, due to the minimum counts, I only capture very few (single digit) events.
It seems that perf understands that the thread under analysis is not generating events at too fast a rate, and thus captures all of them.
Any solution? (I'd like to use operf for its better java JIT handling).
Log in to post a comment.