Messages in this thread | | | Date | Tue, 6 Dec 2011 16:07:21 -0500 | From | Vince Weaver <> | Subject | Re: perf_event self-monitoring overhead regression |
| |
On Thu, 1 Dec 2011, stephane eranian wrote:
> I take it your test is all about self-monitoring, single event, single thread. > Did you try breaking down the cost using TSC and rdtsc() to pinpoint where the > regression is coming from in the 3 perf_event syscalls you're using?
I've started gathering results using rdtsc() and the results are even more puzzling.
See: http://web.eecs.utk.edu/~vweaver1/projects/perf-events/benchmarks/rdtsc_overhead/core2_raw_null_kernel_rdtsc.png for example.
Those results are from a core2 machine, with the CPU scaling governor set to "performance", the test bound to CPU0, and the test run 1000 times.
Part of the issue is that a few of those kernels are Debian unstable distro kernels and not hand-compiled stock kernels. I'll rebuild a full set of kernels myself and see if I can reproduce the results.
Vince
| |