Hi, On 24.09.2018 21:32, Alexey Budankov wrote: > Hi, > > On 24.09.2018 17:29, Jiri Olsa wrote: >> On Mon, Sep 24, 2018 at 04:09:09PM +0300, Alexey Budankov wrote: >>> Hi, >>> >>> On 24.09.2018 10:02, Alexey Budankov wrote: >>>> Hi, >>>> >>>> On 23.09.2018 22:30, Jiri Olsa wrote: >>>>> On Fri, Sep 21, 2018 at 09:13:08AM +0300, Alexey Budankov wrote: >>>>> >>>>> SNIP >>>>> >>>>>> Events: >>>>>> cpu/period=P,event=0x3c/Duk;CPU_CLK_UNHALTED.THREAD >>>>>> cpu/period=P,umask=0x3/Duk;CPU_CLK_UNHALTED.REF_TSC >>>>>> cpu/period=P,event=0xc0/Duk;INST_RETIRED.ANY >>>>>> cpu/period=0xaae61,event=0xc2,umask=0x10/uk;UOPS_RETIRED.ALL >>>>>> cpu/period=0x11171,event=0xc2,umask=0x20/uk;UOPS_RETIRED.SCALAR_SIMD >>>>>> cpu/period=0x11171,event=0xc2,umask=0x40/uk;UOPS_RETIRED.PACKED_SIMD >>>>>> >>>>>> ================================================= >>>>>> >>>>>> Command: >>>>>> /usr/bin/time /tmp/vtune_amplifier_2019.574715/bin64/perf.thr record >>>>>> --threads=T \ >>>>>> -a -N -B -T -R --call-graph dwarf,1024 --user-regs=ip,bp,sp \ >>>>>> -e cpu/period=P,event=0x3c/Duk,\ >>>>>> cpu/period=P,umask=0x3/Duk,\ >>>>>> cpu/period=P,event=0xc0/Duk,\ >>>>>> cpu/period=0x30d40,event=0xc2,umask=0x10/uk,\ >>>>>> cpu/period=0x4e20,event=0xc2,umask=0x20/uk,\ >>>>>> cpu/period=0x4e20,event=0xc2,umask=0x40/uk \ >>>>>> --clockid=monotonic_raw -- ./matrix.(icc|gcc) >>>>> >>>>> hum, so I guess the results suck because of the -a option, >>>>> getting extra samples for all the perf record threads >>>>> >>>>> could you try without the -a? you monitor only user events, >>>>> so you're interested only in ./matrix.* samples, right? >>>> >>>> Ok, trying without -a, in per-process mode. >>> >>> Command: >>> >>> /usr/bin/time ./perf.thr record --threads=T \ >>> -N -B -T -R --call-graph dwarf,1024 --user-regs=ip,bp,sp \ >>> -e cpu/period=P,event=0x3c/Duk,\ >>> cpu/period=P,umask=0x3/Duk,\ >>> cpu/period=P,event=0xc0/Duk,\ >>> cpu/period=0xaae61,event=0xc2,umask=0x10/uk,\ >>> cpu/period=0x11171,event=0xc2,umask=0x20/uk,\ >>> cpu/period=0x11171,event=0xc2,umask=0x40/uk \ >>> --clockid=monotonic_raw -- ./matrix.gcc >>> >>> Workload: matrix multiplication in 128 threads >>> >>> T : 272 >>> P (period, ms) : 0.35 >>> runtime overhead (%) : 13x ~ 87.73 / 6.81 >> >> how do you meassure this? > > This is the ratio of elapsed times: > runtime overhead (%) : elapsed_time_under_profiling / elapsed_time > i.e. > > /usr/bin/time ./matrix.gcc > ... > 767.03user 11.17system 0:06.81elapsed 11424%CPU (0avgtext+0avgdata > 100756maxresident)k > 88inputs+0outputs (0major+139898minor)pagefaults 0swaps > > so elapsed_time = 6.81 sec > > elapsed_time_uder_profiling is elapsed value from output of > > /usr/bin/time ./perf.thr record --threads=T ... > >> >>> data loss (%) : 0 >>> LOST events : 36 >>> SAMPLE events : 8048542 >>> perf.data size (GiB) : 10 >> >> any idea why does it have some much more samples? > > Presumably, this is because period is 350us and this is the smallest > one that perf.thr manages to capture data without data loss (=0) when T=272. > However, during collection, I get message that max sampling frequency > is lowered to 3KHz.
Lowering default frequency rate to 3000. Please consider tweaking /proc/sys/kernel/perf_event_max_sample_rate. Thanks, Alexey > > Thanks, > Alexey > >> >> thanks, >> jirka >> >
