Messages in this thread | | | Date | Thu, 13 Aug 2009 14:13:16 +0530 | From | Balbir Singh <> | Subject | Re: [UPDATED][PATCH][mmotm] Help Root Memory Cgroup Resource Counters Scale Better (v5) |
| |
* Ingo Molnar <mingo@elte.hu> [2009-08-13 10:35:24]:
> > * Balbir Singh <balbir@linux.vnet.ibm.com> wrote: > > > Without Patch > > > > Performance counter stats for '/home/balbir/parallel_pagefault': > > > > 5826093739340 cycles # 809.989 M/sec > > 408883496292 instructions # 0.070 IPC > > 7057079452 cache-references # 0.981 M/sec > > 3036086243 cache-misses # 0.422 M/sec > > > With this patch applied > > > > Performance counter stats for '/home/balbir/parallel_pagefault': > > > > 5957054385619 cycles # 828.333 M/sec > > 1058117350365 instructions # 0.178 IPC > > 9161776218 cache-references # 1.274 M/sec > > 1920494280 cache-misses # 0.267 M/sec > > Nice how the instruction count and the IPC value incraesed, and the > cache-miss count decreased. > > Btw., a 'perf stat' suggestion: you can also make use of built-in > error bars via repeating parallel_pagefault N times: > > aldebaran:~> perf stat --repeat 3 /bin/ls > > Performance counter stats for '/bin/ls' (3 runs): > > 1.108886 task-clock-msecs # 0.875 CPUs ( +- 4.316% ) > 0 context-switches # 0.000 M/sec ( +- 0.000% ) > 0 CPU-migrations # 0.000 M/sec ( +- 0.000% ) > 254 page-faults # 0.229 M/sec ( +- 0.000% ) > 3461896 cycles # 3121.958 M/sec ( +- 3.508% ) > 3044445 instructions # 0.879 IPC ( +- 0.134% ) > 21213 cache-references # 19.130 M/sec ( +- 1.612% ) > 2610 cache-misses # 2.354 M/sec ( +- 39.640% ) > > 0.001267355 seconds time elapsed ( +- 4.762% ) > > that way even small changes in metrics can be identified as positive > effects of a patch, if the improvement is beyond the error > percentage that perf reports. > > For example in the /bin/ls numbers i cited above, the 'instructions' > value can be trusted up to 99.8% (with a ~0.13% noise), while say > the cache-misses value can not really be trusted, as it has 40% of > noise. (Increasing the repeat count will drive down the noise level > - at the cost of longer measurement time.) > > For your patch the improvement is so drastic that this isnt needed - > but the error estimations can be quite useful for more borderline > improvements. (and they are also useful in finding and proving small > performance regressions)
Thanks for the tip, let me try and use the repeats feature. BTW, nice work on the perf counters, I was pleasantly surprised to see a wonderful tool in the kernel with a good set of options and detailed analysis capabilities.
-- Balbir
| |