Messages in this thread | | | Date | Fri, 29 Nov 2013 15:01:21 +0100 | Subject | Re: [BUG] perf stat: explicit grouping yields unexpected results | From | Stephane Eranian <> |
| |
On Fri, Nov 29, 2013 at 2:52 PM, Jiri Olsa <jolsa@redhat.com> wrote: > On Fri, Nov 29, 2013 at 02:43:35PM +0100, Stephane Eranian wrote: >> On Fri, Nov 29, 2013 at 2:33 PM, Jiri Olsa <jolsa@redhat.com> wrote: >> > On Sat, Nov 16, 2013 at 07:41:34PM -0800, Andi Kleen wrote: >> >> > I'd say that the default behavior should be what Jiri implemented: get >> >> > the most out of the situation and inform. But you are right in that >> >> > 'forcing' all elements of a group to be valid should be possible as >> >> > well - if a special perf stat option or event format is used. >> >> >> >> When something is multiplexed it can have a very >> >> large measurement error. For workloads that fluctuate quite a bit, and the >> >> fluctuations do not line up well with the multiplexing interval, >> >> the default scaling does not give good results. >> >> >> >> So you expect to get good data, but you get very bad data. >> >> >> >> When collecting data for a large number of events it is important >> >> to group them correctly, so that events that are directly dependent >> >> on each other in equations are properly grouped. >> >> >> >> When explicit groups were added the user likely considered this >> >> problem, so it's not good to silently override the choices. >> >> >> >> If a user doesn't care they can always not use groups. >> >> >> >> > Even in that second case it shouldn't say <unsupported> for everything >> >> > in the result, but should deny the run immediately and return with an >> >> > error, and should tell the user how many events in the group fit and >> >> > which ones didn't. >> >> >> >> Returning this information would be great, but it would really >> >> need an extended errno, or just a error string reported out. >> > >> > (sry for late reply, I was still ooo, and missed this conversation) >> > >> > I agree, when the last event fails sys_perf_event_open >> > due to the validate_group check, we will get just EINVAL >> > >> > Was there any discussion about the error (or erorr string) >> > propagation from sys_perf_event_open? >> > >> > Something like below? user space supply buffer for error string. >> > >> No. Why do you need kernel changes for that. >> Perf gets the error, knows it is grouping and prints an appropriate > > how does perf know it's grouping and not something else? > Because the group_fd on this syscall is not -1.
>> error message. Why do you need kernel for this? > > like how would you differentiate EINVAL from validate_group or say > from set_ext_hw_attr (got by using unsupported cache event) ? > If you cannot, simply abort and print something like: if (group_fd != -1 && ret == EINVAL) warnx("cannot program event X in group. You may be overcommitting an event group, try reducing the number of events/group"
Or something like that.
| |