Re: [PATCH v2 00/15] Introduce threaded trace streaming for basic perf record operation

[Jiri Olsa <jolsa@redhat.com>]

On Wed, Oct 28, 2020 at 10:35:08AM +0300, Alexey Budankov wrote:
> 
> On 28.10.2020 10:08, Namhyung Kim wrote:
> > Hi,
> > 
> > On Wed, Oct 28, 2020 at 1:02 AM Alexey Budankov
> > <alexey.budankov@linux.intel.com> wrote:
> >>
> >>
> >> On 27.10.2020 15:10, Jiri Olsa wrote:
> >>> On Mon, Oct 26, 2020 at 08:59:01PM +0300, Alexey Budankov wrote:
> >>>>
> >>>> On 24.10.2020 18:43, Jiri Olsa wrote:
> >>>>> On Wed, Oct 21, 2020 at 06:52:43PM +0300, Alexey Budankov wrote:
> >>>>>>
> >>>>>> Changes in v2:
> >>>>>> - explicitly added credit tags to patches 6/15 and 15/15,
> >>>>>>   additionally to cites [1], [2]
> >>>>>> - updated description of 3/15 to explicitly mention the reason
> >>>>>>   to open data directories in read access mode (e.g. for perf report)
> >>>>>> - implemented fix for compilation error of 2/15
> >>>>>> - explicitly elaborated on found issues to be resolved for
> >>>>>>   threaded AUX trace capture
> >>>>>>
> >>>>>> v1: https://lore.kernel.org/lkml/810f3a69-0004-9dff-a911-b7ff97220ae0@linux.intel.com/
> >>>>>>
> >>>>>> Patch set provides threaded trace streaming for base perf record
> >>>>>> operation. Provided streaming mode (--threads) mitigates profiling
> >>>>>> data losses and resolves scalability issues of serial and asynchronous
> >>>>>> (--aio) trace streaming modes on multicore server systems. The patch
> >>>>>> set is based on the prototype [1], [2] and the most closely relates
> >>>>>> to mode 3) "mode that creates thread for every monitored memory map".
> >>>>>
> >>>>> so what I liked about the previous code was that you could
> >>>>> configure how the threads would be created
> >>>>>
> >>>>> default --threads options created thread for each cpu like
> >>>>> in your change:
> >>>>>
> >>>>>   $ perf record -v --threads ...
> >>>>>   ...
> >>>>>   thread 0 monitor: 0 allowed: 0
> >>>>>   thread 1 monitor: 1 allowed: 1
> >>>>>   thread 2 monitor: 2 allowed: 2
> >>>>>   thread 3 monitor: 3 allowed: 3
> >>>>>   thread 4 monitor: 4 allowed: 4
> >>>>>   thread 5 monitor: 5 allowed: 5
> >>>>>   thread 6 monitor: 6 allowed: 6
> >>>>>   thread 7 monitor: 7 allowed: 7
> >>>>
> >>>> Yes, it is configurable in the prototype. Even though this patch set
> >>>> doesn't implement that parameters for --thread option, just because
> >>>> VTune doesn't have use cases for that yet, it has still been designed
> >>>> and implemented with that possible extension in mind so it could then
> >>>> be easily added on top of it.
> >>>
> >>> I'm not sure about vtune extensions, but if we are going to
> >>> have --threads option I believe we should make it configurable
> >>> at least to the extend descibed below
> >>
> >> It employs --threads mode only and there are no use cases
> >> observed so far beyond this mode. Do you have or see such
> >> use cases?
> > 
> > I don't know about vtune and other users, but it's an important
> > feature for better performance so I agree with Jiri's opinion to
> > make it flexible for the system requirement.
> 
> For sure, vtune is not the only one for this threaded streaming
> and it should be well suited for perf tool use cases equally.
> And for perf it would be beneficial to document some examples in
> perf-record.txt as a part of this configuration implementation.
> I am not just aware of such examples and that is why I am asking
> you guys.

I saw no LOST events on big servers for some tests with --threads=numa
option, so there was no reason to spawn 200+ threads

jirka