lkml.org 
[lkml]   [2020]   [Oct]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH] perf data: Allow to use stdio functions for pipe mode
Hi Ian,

On Thu, Oct 29, 2020 at 1:37 AM Ian Rogers <irogers@google.com> wrote:
>
> On Wed, Oct 28, 2020 at 1:56 AM Namhyung Kim <namhyung@kernel.org> wrote:
> >
> > When perf data is in a pipe, it reads each event separately using
> > read(2) syscall. This is a huge performance bottleneck when
> > processing large data like in perf inject. Also perf inject needs to
> > use write(2) syscall for the output.
> >
> > So convert it to use buffer I/O functions in stdio library for pipe
> > data. This makes inject-build-id bench time drops from 20ms to 8ms.
>
> This is great! Reading the numbers below the times are reduced to
> 67.9% of their previous, this is a little less than 20ms to 8ms so
> perhaps check that number.

The numbers before this patch:

$ ./perf bench internals inject-build-id
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 20.594 msec (+- 0.034 msec)
Average time per event: 2.019 usec (+- 0.003 usec)
Average memory usage: 8319 KB (+- 0 KB)
Average build-id-all injection took: 19.443 msec (+- 0.177 msec)
Average time per event: 1.906 usec (+- 0.017 usec)
Average memory usage: 7490 KB (+- 0 KB)

>
> > $ perf bench internals inject-build-id
> > # Running 'internals/inject-build-id' benchmark:
> > Average build-id injection took: 8.074 msec (+- 0.013 msec)
> > Average time per event: 0.792 usec (+- 0.001 usec)
> > Average memory usage: 8328 KB (+- 0 KB)
> > Average build-id-all injection took: 5.490 msec (+- 0.008 msec)
> > Average time per event: 0.538 usec (+- 0.001 usec)
> > Average memory usage: 7563 KB (+- 0 KB)
> >
> > This patch enables it just for perf inject when used with pipe (it's a
> > default behavior). Maybe we could do it for perf record and/or report
> > later..
>
> For perf record there is also the async IO used for writing. I like
> that this patch is adding the reading to perf_data. Should the async
> IO code live in perf_data rather than record? Should async IO be used
> for reading in that case? What would be the policy of using stdio,
> not-stdio and async IO?

I'm not sure but using async IO for read is meaningless since we cannot
do anything without the data. The reason I added this was that the pipe
mode does so many small I/O with syscalls. But perf record will mostly
do large I/O so I don't expect big speedup for that.

Thanks
Namhyung

\
 
 \ /
  Last update: 2020-10-29 12:17    [W:1.912 / U:0.004 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site