lkml.org 
[lkml]   [2017]   [Jun]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    SubjectRe: [PATCH v3 1/n] perf/core: addressing 4x slowdown during per-process profiling of STREAM benchmark on Intel Xeon Phi
    From
    Date
    On 19.06.2017 16:38, Mark Rutland wrote:
    > On Mon, Jun 19, 2017 at 01:46:39PM +0100, Mark Rutland wrote:
    >> On Fri, Jun 16, 2017 at 05:22:29PM +0300, Alexey Budankov wrote:
    >>> On 16.06.2017 17:08, Alexey Budankov wrote:
    >>>> On 16.06.2017 12:09, Mark Rutland wrote:
    >>>>> There's a --per-thread option to ask perf record to not duplicate the
    >>>>> event per-cpu.
    >>>>>
    >>>>> If you use that, what amount of slowdown do you see?
    >>>
    >>> After applying all three patches:
    >>>
    >>> - system-wide collection:
    >>>
    >>> [ perf record: Woken up 1 times to write data ]
    >>> [ perf record: Captured and wrote 303.795 MB perf.data (~13272985 samples) ]
    >>> 2162.08user 176.24system 0:12.97elapsed 18021%CPU (0avgtext+0avgdata
    >>> 1187208maxresident)k
    >>> 0inputs+622624outputs (0major+1360285minor)pagefaults 0swaps
    >>>
    >>> - per-process collection:
    >>>
    >>> [ perf record: Woken up 5 times to write data ]
    >>> [ perf record: Captured and wrote 1.079 MB perf.data (~47134 samples) ]
    >>> 2102.39user 153.88system 0:12.78elapsed 17645%CPU (0avgtext+0avgdata
    >>> 1187156maxresident)k
    >>> 0inputs+2272outputs (0major+1181660minor)pagefaults 0swaps
    >>>
    >>> Elapsed times look similar. Data file sizes differ significantly.
    >>
    >> Interesting. I wonder if that's because we're losing samples due to
    >> hammering the rb, or if that's a side-effect of this patch.
    >>
    >> Does perf report describe any lost chunks?
    >>
    >> For comparison, can you give --per-thread a go prior to these patches
    >> being applied?
    >
    > FWIW, I had a go with (an old) perf record on an arm64 system using
    > --per-thread, and I see that no samples are recorded, which seems like a
    > bug.
    >
    > With --per-thread, the slwodown was ~20%, whereas with the defaults it
    > was > 400%.

    That looks similar to what I am observing in per-process single thread
    profiling >4x slowdown.

    >
    > Thanks,
    > Mark.
    >

    \
     
     \ /
      Last update: 2017-06-19 16:09    [W:2.233 / U:0.220 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site