lkml.org 
[lkml]   [2018]   [Aug]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    SubjectRe: [PATCH v4 0/2]: perf: reduce data loss when profiling highly parallel CPU bound workloads
    From
    Date

    Hi Kim,

    On 28.08.2018 22:43, Kim Phillips wrote:
    > On Tue, 28 Aug 2018 18:44:57 +0300
    > Alexey Budankov <alexey.budankov@linux.intel.com> wrote:
    >
    >> Experiment with profiling matrix multiplication code executing 128
    >> threads on Intel Xeon Phi (KNM) with 272 cores, like below,
    >> demonstrates data loss metrics value of 98%:
    >
    > So I took these two patches for a quick test-drive on intel-pt.

    Thanks for testing that out in this scenario! It hasn't been tested yet.

    >
    > BEFORE (acme's today's perf/core branch):
    >
    > $ sudo perf version
    > perf version 4.18.rc7.g55fc647
    > $ sudo perf record -e intel_pt// dd if=/dev/zero of=/dev/null count=100000
    > 100000+0 records in
    > 100000+0 records out
    > 51200000 bytes (51 MB, 49 MiB) copied, 0.0868081 s, 590 MB/s
    > [ perf record: Woken up 21 times to write data ]
    > [ perf record: Captured and wrote 41.302 MB perf.data ]
    > $
    >
    > AFTER (== BEFORE + these two patches):
    >
    > $ sudo ./perf version
    > perf version 4.18.rc7.gbc1c99
    > $ sudo perf record -e intel_pt// dd if=/dev/zero of=/dev/null count=100000
    > 100000+0 records in
    > 100000+0 records out
    > 51200000 bytes (51 MB, 49 MiB) copied, 0.0931142 s, 550 MB/s
    >
    > ...and it's still running, minutes afterwards. Before I kill it,
    > here's some strace output:
    >
    > nanosleep({tv_sec=0, tv_nsec=500000}, NULL) = 0
    > lseek(3, 332556518, SEEK_SET) = 332556518
    > write(3, "D\0\0\0\0\0\10\0", 8) = 8
    > lseek(3, 0, SEEK_CUR) = 332556526
    > futex(0x7f221e7252c8, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e7252cc, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e7252c8, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e7252cc, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e7252c8, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e7252cc, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e7252c8, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > nanosleep({tv_sec=0, tv_nsec=500000}, NULL) = 0
    > lseek(3, 332578462, SEEK_SET) = 332578462
    > write(3, "D\0\0\0\0\0\10\0", 8) = 8
    > lseek(3, 0, SEEK_CUR) = 332578470
    > futex(0x7f221e7252cc, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > nanosleep({tv_sec=0, tv_nsec=500000}, NULL) = 0
    > lseek(3, 332598822, SEEK_SET) = 332598822
    > write(3, "D\0\0\0\0\0\10\0", 8) = 8
    > lseek(3, 0, SEEK_CUR) = 332598830
    > futex(0x7f221e7252c8, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e7252cc, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e7252c8, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e7252cc, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e7252c8, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e725200, FUTEX_WAKE_PRIVATE, 1) = 1
    > futex(0x7f221e7252cc, FUTEX_WAKE_PRIVATE, 1) = 1
    > ^Cstrace: Process 3597 detached
    >
    > I can't prove that it's these two patches that create the hang, but
    > this does look like a livelock situation...hm, hitting ^C doesn't stop
    > it...had to kill -9 it...erm, does 'perf record -e intel_pt// dd...'
    > work for you on a more standard machine?:
    >
    > $ dmesg | grep Perf
    > [ 0.044226] Performance Events: PEBS fmt3+, Skylake events, 32-deep LBR, full-width counters, Intel PMU driver.

    Strace patterns look similar to the ones implemented in the patches.
    Let me reproduce and investigate the hang locally.

    Thanks,
    Alexey

    >
    > Thanks,
    >
    > Kim
    >

    \
     
     \ /
      Last update: 2018-08-29 10:41    [W:4.915 / U:0.080 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site