Hello,
This is to optimize the event synthesis during perf record.
The first patch is to reduce memory usage when many threads are used.
The second is to avoid unncessary syscalls for kernel threads. And
the last one is to reduce the number of threads to iterate when new
threads are being created at the same time.
Unfortunately there's no dramatic improvement here but I can see ~5%
gain in the 'perf bench internals synthesize' on a big machine.
(The numbers are not stable though)
Before:
# perf bench internals synthesize --mt -M1 -I 100
# Running 'internals/synthesize' benchmark:
Computing performance of multi threaded perf event synthesis by
synthesizing events on CPU 0:
Number of synthesis threads: 1
Average synthesis took: 68831.480 usec (+- 101.450 usec)
Average num. events: 9982.000 (+- 0.000)
Average time per event 6.896 usec
After:
# perf bench internals synthesize --mt -M1 -I 100
# Running 'internals/synthesize' benchmark:
Computing performance of multi threaded perf event synthesis by
synthesizing events on CPU 0:
Number of synthesis threads: 1
Average synthesis took: 65036.370 usec (+- 158.121 usec)
Average num. events: 9982.000 (+- 0.000)
Average time per event 6.515 usec
Thanks,
Namhyung
Namhyung Kim (3):
perf tools: Use /proc/<PID>/task/<TID>/status for synthesis
perf tools: Skip MMAP record synthesis for kernel threads
perf tools: Use scandir() to iterate threads
tools/perf/util/synthetic-events.c | 88 ++++++++++++++++++++----------
1 file changed, 58 insertions(+), 30 deletions(-)
--
2.30.0.365.g02bc693789-goog
