Details
Major
Description
Even though 10.1 has quite a few generic optimizations it still shows slightly worse performance compared to 10.0. This trend can be reproduced quite reliably on different platforms (10.0 -> 10.1):
1 thread avg TPS: 805.91 -> 780.83 (3.11%) 1 thread max TPS: 810.60 -> 788.50 (2.73%) 64 threads avg TPS: 23026.92 -> 22346.58 (2.95%) 64 threads max TPS: 23245.01 -> 22666.61 (2.49%)
One of the biggest contributors to this slowdown (~2%) seem to be Exec_time_tracker. I get the following numbers if it is commented out (10.0 -> 10.1):
1 thread avg TPS: 805.91 -> 796.44 (1.18%) 1 thread max TPS: 810.60 -> 806.29 (0.53%)
Gliffy Diagrams
Attachments
Activity
- All
- Comments
- Work Log
- History
- Activity
- Transitions
(Note to self: before this patch is pushed, sort out how it works with UNIONs)
The patch seem to be incomplete: ANALYZE_START_TRACKING /ANALYZE_STOP_TRACKING in sql_delete.cc and sql_update.cc were not updated. I'll call those with do_tracking= 1.
I can see little improvement, almost as good as if it was when I commented it out.
perf proof (event: cycles):
ha_index_next() 0.38% -> 0.25% my_timer_cycles() - 0.36% -> out of radar
So we freed up 0.49% CPU cycles.
When tracker functionality is removed from TABLE_IO_WAIT perf says:
ha_index_next() 0.38% -> 0.18% my_timer_cycles() - 0.36% -> out of radar
That is 0.56%.
Ideally we should have different method with out-of-loop condition check. But that may be complex. Reducing number of conditions from 2 to 1 is trivial and should be worthy:
#define TABLE_IO_WAIT(TRACKER, PSI, OP, INDEX, FLAGS, PAYLOAD) \
{ \
if (unlikely(tracker)) \
{ \
tracker->start_tracking(); \
MYSQL_TABLE_IO_WAIT(PSI, OP, INDEX, FLAGS, PAYLOAD); \
tracker->stop_tracking(); \
} \
else \
MYSQL_TABLE_IO_WAIT(PSI, OP, INDEX, FLAGS, PAYLOAD); \
}
Following the IRC discussion, I've made a patch that make ANALYZE timing hooks not to be invoked for non-ANALYZE statements.
However, it made no difference in execution on my system. Sergey Vojtovich, could you please check if the patch fixes the problem?