saintstack commented on pull request #3371:
URL: https://github.com/apache/hbase/pull/3371#issuecomment-859169361
Thread local was used to avoid threads having to coordinate around a
resource; this plus ringbuffer was meant to get us a write that was w/o locking
(till we hit dfsclient at least). This impl is old though now given redo in
asyncwal.
Thanks for doing the nice graphs and perf runs above.
Interesting on caffeine vs guava; probably this workload (but in compare of
caffeine against our block cache CHM, CHM seemed better though caffeine looked
to have much nicer 'locking' profile).
It is hard to argue w/ a perf improvement (and improved safety). I'd be in
favor of commit.
Funny though how its 103 appends per sync w/ patch and w/o (buffer limit?).
You might try using get/put instead of getIfPresent or whatever it is you
are using just to see what is possible perf-wise.
Have you looked at flamegraphs for asyncwal? Perhaps you'll see where the 1%
is going? Yeah, would be good to get @Apache9 input.
Nice find @bharathv (after reading the JIRA)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
