mikemccand commented on issue #15662:
URL: https://github.com/apache/lucene/issues/15662#issuecomment-3862283921

   Thanks @msokolov -- yes please!  If anyone can repro that would help!  
Concurrent binary search / git-bisect!
   
   I've already tried to reduce my cycle time -- cut out the not-needed 
benchmarks in my iterations, and I can stop the benchy early (don't need all 20 
JVM iterations for confidence measure since this signal (~10-48% slowdown!!) is 
quite confidently apparent after one JVM), but all of that is still slow, 
especially building the deterministic fixed search index.  I have thought about 
simply re-using that index --that would be really fast, and probably (almost 
certainly?) this slowdown is a search-time thing.
   
   It is odd that @shubhamsrkdev couldn't repro -- it does make it seem likely 
it's something in all the package updates, or something about many, many core 
CPUs (that's why I liked the shared-`AtomicInteger` theory, but, surprise! 
😮😮😮😲😲😲😱😱😱, it's not that!).  We've seen such things in the past, where 64/128 
core `beast3` shows different results (though usually not THIS different -- a 
small gain for @shubhamsrkdev runs, and a massive lost for nightly box).
   
   The [beast3 
box](https://blog.mikemccandless.com/2021/01/apache-lucene-performance-on-128-core.html)
 is also oldish now -- 150 human years equivalent!! (one computer year = 30 
human years).  CPU is Ryzen Threadripper 3990X, first released ~6 years ago!  
Probably/maybe `beast4` will finally move to the cloud... and then we could do 
cloudy/rainy things more easily (run many benchmarks, other people can run same 
nightlies, repro on different hardware, ...).


-- 
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.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to