abstractdog commented on pull request #1280:
URL: https://github.com/apache/hive/pull/1280#issuecomment-669038358


   @zabetak : let me grab the opportunity to thank you for your [JMH 
benchmarks](https://issues.apache.org/jira/browse/HIVE-23880?focusedCommentId=17163111&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-17163111)!
 it helped a lot, some of my findings from the last 2 weeks:
   
   1. on cluster, JDK11 is better, in every scenario, we'll have to switch to 
that in LLAP daemons
   
   2. more threads doesn't make any serious improvement <- that's the most 
important what I've found in the last two weeks...basically, my implementation 
was wrong, and the results got distorted by the improper usage of executor 
service (that's what is fixed in the new, squashed commit), so now, on the 
cluster I can see results which are in line with your JMH findings
   
   3. removed automatic thread calculation: performance tests revealed that 1 
thread is the most optimal, and can lead to serious improvements, this is 
something that cannot be measured from JMH easily because the advantage of 1 
thread (which is the main task thread + 1 thread) is to decouple from the main 
thread, and let it handle other, probably CPU heavy stuff (waiting for inputs 
one by one, build vectorized row batches one by one, etc.), by this I reduced 
the task runtime by 50-60 seconds (170s -> 110s)
   
   4. as agreed with @ashutoshc, I've left the support of multiple threads in 
the code, because we don't know if we can have the advantage of it later, and 
the split logic doesn't consume significant amount of resources...but I've set 
default 1 thread in HiveConf in order to let the user know that this is the 
recommended, optimal usage
   
   cc: @pgaref , @ashutoshc 


----------------------------------------------------------------
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:
[email protected]



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

Reply via email to