ciyongch commented on issue #15429: Operator Performance Regression on CPU
URL: 
https://github.com/apache/incubator-mxnet/issues/15429#issuecomment-508381835
 
 
   @sandeep-krishnamurthy As most of the MXNet's ops will do the parallelism 
via openmp on CPU, from performance perspective, binding CPU cores will reduce 
the performance degradation due to cache misses and get better thread 
scheduling. 
   Without binding the cores,  the performance variance could be bigger due to 
thread/context switch among those available cores, and plus the time of 
benchmarking on an individual op is usually small, like several to a hundred 
microsecond as showed in the above table (in the case of 10 warmup + 100 runs 
for each op).
   BTW, as op level performance tuning is not always reflect its real 
performance in a real model (for example, data is more easily/likely cached in 
op benchmarking, which results the better performance compared to a real model 
scenario). It's better if the future CI will combine some basic/general models 
benchmarking to determine the regression. What do you think?
   
    
   

----------------------------------------------------------------
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]


With regards,
Apache Git Services

Reply via email to