Quantcast
Channel: Intel® Cilk™ Plus
Viewing all articles
Browse latest Browse all 77

Timing and benchmarking reproducebility

$
0
0

This is not really a cilk question but I will try anyway because I think you INTEL must have a lot experience.  I have read the section https://software.intel.com/en-us/node/522641 but it is a bit vague.

How do you get reliable timing results when benchmarking cilk programs? Do you use a particular OS in a particular setup for instance?

My experience on Windows is times can varies a lot for even for single threaded runs when the same program is run different points times. A 10% difference can easily be measurement error. Our experience seems to indicate Linux is not much better.

Btw I have disabled hyperthreading and is using a server and not a laptop. I am only user and shut down unneeded applications before running my test. Maybe I should compute averages and variance of run times and apply statistical tests to the results. And use that to conclude about performance.

 


Viewing all articles
Browse latest Browse all 77

Trending Articles



<script src="https://jsc.adskeeper.com/r/s/rssing.com.1596347.js" async> </script>