Re: Cache measurement pollution by Oprofile

[Maynard J. <may...@us...>]

Masoud Sajadieh wrote:
> 
> Hello,
> 
> I'm trying to profile cache-related metrics of an application running on a
> 2-thread Atom processor over Linux. I constantly run into the probelm that
> Oprofile makes a lot of cache references and hence can completely distort
> the cache behavior of the system when Oprofile is not running. Oprofile
> seems to run mostly on one thread and the application on the other one. This
> is the typical problem of observation process polluting the measured and
> should have been investigated in the context of run-time cache measurements. 
> 
> I particularly do not have constraints over the speed of measurement and can
> slow down the process if it improves the accuracy of cache access/miss
> results. Are there ways to resolve this using Oprofile? Any lead and pointer
> is appreciated.
I personally have not done the kind of detailed analysis that you are apparently doing here.  So, just looking at proportion of cache event samples in the different areas of your program is not helpful?  Can you explain what you really want to do with these cache-related metrics?  Maybe oprofile isn't the best tool.

But if you have a multi-core system and your main interest is with L1 cache metrics, you could try the technique of setting processor affinity for the oprofile daemon process (oprofiled) to try to isolate its L1 cache usage from the app under analysis.

-Maynard
> 
> Regards,
> -- Masoud