Identifying the cause of a high CPI ratio

[olivier_m]

Hi all,

I try to identify why the same function has a higher CPI ratio when executed in parallel than when executed by a single thread and I don't understand why !
I use vtune as my performance analyser, I tried to see if there was a problem of cache miss using L1 and L2 data/instruction cache miss rate but there is nothing special. Now, I don't know what looking for.

Any idea / suggestion would be greatly appreciated.

Thanks

Olivier

[TimP]

When the combined memory bus utilization of the threads approaches the maximum capacity, this will happen even without any problems in caching.

[olivier_m]

Thanks, that's what I guess but is there any way to confirm it ? I don't know what vtune event to monitor or what other tool to use to be certain that's the reason of the overhead !


Thanks again.

Olivier

[TimP]

Check the posts on bus utilization in the VTune forum, or just correlate yourself, comparing your performance scaling by function with the memory traffic rates.