Resolution of performance counters

Resolution of performance counters

sid9559's picture

hello Everyone,

Is there a way to tweak with the resolution of the performance counter (event counters). Basically I am doing hotspot analysis for an algorithm and I feel that the execution time is really small that i am not able to get proper results. So, i was wondering if there is a way to reduce the threashold for the hwardware events to trigger.

regards,
siddharth

8 posts / 0 new
Last post
For more complete information about compiler optimizations, see our Optimization Notice.
Tim Prince's picture

You may need to reduce sample after values below default, if running less than 10 seconds. This could cause VTune overheads to become significant.

Peter Wang (Intel)'s picture

Click on existing predefined analysis type (e.g. lightweight-hotspots), right-click to select "Copy from current" to create a new analysis type, you can modify "Sample After" Value.

sid9559's picture

Thank you for the anwers. I am assuming that "sample after" is same as the "CPU sampling time innterval". Also, I could reduce it to 1ms. But the problem is that my algorithm has the total cpu time for 70-100 ms . Is there a way that i can analyse it in a much better way?

The problem is that my hotspot functions are consistent and I am assuming that the reason for this is the resolution.

regards,
siddharth

Peter Wang (Intel)'s picture

"Sample After" value is for lightweight-hotspots, and "CPU sample interval" is for hotspots.

If you use hotspots, please don't usesample interval to 1ms, I suggest to use 3ms. I think that you can run your critical functions several times,for example, put them into loop - in order to collect more samples.

Regards, Peter

sid9559's picture

Thank you for your help. I think i will try to put my critical functions in a loop and then do some analysis.

regards,
siddharth

mark-dewing (Intel)'s picture

There are two sampling technologies used in Amplifier XE - one is based on software timers and the other is based on hardware event counters.

  • The hotspots, concurrency, and locks and waits analysis types are all based on software timers. The sampling interval is set with the 'CPU sampling interval', and defaults to a 10 ms interval. One advantage of these analysis types is stack information.
  • The other analysis types (lightweight hotspots, and all the others) are based on the hardware event counters. The sampling interval for these types are set with the 'Sample-After' value (can be adjusted with custom analysis types). This sampling interval can also be indirectly adjusted with the 'Duration time estimate' item in the Project Properties menu. Or -target-duration-type on the command line. It basically applies a multiplier to all the 'sample-after' values in an analysis types.).

Hopefully this should give a little background to the other recommendations in this thread.

iliyapolak's picture

>>>Is there a way to tweak with the resolution of the performance counter (event counters>>>

The time mesurement resolution should be ~3ms.If your code executes faster you need to run it in small loop.

Login to leave a comment.