Latency values in perf mem report

Latency values in perf mem report

[Milan Radulovic]

Hello all,

I measured load latency using perf mem rec and I have some doubts regarding
the results.
It is a 2 socket machine, with 2x Intel E5-2620 (2x 6 cores) at 2.3 GHz.
I sampled the application which runs on all 12 cores for half an hour.
When I print out the results with perf mem rep I get a very broad range of
latencies for certain events. Specifically, Local RAM hits have latencies
from ~20 cycles up to ~2000 cycles.
It is similar for Remote RAM hits (which happen rarely).
First, I used 200 samples/second, but it is the same when I use 50
samples/second.

Is there some useful advice how could I filter these results?
I guess RAM hits should not have latencies bellow ~200 cycles.

Thanks for your help,
Milan

Re: Latency values in perf mem report

[Vince Weaver]

On Wed, 5 Aug 2015, Milan Radulovic wrote:

> I measured load latency using perf mem rec and I have some doubts regarding
> the results.
> It is a 2 socket machine, with 2x Intel E5-2620 (2x 6 cores) at 2.3 GHz.

I'm not sure if it applies, but you might want to take a look at this:
	
https://software.intel.com/en-us/articles/performance-monitoring-on-intel-xeon-processor-e5-family

Vince

Re: Latency values in perf mem report

[Milan Radulovic]

Vince Weaver <vincent.weaver <at> maine.edu> writes:
>
https://software.intel.com/en-us/articles/
performance-monitoring-on-intel-xeon-processor-e5-family
> 
> Vince

Hello Vince,

I took a look at the document you suggested. 
It seems that the counter which measures
latency on SandyBridge may undercount...
If you apply some of the workarounds which they suggest, I get better results.
There are still some small values but they appear far less.

Thanks for the help!

Best regards,
Milan