how detect qpi and snoop traffic from system with 2way or more way westmere ex

[SB17]

Hi all

I study NUMA effects and glue-less server platforms. I want to see cross-socket (cross-NUMAnode) traffic and work MESIF protocol as Snoop requests. Can I see it with hardware counters on the westmere ex ?

Thank for your attention to my problem.

[Patrick_F_Intel1]

Yes, you can.

Pat

[SB17]

:)

How ?

Maybe I was not attentive, but I have not found a counters show cross-socket traffic and are not sure which counters allow to see the cross-node snooping

[Patrick_F_Intel1]

Where have you looked so far? What OS are you running? Which chip are you using (cpuid brand string)? Is this just for curiousity or are you trying to resolve an issue? If an issue, what issue are you seeing?

pat

[SB17]

Looked information in the intel pmu counters specifications and other docs. OS Linux. Kernel 2.6.32. CPU Intel 48xx. I use ocperf. It's my curiosity.  Solution - would like to know what exactly counters could be used and/or reference to the documents containing such information

[SB17]

No one really has no ideas or what I said, that is not correct? My english not so good

[Patrick_F_Intel1]

Please see http://www.intel.com/content/dam/www/public/us/en/documents/performance-briefs/xeon-e7-family-uncore-performance-programming-guide.pdf

[SB17]

Thank for your attention to my problem.

[SB17]

Hi

'm terribly sorry, but I have again raised the question.

I found how to get a sokets read bandwidth: 

((offcore_response.any_data.any_llc_miss_0 + offcore_response.any_rfo.any_llc_miss_0  + offcore_response.any_data.any_llc_miss_1 + offcore_response.any_rfo.any_llc_miss_1) * 64 )/1024/1024/ time sec = read bw in MB/sec

it's ok.

Similarly apparently can receive qpi traffic (offcore_response.any_data.remote_dram_0 or offcore_response.any_data.any_llc_miss_0)

But how i can get write bandwidth ?

I tried using the MSR (msr-tools and rdmsr) but did not find a clear description of how to use MSR (which registers) in bash in westmereEX CPU

Sorry for the my English and please do not kick me and Thanks for your time

[Patrick_F_Intel1]

Have you tried Intel PCM utility? I believe it will display the read and write bandwidth per socket. PCM uses events from the uncore performance doc I mentioned previously. For PCM see: http://software.intel.com/en-us/articles/intel-performance-counter-monitor-a-better-way-to-measure-cpu-utilization

Pat

[SB17]

Utility certainly looked, but, first, I did not understand how there use non-standard events (snoop, etc.). Secondly, I would like to do it myself, to better understand how the CPU and profiling of other events.
A link is also seen. A very useful place to start learning PCM, but it does not describe how to calculate this performance (specialy for westmere-ex). Looked like it is done in a likwid-perfctrl (http://code.google.com/p/likwid/) but it is not clear immediately. it is clear that they are get  this metrics from MSR (although I could be wrong)

Thanks for your time