How can I precisely measure cache related information on Phi?

[YW]

HI,

I know General Exploration of vTUne provides some cache related information of programs running on Phi. But it's too coarse, for example, it doesn't have L2 miss ratio, and it contains some hit ratio more than 100%...

Can anyone suggest a better tool to precisely measure the cache related information of programs running on Phi? Basically, I would like to know the L1, L2 cache hit/miss ratio and the penalty of a cache miss.

Thanks!

[Sumedh_N_Intel · Accepted solution]

Hi, 

Intel VTune Amplifier XE uses hardware counters present in the Performance Monitoring Units (PMUs) to calculate various performance metrics. These counters can programmed to monitor various hardware events such as cache misses and the resulting counts can be used in conjunction with other event counts to calculate metrics such as hit ratios. However, the events that are available on a processor are dependent on the underlying hardware. The hardware on the Intel Xeon Phi coprocessor does not support hardware events that can be reliably used to calculate L2 hit ratios. As a result, the you do not see these metrics in the general exploration analysis in Intel VTune Amplifier XE. I believe that as long as you rely on PMUs for performance analysis on the coprocessor, the general exploration on the Intel VTune Amplifier XE analysis is going to be as good as it gets. 

Another point to note is that Intel VTune Amplifier XE uses statistical information and hence the results is always an approximation. There are several reasons due to which the data collected by Intel VTune Amplifier XE may be statistically incorrect. These may result in errors such as cache ratios > 100, etc. You can read more about statistical validity of results at : https://software.intel.com/en-us/blogs/2013/05/29/sanity-check-statistical-data-validity-of-intel-vtune-amplifier-xe-results