Using Vtune for monitoring multiple co-running functions

[YzhangUCSC]

Hi Community!

I'm a beginner at using Vtune. I have a question about how to measure the performance for each functions in co-running functions groups? Here's the example:

1. We have function A and B. If we use Vtune to monitor the LLC/Mem Bandwidth performance under stand alone setting, we can easily do this by using vtune command lines.
2. Now, when function A and B running together, and say that function A is running on CPU 0 of Numa Node 0; function B running on CPU 1-23 on Numa node 0. The results of profiling is the overall performance of two functions, but not result of A and result of B separately.

How can I check the performance deduction on function A? The way of using PID not works here since function A is short-live, like it will stop in less than 1 sec.

Thanks any possible help in advance!

[YzhangUCSC]

Ohh I guess mabe ittpause() and ittresume() could work. I'll give it a try and then get back

[yuzhang3_intel]

I think there are some methods to try:

1.  Yes, ITT API can control the collection as you mentioned.

2.  You can specify the CPU(s) for the collection

https://www.intel.com/content/www/us/en/docs/vtune-profiler/user-guide/2024-0/cpu-mask.html

3. You can set a small value for the sampling-interval. 

https://www.intel.com/content/www/us/en/docs/vtune-profiler/user-guide/2024-0/sampling-interval.html

[YzhangUCSC]

Got it! BTW, should I place the vtune profiler on another numa node's CPU, like CPU 24-47? Since functions A and B used all CPUs in Numa 0. If I still play the vtune on Numa 0, there could be a possible interference in performance right?

[yuzhang3_intel]

Generally, the VTune is running with workload together. Maybe you can try to specify VTune to run on cpu0 using the taskset tool, and function A on cpu1, function B on cpu2, etc., However, VTune sampling still has an impact on other functions.