TIL: PAPI

For my course project in the course Parallel Programming at IIT Kanpur, I had to simulate an MPI program on the Lab cluster managed by the CSE Deapartment. Not going into too many details, I wanted to get a measure of time taken by the process between two functions even when the process is being switched in and out of RAM.

If process scheduling wasn’t an issue, I could have directly timed between the two function calls and then used that. Since process scheduling comes into play, there is no reliable way (that I could find) to get the time spent by the process running on the CPU. To solve this problem, I decided that it would be better to just get a count of the number of assembly instructions executed between the two functions and then using the average CPI of the architecture and the clock speed of the processor installed in the CSE lab, calculate the time taken.

While searching how to do this, I found out about Performance Application Programming Interface (PAPI). This was an awesome library which allows measuring exactly what I wanted with a simple to use API. For example, in my use-case, all I needed to do was this:

long long ins_count;
float rtime, ptime, ipc;
handle_papi_error(PAPI_ipc(&rtime, &ptime, &ins_count, &ipc));

The above would give me the instruction count when needed in the ins_count variable.

I haven’t explored PAPI fully, but found it pretty interesting so that’s all.

comments powered by Disqus