PLASMA running on multiple cores and FLOPS

Open forum for general discussions relating to PLASMA.

PLASMA running on multiple cores and FLOPS

Postby luiceur » Fri Apr 19, 2013 6:32 am

I am trying to benchmark MKL and PLASMA. With MKL there is no doubt that increasing the number of cores, the performance will increase. Now, what about PLASMA?
I've managed to set an SGEMM benchmark suite. Following the examples and testing cases provided by PLASMA, I can set the number of cores. My question is, does the application run on the number of cores provided or creates threads within the same core? I am running on an interective node with access to up to 6CPUs x 4 Cores/CPU.
My other question is, how do you meassure the flops? I am currently using the one provided on the flops.h file (FLOPS_SGEMM(M,N,K)). Would this function give me the right number of FLOPS?
Thanks a lot!
luiceur
 
Posts: 3
Joined: Fri Apr 19, 2013 4:52 am

Re: PLASMA running on multiple cores and FLOPS

Postby admin » Mon Apr 22, 2013 10:41 am

PLASMA is a multithreaded software library, so, yes, the objective is that the performance increases with more cores.
That being said, it obviously depends on the workload.
The right way to use PLASMA is to launch as many threads as hardware cores.
Not to be confused with hardware threads.
Hyperthreading is unhealthy for PLASMA.
DO NOT OVERSUBSCRIBE CORES.
If you don't specify the number of threads, PLASMA will launch as many as cores.
If you installed PLASMA with hwloc, PLASMA will correctly avoid hyperthreading.
To answer the second question, we measure performance using the routines in the /timing subdirectory.
Jakub
admin
Site Admin
 
Posts: 79
Joined: Wed May 13, 2009 1:27 pm


Return to User discussion

Who is online

Users browsing this forum: Google [Bot], Yahoo [Bot] and 2 guests

cron