Is MAGMA really hybrid?

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
Posts: 2
Joined: Thu Oct 07, 2010 6:34 am

Is MAGMA really hybrid?

Post by jreybert » Thu Oct 07, 2010 10:50 am


According to magma-v02 documentation, magma should be able to balance work upon cores and GPU. Is it functionnal with magma 0.2?

I am trying MAGMA 0.2 on multicore + GPU machines. The configuration is:
2 x Intel Xeon E5520 => 8 cores
2 x Tesla C1060 (I know Magma only use one GPU)
HyperThreading disabled

I run some test programs provided with magma, with 8 and 4 OpenMP threads.

e.g, for testing_cgeqrf, which time both cgeqrf for MKL and MAGMA libraries, I get these timings:
  • with 8 omp threads: MKL: 119.38 GFlops - MAGMA 69.66 GFlops
  • with 4 omp threads: MKL: 62.33 GFlops - MAGMA 69.29 GFlops
=> Number of threads does not affect MAGMA performances.

I did not run these tests with all the MAGMA functions. So my questions are:
  • are all the MAGMA functions hybrid?
  • if the tested function is, why the number of threads does not affect the performances?
Extra question: is a v3 is scheduled, with the management of several GPUs?

Posts: 1
Joined: Mon Oct 11, 2010 5:19 pm

Re: Is MAGMA really hybrid?

Post by KevinWilson88 » Mon Oct 11, 2010 5:28 pm


I'm not sure myself but I'm also interested in finding out the answer to your question. I found the following document on Magma 0.2 quite informative. May be you will also find your answers there. Check it out ... A-2009.pdf

Good luck!

Posts: 11
Joined: Tue Dec 07, 2010 4:21 am

Re: Is MAGMA really hybrid?

Post by mtacconi » Tue Dec 07, 2010 4:49 am

A rapid look to the library symbol table shows no evidence of undefined reference to whatsoever calls to OpenMP or multithread runtime library.
I would say that this binary release of magma is single CPU threaded / single GPU.

Posts: 1
Joined: Thu Jan 06, 2011 5:59 pm

Re: Is MAGMA really hybrid?

Post by kusario » Fri Jan 14, 2011 4:27 am

that would be a great pitty cause of the speed we could achieve by that with hybrid multicores.

Post Reply