mkl=parallel

Run-time issues concerning Yambo that are not covered in the above forums.

Moderators: myrta gruning, andrea marini, Daniele Varsano, Conor Hogan

Post Reply
martinspenke
Posts: 149
Joined: Tue Apr 08, 2014 6:05 am

mkl=parallel

Post by martinspenke » Mon Dec 01, 2014 9:12 am

Dear Daniele,

Is yambo faster with setting mkl = parallel or sequential for gw and bse calculations ?
If it is faster for mkl = parallel, can you tell how large is the speed up factor vs. the sequential version generally ?

Bests
Martin
Martin Spenke, PhD Student
Theoretisch-Physikalisches Institut
Universität Hamburg, Germany

andrea.ferretti
Posts: 214
Joined: Fri Jan 31, 2014 11:13 am

Re: mkl=parallel

Post by andrea.ferretti » Mon Dec 01, 2014 2:19 pm

Dear Martin,

time ago I've run some benchmark tests on the devalue version of yambo to test the efficiency of parallel libs such as MKL.
I don't have the results at hand but I kind of remember that, yes, you gain something by using parallel against sequential MKL
(yambo extensively relies on blas usage), but the gain is not massive. Instead, better performances can be obtained by an explicit inclusion of multithreading parallelism (e.g. through openmp), which is currently under testing in the devel version of yambo and is going to be available in one of the next releases.

take care
Andrea
Andrea Ferretti, PhD
CNR-NANO-S3 and MaX Centre
via Campi 213/A, 41125, Modena, Italy
Tel: +39 059 2055322; Skype: andrea_ferretti
URL: http://www.nano.cnr.it

Post Reply