[petsc-users] Building with MKL 10.3

[Robert Ellis]

All,

Coincidentally, I have just spent much of the last two weeks testing the effect of the latest MKL blas-lapack release on a large domain decomposition style PETSc application, MPICH2, Windows, Intel Fortran, VS 2010 C++, all static linked, no hyperthreading, multi-blade, latest dual hex core Xeons, 25GB ram. Each core contains a domain and a shell is used for the matrix operations. Regardless of setting the number of threads for MKL or OMP, the MKL performance was worse than simply using --download-f-blas-lapack=1. My interpretation is that the decomposition of one domain per core effectively saturates the hardware and performance is actually degraded using the more sophisticated MKL. 

And, yes, I know that Windows is less than ideal for this type of work but there are other constraints...

Cheers,
Rob 


-----Original Message-----
From: petsc-users-bounces at mcs.anl.gov [mailto:petsc-users-bounces at mcs.anl.gov] On Behalf Of Preston, Eric - IS
Sent: Tuesday, March 15, 2011 1:55 PM
To: petsc-users at mcs.anl.gov
Subject: Re: [petsc-users] Building with MKL 10.3


As the original poster, I can say I didn't give much thought to using the threaded vs. non-threaded MKL libraries, I was just sharing a solution to building with a different MKL variant (which should probably be incorporated into the config scripts.) However, the discussion it spawned was informative.  The main consideration there was not performance but compatibility with your other libraries. I know on windows, for instance, different run-time libraries are used in each case and all your components must be compiled with the same option. Thankfully, I'm not using windows for this project, so it might not make any difference.

Eric Preston


> Hello,
>    I am neither a regular Petsc user nor contributor so preemptive
>apologies if I am completely off the line here.
>
>I am not sure if the original poster had hyper-threading in mind when he
>asked about multi-threading, in case that was the idea, I don't think using
>petsc with MKL (HT) is going to give any benefit, I don't think MKL is
>really resource insensitive.
>
>Also I wonder what percentage of the code is actually blas/lapack intensive
>to make any significant dent in wall cock?
>
>of course +1 to everything else posed above.
>
>Cheers,
>
>C.S.N


This e-mail and any files transmitted with it may be proprietary and are intended solely for the use of the individual or entity to whom they are addressed. If you have received this e-mail in error please notify the sender.
Please note that any views or opinions presented in this e-mail are solely those of the author and do not necessarily represent those of ITT Corporation. The recipient should check this e-mail and any attachments for the presence of viruses. ITT accepts no liability for any damage caused by any virus transmitted by this e-mail.