On Mon, Mar 14, 2011 at 7:48 PM, Lisandro Dalcin <span dir="ltr"><<a href="mailto:dalcinl@gmail.com">dalcinl@gmail.com</a>></span> wrote:<br><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
On 14 March 2011 20:48, Barry Smith <<a href="mailto:bsmith@mcs.anl.gov">bsmith@mcs.anl.gov</a>> wrote:<br>
><br>
> Eric,<br>
><br>
> With the current PETSc design that uses one MPI process per core I am not sure that it makes sense to use multi-threaded MKL since the processes are already using all the cores. I could be wrong.<br>
><br>
> Barry<br>
><br>
<br>
Well, I think that PETSc should support sequential codes using<br>
multi-threaded blas-lapack and perhaps some OpenMP. Moreover, perhaps<br>
we could get some little speedup in MatMult for these kinds of<br>
users/applications.<br></blockquote><div><br></div><div>I have never ever seen convincing evidence of this. First, you would need enough bandwidth to satisfy</div><div>2+ cores. This is almost never the case. But suppose you do have this. Then you would need a convincing</div>
<div>reason to use threads instead of MPI processes, which would mean data reuse. But there is very little</div><div>reue here; it is mostly streaming.</div><div><br></div><div> Matt</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
--<br>
<font color="#888888">Lisandro Dalcin<br>
---------------<br>
CIMEC (INTEC/CONICET-UNL)<br>
Predio CONICET-Santa Fe<br>
Colectora RN 168 Km 472, Paraje El Pozo<br>
3000 Santa Fe, Argentina<br>
Tel: <a href="tel:%2B54-342-4511594">+54-342-4511594</a> (ext 1011)<br>
Tel/Fax: <a href="tel:%2B54-342-4511169">+54-342-4511169</a><br>
</font></blockquote></div><br><br clear="all"><br>-- <br>What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.<br>-- Norbert Wiener<br>