<br><br>On Thursday, April 26, 2012 5:30:23 AM UTC-4, c.pignedoli wrote:<blockquote class="gmail_quote" style="margin: 0;margin-left: 0.8ex;border-left: 1px #ccc solid;padding-left: 1ex;">Dear all,
<br>I have the impression I am missing many recent improvements of the code,
<br>I was not able to follow them.
<br>
<br>I usually run standard DFT calculations for bulk/surface systems with
<br>700-1500 atoms (10 to 18 electrons each)
<br>on a Cray Xe6 (AMD Interlagos 2 x 16-core 64-bit CPUs, 32 GB per compute node)
<br>
<br>Do you have suggestions for parameters that would allow for an optimal use of
<br>this architecture? (a problem I have is the limited amount of memory
<br>per core 1GB)
<br></blockquote><div><br></div><div>you should compile with OpenMP support (which also means,</div><div>that you have to ditch the crappy default PGI programming environment</div><div>and use gcc/gfortran instead).</div><div><br></div><div>for plain quickstep calculations there should be a</div><div>significant performance increase when running with</div><div>4 MPI tasks per node and 8 threads per MPI task.</div><div><br></div><div>due to the massively reduced need for data</div><div>replication, you should not run out of memory.</div><div><br></div><div>since interlagos CPUs do some kind of improved</div><div>hyperthreading and two tasks will have to share</div><div>a floating point unit, i would expect that you should</div><div>get a large improvement already by just using</div><div>only half the cores when using plain MPI only.</div><div>make sure that you use the proper "distance"</div><div>flag to aprun, however, since you need to make</div><div>sure you have exactly one MPI task per </div><div>"bulldozer" subunit.</div><div><br></div><div>i would talk to the user support specialists</div><div>of the machine, to make sure you get it right.</div><div><br></div><div>cheers,</div><div> axel.</div><div> </div><blockquote class="gmail_quote" style="margin: 0;margin-left: 0.8ex;border-left: 1px #ccc solid;padding-left: 1ex;">
<br>Thanks in advance for your help
<br>
<br>Carlo
<br></blockquote>