<html>
<head>
<style><!--
.hmmessage P
{
margin:0px;
padding:0px
}
body.hmmessage
{
font-size: 10pt;
font-family:Verdana
}
--></style>
</head>
<body class='hmmessage'>
<pre>Hi,<br><br>This of course depends very much on all your settings.<br>Up to now I have found that a 3:1 ratio is optimal for rectangular boxes,<br>whereas 2:1 is optimal for rhombic dodecahedra.<br>64 cores was the limit for the benchmark system on that hardware,<br>so I had to increase the number of PME beyond the 1/3 to get optimal performance.<br><br>Berk<br><br><br>Hi,<br><br>I recently got a nice Sun X6275 blade system <br><a href="http://nf.nci.org.au/facilities/vayu/hardware.php" target="_blank">http://nf.nci.org.au/facilities/vayu/hardware.php</a> (ok, not just for me!) <br>whose hardware ought to be better, but very comparable to that used in <br>the GROMACS 4 paper (<a href="http://pubs.acs.org/doi/abs/10.1021/ct700301q" target="_blank">http://pubs.acs.org/doi/abs/10.1021/ct700301q</a>). <br>Figure 5 there demonstrates good scaling out to 38 cpus, but it drops <br>away from linearity by 64. On the same test system on this hardware (pdf <br>attached; blue arrow is ideal linearity) the drop-off is still about the <br>same.<br><br>For 64 cpus, that paper reports using 28 PME nodes, which is an <br>appreciably lower ratio than the 2:1 for the rest of the tests. On my <br>hardware I found that that choice cost about 30% performance, compared <br>with a more reasonable use of 19 PME nodes (64 vs 86 ns/day, pdf <br>attached). Is that 28 a typo for 18?<br><br>Mark<br></pre> <br /><hr />What can you do with the new Windows Live? <a href='http://www.microsoft.com/windows/windowslive/default.aspx' target='_new'>Find out</a></body>
</html>