<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
</head>
<body style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;" class="">
Dear Jon,
<div class=""><br class="">
</div>
<div class="">we run a bunch of AMD EPYC Naples Dual Socket servers with GPFS in our TSM Server Cluster. From what I can say it runs stable, but IO performance in general and GPFS performance in particular - even compared to an Xeon E5 v3 system - is rather
poor. So to put that into perspective on the Xeon Systems with two EDR IB Links, we get 20GB/s read and write performance to GPFS using iozone very easily. On the AMD systems - with all AMD EPYC tuning suggestions applied you can find in the internet - we
get around 15GB/s write but only 6GB/s read. We also opened a ticket at IBM for this but never found out anything. Probably because not many are running GPFS on AMD EPYC right now? The answer from AMD basically was that the bad IO performance is expected in
Dual Socket systems because the Socket Interconnect is the bottleneck. (See also the IB tests DELL did <a href="https://www.dell.com/support/article/de/de/debsdt1/sln313856/amd-epyc-stream-hpl-infiniband-and-wrf-performance-study?lang=en" class="">https://www.dell.com/support/article/de/de/debsdt1/sln313856/amd-epyc-stream-hpl-infiniband-and-wrf-performance-study?lang=en</a> as
soon as you have to cross the socket border you get only half of the IB performance)</div>
<div class=""><br class="">
</div>
<div class="">Of course with ROME everything get’s better (that’s what AMD told us through our vendor) but if you have the chance then I would recommend to benchmark AMD vs. XEON with your particular IO workloads before buying.</div>
<div class=""><br class="">
</div>
<div class="">Best Regards,</div>
<div class="">Stephan Peinkofer</div>
<div class="">
<div class="">
<div>-- <br class="">
Stephan Peinkofer<br class="">
Dipl. Inf. (FH), M. Sc. (TUM)<br class="">
<br class="">
Leibniz Supercomputing Centre<br class="">
Data and Storage Division<br class="">
Boltzmannstraße 1, 85748 Garching b. München<br class="">
URL: <a href="http://www.lrz.de" class="">http://www.lrz.de</a></div>
</div>
<div><br class="">
<blockquote type="cite" class="">
<div class="">On 22. Oct 2019, at 11:12, Jon Diprose <<a href="mailto:jon@well.ox.ac.uk" class="">jon@well.ox.ac.uk</a>> wrote:</div>
<br class="Apple-interchange-newline">
<div class="">
<div class="">Dear GPFSUG,<br class="">
<br class="">
I see the faq says Spectrum Scale is supported on "AMD Opteron based servers".<br class="">
<br class="">
Does anyone know if/when support will be officially extended to cover AMD Epyc, especially the new 7002 (Rome) series?<br class="">
<br class="">
Does anyone have any experience of running Spectrum Scale on Rome they could share, in particular for protocol nodes and for plain clients?<br class="">
<br class="">
Thanks,<br class="">
<br class="">
Jon<br class="">
<br class="">
-- <br class="">
Dr. Jonathan Diprose <<a href="mailto:jon@well.ox.ac.uk" class="">jon@well.ox.ac.uk</a>> Tel: 01865 287837<br class="">
Research Computing Manager<br class="">
Henry Wellcome Building for Genomic Medicine Roosevelt Drive, Headington, Oxford OX3 7BN<br class="">
<br class="">
_______________________________________________<br class="">
gpfsug-discuss mailing list<br class="">
gpfsug-discuss at <a href="http://spectrumscale.org" class="">spectrumscale.org</a><br class="">
<a href="http://gpfsug.org/mailman/listinfo/gpfsug-discuss" class="">http://gpfsug.org/mailman/listinfo/gpfsug-discuss</a><br class="">
</div>
</div>
</blockquote>
</div>
<br class="">
</div>
</body>
</html>