<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
</head>
<body>
<div class="moz-cite-prefix">Hello all,</div>
<div class="moz-cite-prefix"><br>
</div>
<div class="moz-cite-prefix">TLDR: Try the combination intel/19.1.1.217
and openmpi/3.1.6 on expanse.<br>
</div>
<div class="moz-cite-prefix"><br>
</div>
<div class="moz-cite-prefix">Long version:</div>
<div class="moz-cite-prefix"><br>
</div>
<div class="moz-cite-prefix">I had a similar issue on expanse during
the EUP with great performance on 1 node and worse on multiple
nodes. I tried many combinations of the available (at the time)
mpi implementations and compilers. <br>
</div>
<div class="moz-cite-prefix"><br>
</div>
<div class="moz-cite-prefix">Here is what I found:</div>
<div class="moz-cite-prefix">
<pre class="moz-quote-pre" wrap="">* intel19 + intel mpi : great 1 node speed, poor scaling to >1 node, poor weak scaling
* intel19 + openmpi 4 : same as above
* gcc 10 + openmpi 4 : 40% slower 1 node speed, poor scaling to >1 node, poor weak scaling
* gcc 9 + openmpi 3.1.6 : 40% slower 1 node speed, good scaling to >1 node, acceptable weak scaling
* using intel19 + openmpi 4 executable but with gcc9/openmpi 3.1.6 modules loaded at runtime : great 1 node speed, good scaling to >1 node, acceptable weak scaling</pre>
</div>
<div class="moz-cite-prefix">The MPI implementation that worked was
using openmpi 3.1.6. At the time this was only available if you
used the gcc 9 compilers. However, I contacted support and they
installed a version for the intel compilers. Here is what support
said:</div>
<div class="moz-cite-prefix">
<pre class="moz-quote-pre" wrap="">"I think I can install openmpi/3.1.6 with intel compilers. I have to go back and check but I think the main difference is we are using ibverbs on the openmpi/3.1.6 build and ucx on the openmpi/4.0.4. For most codes ucx has been the faster option but in your case it seems different. I will let you know once the compilers are in place."
</pre>
</div>
<div class="moz-cite-prefix">After he installed it, the combination
of intel19 compilers with openmpi 3.1.6 gives acceptable scaling.
I am not familiar enough with ucx vs ibverbs to comment on if that
is the issue with the AMD clusters. I also have this same issue
on bridges2 which uses the same AMD nodes as expanse, and have not
been able to get the code to perform well on >1 node.</div>
<div class="moz-cite-prefix"><br>
</div>
<div class="moz-cite-prefix">At least for expanse, I'd suggest
trying to load intel 19 and openmpi and see if you get better
scaling. Then if you do, we could inquire with support on the
differences in configurations for openmpi 3.1.6 and openmpi 4.0.4
to see if there is more beyond ucx vs ibverbs. Then, the next
step would be seeing if we can replicate this elsewhere (like
bridges2 or anvil).<br>
</div>
<div class="moz-cite-prefix"><font face="monospace"><br>
</font></div>
<div class="moz-cite-prefix"><font face="monospace">module load
intel/19.1.1.217<br>
module load openmpi/3.1.6</font></div>
<br>
<div class="moz-cite-prefix">Best,<br>
Jim Healy</div>
<div class="moz-cite-prefix">CCRG Research Associate<br>
</div>
<div class="moz-cite-prefix"><br>
</div>
<div class="moz-cite-prefix">On 8/27/21 12:44 PM, Gabriele Bozzola
wrote:<br>
</div>
<blockquote type="cite"
cite="mid:CAMLD_5HhitAvrYcCsdkCytvAGkWd+pnTjzXYw1j+vzmOQrr4mw@mail.gmail.com">
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
<div dir="ltr">Hello,
<div><br>
</div>
<div>Last week I opened a PR to add the configuration files </div>
<div>for Expanse to simfactory. Expanse is an example of</div>
<div>the new generation of AMD supercomputers. Others are</div>
<div>Anvil, one of the other new XSEDE machines, or Puma, </div>
<div>the newest cluster at The University of Arizona.</div>
<div><br>
</div>
<div>I have some experience with Puma and Expanse and</div>
<div>I would like to share some thoughts, some of which come</div>
<div>from interacting with the admins of Expanse. The problem</div>
<div>is that I am finding terrible multi-node performance on
both </div>
<div>these machines, and I don't know if this will be a common</div>
<div>thread among new AMD clusters. </div>
<div><br>
</div>
<div>These supercomputers have similar characteristics.<br>
</div>
<div><br>
</div>
<div>First, they have very high cores/node count (typically </div>
<div>128/node) but low memory per core (typically 2 GB / core).</div>
<div>In these conditions, it is very easy to have a job killed
by </div>
<div>the OOM daemon. My suspicion is that it is rank 0 that </div>
<div>goes out of memory, and the entire run is aborted.</div>
<div><br>
</div>
<div>Second, depending on the MPI implementation, MPI collective</div>
<div>operations can be extremely expensive. I was told that</div>
<div>the best implementation is mvapich 2.3.6 (at the moment).</div>
<div>This seems to be due to the high core count.</div>
<div><br>
</div>
<div>I found that the code does not scale well. This is
possibly </div>
<div>related to the previous point. If your job can fit on a
single node, </div>
<div>it will run wonderfully. However, when you perform the
same </div>
<div>simulation on two nodes, the code will actually be slower. </div>
<div>This indicates that there's no strong scaling at all from </div>
<div>1 node to 2 (128 to 256 cores, or 32 to 64 MPI ranks).</div>
<div>Using mvapich 2.3.6 improves the situation, but it is still</div>
<div>faster to use fewer nodes. </div>
<div><br>
</div>
<div>(My benchmark is a par file I've tested extensively on
Frontera)<br>
</div>
<div><br>
</div>
<div>I am working with Expanse's support staff to see what we
can</div>
<div>do, but I wonder if anyone has had a positive experience
with </div>
<div>this architecture and has some tips to share.</div>
<div>
<div><br>
</div>
</div>
<div>Gabriele</div>
<div><br>
</div>
</div>
<br>
<fieldset class="mimeAttachmentHeader"></fieldset>
<pre class="moz-quote-pre" wrap="">_______________________________________________
Users mailing list
<a class="moz-txt-link-abbreviated" href="mailto:Users@einsteintoolkit.org">Users@einsteintoolkit.org</a>
<a class="moz-txt-link-freetext" href="http://lists.einsteintoolkit.org/mailman/listinfo/users">http://lists.einsteintoolkit.org/mailman/listinfo/users</a>
</pre>
</blockquote>
<p><br>
</p>
</body>
</html>