<html>
<head>
<meta http-equiv="content-type" content="text/html; charset=UTF-8">
</head>
<body>
<p>In class I had you run mpi from the command line across random
nodes of your own choosing. I wanted you to see what happened
when you ran a calculation and potentially ran into others
possibly sharing the node and or network. Several of you got what
seemed to be random results with the high number of nodes actually
taking significantly longer.</p>
<p>By utilizing PBS/Torque "correctly", we are eliminating the
chance of multiple people using the same node and also -- if the
code is written correctly -- we will be minimizing the network
traffick. To help you understand this phenomena, I ran some test
jobs last night on the cluster using the both the "allcopy"
version and the improved mmm_mpi version of the code that only
passes portions of the matrices. I then graphed megaflops vs.
number of processors.</p>
<p><img moz-do-not-send="false"
src="cid:part1.vsyRqgh0.fBzu19Ek@mercer.edu" alt="plot"
width="640" height="480"></p>
<p><br>
</p>
<p>Notice what happens -- the allcopy version is faster when there
are fewer than 6 nodes -- but it scales POORLY and after using six
nodes the performance actually starts to decrease due to the
increase in network traffick. In contrast, the mpi_mmm code is
still increasing in performance up to 10 nodes. Recognize, this
graph will vary based on the dimension of the matrix -- but it
demonstrates a key component of distributed HPC computing --
processor power vs. network bandwidth and the necessity of
thinking about this when designing HPC algorithms.<br>
</p>
<p><br>
</p>
<div class="moz-signature">-- <br>
<b>Andrew J. Pounds, Ph.D.</b><br>
<i>Professor of Chemistry and Computer Science</i><br>
<i>Director of the Computational Science Program</i><br>
<i>Mercer University, Macon, GA, 31207 (478) 301-5627 </i></div>
</body>
</html>