<html>
<head>
<meta http-equiv="Content-Type" content="text/html;
charset=windows-1252">
</head>
<body text="#000000" bgcolor="#FFFFFF">
<p><br>
</p>
<div class="moz-cite-prefix">On 4/10/20 10:19 PM, wrote:<br>
</div>
<blockquote type="cite"
cite="mid:a82c73d57a5d4a94972b7e15cc15e69d@MWHPR01MB2237.prod.exchangelabs.com">
<meta http-equiv="Content-Type" content="text/html;
charset=windows-1252">
<style type="text/css" style="display:none;"> P {margin-top:0;margin-bottom:0;} </style>
<div style="font-family: Calibri, Arial, Helvetica, sans-serif;
font-size: 12pt; color: rgb(0, 0, 0);">
Dr. Pounds,</div>
<div style="font-family: Calibri, Arial, Helvetica, sans-serif;
font-size: 12pt; color: rgb(0, 0, 0);">
<br>
</div>
<div style="font-family: Calibri, Arial, Helvetica, sans-serif;
font-size: 12pt; color: rgb(0, 0, 0);">
Am I being stupid by recording my matrix timings from 1000-8000
with 1000 increments? Everything is queued up and expected to
finish Sunday morning, but if a smaller range is acceptable
please let me know. I don't want to hog hammer, and if a smaller
range is acceptable I think everyone would benefit from an email
with the range specifications. </div>
<div style="font-family: Calibri, Arial, Helvetica, sans-serif;
font-size: 12pt; color: rgb(0, 0, 0);">
<br>
</div>
<div style="font-family: Calibri, Arial, Helvetica, sans-serif;
font-size: 12pt; color: rgb(0, 0, 0);">
Just so you know 1000-10000 takes 3 hours, and 1000-8000 only
takes 1.5 hours. Smaller than that might be better, but it's up
to you. Thanks,</div>
<div style="font-family: Calibri, Arial, Helvetica, sans-serif;
font-size: 12pt; color: rgb(0, 0, 0);">
<br>
</div>
<div style="font-family: Calibri, Arial, Helvetica, sans-serif;
font-size: 12pt; color: rgb(0, 0, 0);"><br>
</div>
</blockquote>
<p><br>
</p>
<p>Timing takes time. Period. Now, as we have seen over and over
again, most of your codes reached a "leveling off" on the number
of megaflops and with the matrix size. If you can demonstrate
that memory-to-cache bandwidth is saturated and not changing (like
with a plot of megaflops vs. matrix size) then you have
demonstrated that it is safe to use matrices of smaller dimensions
in your testing. Recognize that you will have to do this type of
plot using different different numbers of processors. We have
already demonstrated numerous times that on our particular
hardware there is no real benefit to using more than half of the
processors (one of your BabyBLAS graphs will prove that). So on
hammer for this particular test I would run tests using 4, 8, 12,
16, 24 processor with each test using a fixed number of tests
producing a different curve on your graph.<br>
</p>
<p><br>
</p>
<p>If I see that everyone -- and I mean everyone -- has at least
done some timing on hammer by Sunday then I can extend the
deadline by a few days to give the class sufficient time to do all
the timings. I anticipate the queues on hammer being very full.</p>
<p><br>
</p>
<pre class="moz-signature" cols="72">--
Andrew J. Pounds, Ph.D. (<a class="moz-txt-link-abbreviated" href="mailto:pounds_aj@mercer.edu">pounds_aj@mercer.edu</a>)
Professor of Chemistry and Computer Science
Mercer University, Macon, GA 31207 (478) 301-5627
<a class="moz-txt-link-freetext" href="http://faculty.mercer.edu/pounds_aj">http://faculty.mercer.edu/pounds_aj</a>
</pre>
</body>
</html>