<div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Thu, May 7, 2015 at 1:25 PM, Jed Brown <span dir="ltr"><<a href="mailto:jed@jedbrown.org" target="_blank">jed@jedbrown.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><span class="">Matthew Knepley <<a href="mailto:knepley@gmail.com">knepley@gmail.com</a>> writes:<br>
<br>
> On Thu, May 7, 2015 at 9:23 AM, Justin Chang <<a href="mailto:jychang48@gmail.com">jychang48@gmail.com</a>> wrote:<br>
><br>
>> So to summarize, if I understand everything, I should do the following:<br>
>><br>
>> 1) calculate the flop/byte ratio for various problem sizes and solver<br>
>> methods on one process and:<br>
<br>
</span>It is impossible to define "arithmetic intensity" (flops/byte) without<br>
selecting a cache model. And my "cache model", I include register<br>
reuse. Assuming no reuse at all is not even close to representative<br>
(off by orders of magnitude).</blockquote><div><br></div><div>You are making assumptions about the algorithm here. Pure streaming computations,</div><div>like VecAXPY do not depend on the cache model.</div><div><br></div><div> Matt</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><span class=""><br>
> Should be roughly invariant to problem size.<br>
<br>
</span>Except (a) insofar as it depends on cache sizes and (b) the algorithm<br>
characteristics depend on data size (e.g., the cost of supernode<br>
factorization for sparse direct solvers scales superlinearly, so<br>
arithmetic intensity will drift as you increase the problem size).<br>
<br>
<br>
You can see how we compare "pessimal" versus "perfect" caching here:<br>
<br>
<a href="http://59A2.org/na/MayBrownLePourhiet-pTatin3d-2014.pdf" target="_blank">http://59A2.org/na/MayBrownLePourhiet-pTatin3d-2014.pdf</a><br>
</blockquote></div><br><br clear="all"><div><br></div>-- <br><div class="gmail_signature">What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.<br>-- Norbert Wiener</div>
</div></div>