<div dir="ltr">A general idea of the kind of applications that run on Roadrunner can be seen in the SC'08 Initial report. The application described there is a neutron-transport model. The kind of the communication pattern seen in this application is described as a wavefront. For the Roadrunner it was ported from a pure MPI implementation. On Roadrunner the Opterons are used for communication and the compute intensive part is ported on to the PowerXCell.<br>
<br><a href="http://www.c3.lanl.gov/pal/publications/papers/barker08:SC08_Roadrunner.pdf">http://www.c3.lanl.gov/pal/publications/papers/barker08:SC08_Roadrunner.pdf</a><br></div><div class="gmail_extra"><br><br><div class="gmail_quote">
On Fri, Apr 5, 2013 at 9:18 AM, Vincent Diepeveen <span dir="ltr"><<a href="mailto:diep@xs4all.nl" target="_blank">diep@xs4all.nl</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
Note that as for nuke explosions i have no idea how those look like -<br>
maybe someone more knowledgeable wants to comment on that.<br>
<br>
As a total layman there i would suspect that it's important where the<br>
protons/neutrons/whatever-tons/supertiny-tons are located. I'd be<br>
modelling that naively using<br>
matrixcalculations.<br>
<br>
So that would mean the only low level library you need is a<br>
matrixcalculation and some relative simple functions - with the<br>
matrixcalculations<br>
eating 99% of all system time on that massive supercomputer out of<br>
all calculations done on it.<br>
<br>
In such case one would need surprisingly little very well optimized<br>
code to make optimal usage out of such massive supercomputer.<br>
<br>
Any other 'secret' batchjob i'd be running on a different<br>
supercomputer. If there is no need to run a massive vector oriented<br>
workload type matrixcalculation -<br>
one just shouldn't run on such type of supercomputer i feel. NASA<br>
still had that 10240 socket supercomputer back then if i remember well,<br>
to give one example...<br>
<div class="HOEnZb"><div class="h5"><br>
<br>
<br>
On Apr 4, 2013, at 5:39 PM, Max R. Dechantsreiter wrote:<br>
<br>
> Vincent,<br>
><br>
>> It doesn't matter whether you code for blue gene, cuda or phi - from<br>
>> a software viewpoint it's all vector type coding you've got to do.<br>
>> the price of 1 coder is total peanuts compared to<br>
>> the price of those supercomputers. So specialistic written software<br>
>> is what you need anyway.<br>
><br>
> Simply porting an application shouldn't take much effort,<br>
> as long as vendor-specific libraries aren't involved.<br>
> However, special and sometimes intensive efforts are often<br>
> required to achieve good performance (high utilization).<br>
><br>
> The unfortunate fact is that funds are more easily spent<br>
> on hardware than on the human resources needed to utilize<br>
> them effectively.<br>
><br>
> Max<br>
> ---<br>
> <a href="http://www.linkedin.in/in/benchmarking" target="_blank">http://www.linkedin.in/in/benchmarking</a><br>
<br>
_______________________________________________<br>
Beowulf mailing list, <a href="mailto:Beowulf@beowulf.org">Beowulf@beowulf.org</a> sponsored by Penguin Computing<br>
To change your subscription (digest mode or unsubscribe) visit <a href="http://www.beowulf.org/mailman/listinfo/beowulf" target="_blank">http://www.beowulf.org/mailman/listinfo/beowulf</a><br>
</div></div></blockquote></div><br></div>