[Halld-mc] [GlueX] computing power required

[Blake Leverington]

Hi Graham,

I'm currently using a computing cluster at Indiana University. Looking 
at the output of /proc/cpuinfo (attached as cpuinfo.txt) it's a duel 
core 1.6 GHz AMD Athlon. So not a new system, but reasonable. I assume 
newer systems are a factor of 2 or so faster. I'm not sure what GlueX 
numbers are based on but those should be taken as the more official 
numbers.  I'm running PYTHIA events which has a cross section of 124ub. 
These have something like 3-15 particles per event so it takes a bit of 
time  per event to simulate and analyze. The 1.5 seconds is also a sum 
of time for 5 different programs I have to run for generating, 
filtering, simulating, pruning and analyzing and maybe a 25% 
overestimate. 1.2 sec/event is what the logs file times add up to.

Regards,
Blake Leverington




Graham Heyes wrote:
> Blake,
>     I am the computing coordinator for Physics division at JLab. What 
> performance of CPU gives you 1.5 seconds per event? I am concerned 
> because the numbers I have for GLUEX simulation are much less than a 
> second per event and I need to calculate the computing requirements 
> for 2012 through 2015. If the seconds per event is out be an order of 
> magnitude it's a big deal.  On the other hand if you are benchmarking 
> on slow machines then I may be able to help. I have just got funding 
> to add another 34 compute nodes to the JLab cluster. This doesn't 
> sound like a lot but each node will be powerful enough to run 20 
> analysis jobs. So that's 680 more parallel jobs. The current cluster 
> runs about 1200 parallel jobs.
>
>     Regards,
>         Graham Heyes
>
>
> On Oct 8, 2009, at 6:13 pm, Blake Leverington wrote:
>
>> Hi everyone,
>>
>> I am attempting to simulate a week's worth of experimental runtime,
>> which means a background data set and whatever signal I might come up
>> with. What this works out to is 5.5x10^7 events (after copious
>> filtering) that I need to run through HDGEANT and my analysis. At an
>> average of 1.5 seconds to simulate and analyze each event, this works
>> out to about 1000 days of CPU time. Now, since I would like to finish my
>> PhD long before then, I was wondering if anyone had a spare computing
>> cluster or 10 that I could use for a week or more to cut my simulations
>> down to a reasonable time. Including repeats for crashes, errors, 
>> bugs, etc.
>>
>> Anyone have anything available?
>>
>> Cheers,
>> -Blake Leverington
>> _______________________________________________
>> GlueX mailing list
>> GlueX at dustbunny.physics.indiana.edu
>> http://dustbunny.physics.indiana.edu/mailman/listinfo/gluex
>
>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: cpuinfo.txt
Url: https://mailman.jlab.org/pipermail/halld-mc/attachments/20091009/b24c5637/attachment.txt