Seti@Home optimized science apps and information
Optimized Seti@Home apps => Linux => Topic started by: khanna on 30 Apr 2007, 08:32:55 am
-
Does anyone have any instructions on building SETI on PowerPC Linux, with AltiVec support? I am able to build without AltiVec, but when I enable it, I get into trouble because the source expects to be on Mac OS X.
Thanks.
-
Hi khanna,
we're using ICC/IPP by Intel to compile our optimized apps. Because of this, PPC-Linux is not really supported from our side, because Intel tools work only on Intel CPUs, not IBM/Motorola ones ;)
Sorry, I can't help you.
Regards,
Simon.
-
Ok, thanks. Maybe you can answer a more general question for me about the SETI source code. How much time is spent on calculating FFTs in the code? Is that the most computationally intensive part? Or are there other parts of the code that are a bottleneck?
-
There are other parts, but FFTs are what most time is spent on.
Other large time consumers are pulse folding, PowerSpectrum, chirping and some summing/averaging functions.
HTH,
Simon.
-
Thanks. I need to run some tests, both for benchmarking and validity. Can you suggest any short but relevant workunits? I'm not quite patient enough to wait for several hours every time I made a minor change and want to test the outcome.
Thanks.
-
My recommendation for you - grab "live" WUs off any of your hosts, with a spread of ARs reflecting the most common ones.
There's a set of WUs we use to benchmark here, most of them were also grabbed off BOINC. The only modification I made was to reduce two parameters inside the WU files - they have an XML header, if you open them up in a text editor, you'll see.
There are two occurrences of the same setting, both need to be changed (only a few lines apart) - "<chirp_limit>x.xx</chirp_limit>". Reduce both by the same factor to reduce overall crunch time for the WU.
However, this makes exact performance projection difficult - for the most part, if you take the perceived benchmark speedup and use 50-60% of the value, that'll be what it does on BOINC itself. The only real or rather relevant benchmark is taking two identical systems and putting differing apps on them, then comparing times for similar AR WUs (RAC is also not a great tool for comparison).
Here's a link to a Linux benchmark script (http://lunatics.at/index.php?module=Downloads;sa=dlview;id=84). It includes our test WUs as well as the default i386 Linux app (which probably won't be of much use to you, just replace it with the stock PPC one). Simply edit the variables at the start of the script to suit your needs and run it.
HTH,
Simon.