Seti@Home optimized science apps and information
Optimized Seti@Home apps => Linux => Topic started by: sbalfour on 23 Jun 2009, 02:44:01 am
-
Ps3 (PPC 970 3.2GHZ PPE) has Whetstone 376; my aging Opteron 285 has Whetstone 1952 @2.4Ghz. Opteron 285
finished Seti MB in average 1.5hrs CPU time now. PS3 benchmarks only 5X slower than Opteron, but runs the app
16 times slower? I eliminated the obvious - CPU is 100% utilized, 97% for user, no I/O wait time. The Khanna uses
the SPE's for FFT. Maybe FFT isn't the cpu hog? How long did WUs take on the standard app? There's something
terribly wrong. Look at the performance of folding@home on PS3: 50-100 times faster than desktop CPU's. We
need to think big here; my gut says we're missing a factor of 50. No wonder PS3 owners are wiping their boots of us.
Stuart
BTW, boincmgr reports the PPE as 1 cpu, though it is a 2xSMT processor. YDL displays 2 penguins, so we've got OS
support. How do I fix that?
-
I found that performance on the PS3 left much to be desired.
If you get Boinc for Linux Power/PowerPC - 32 Bit from
http://www.dotsch.de/boinc/Home.html
then the PS3 will crunch 2 at a time but still very slow.
-
Actually, Dotsch's page is where I got the PS3 Seti MB app. It's the Khanna one. The info that accompanies the
download is to start your own boinc client process (not as a daemon) for each instance of the app (so start two).
But I already have a boinc daemon running. This machine has only 256MB of memory to run the hypervisor,
YDL OS, the boinc daemon, Seti app(s), Firefox, boincmgr, xterms, etc. So I shall not like to start process-level
boinc clients. It's not necessary. The Boinc manager UI recognizes Pentium/netburst hyperthreaded CPUs as
multiple, as well as Core I7 HT CPUs. So why not PPC 970? I think it's a bug in boincmgr. If it gets CPU info from
some init file, I might be able to work around it. But if it runs a program to query the CPU interactively, I'm hosed.
Stuart
-
The Platform string is what you need to look at when choosing which version Boinc. This is what is reported as the processor type to the project servers either PS3Grid or Seti.
Platform string powerpc64-ps3-linux-gnu which would be used by PS3Grid. 1 CPU
Platform string powerpc-linux-gnu for use on Seti. 2 CPU
Hope this is a better explanation. Also please note I am talking about Boinc version, not the version of the science app that actually crunches the work.
-
...
The Boinc manager UI recognizes Pentium/netburst hyperthreaded CPUs as
multiple, as well as Core I7 HT CPUs. So why not PPC 970? I think it's a bug in boincmgr. If it gets CPU info from some init file, I might be able to work around it. But if it runs a program to query the CPU interactively, I'm hosed.
Stuart
The BOINC core client always gets information from the OS if possible, though there are a few cases where they must go deeper. For finding the number of CPUs available in Linux, they use a sysconf(_SC_NPROCESSORS_ONLN) call. Whether there's some init you can modify so YDL will say you have 2 CPUs I don't know. I did run across a thread from 2002 (http://us.fixstars.com/lists/yellowdog-devel/March02/0000.html) saying at that time YDL was returning zero for that call, nothing more recent.
Joe