Hello,
thanx for ur ans. I don't have any parallel m/c and iam not able to configure the Lam on workstation clustors. Can I run the applciation on
dual proceesor which is maintained as a server for the thin clients here?
thanx in advance ..
bye.
On Tue, 29 Jun 2004 Jeff Squyres wrote :
>On Mon, 28 Jun 2004, Narayan D. G wrote:
>
>> if we run simple programs with single process (mpirun pgm -np 1) on SIGNGLE m/c, the prg takes less time. If I run with MANY processeson SINGLE m/c, it will take more time as compared to 1 process. Actually, the time should be less if we run with multiple proceses. Why it is so ?
>
>I'm not sure what "m/c" is, but I think I understand the gist of your question. This has been discussed on the list several times in the past -- you might want to search the list archives for more complete answers.
>
>The short version is that you are overloading your CPU and the OS has to keep switching between all your different processes. This adds considerable overhead as compared to when you are only running one process. You're likely also thrashing your different levels of cache, which can kill performance.
>
>To get speedup, you need to spread those multiple processes across multiple CPUs (and potentially multiple nodes).
>
>-- {+} Jeff Squyres
>{+} jsquyres@lam-mpi.org
>{+} http://www.lam-mpi.org/
>_______________________________________________
>This list is archived at http://www.lam-mpi.org/MailArchives/lam/
Hello,
thanx for ur ans. I don't have any parallel m/c and iam not able to configure the Lam on workstation clustors. Can I run the applciation on
dual proceesor which is maintained as a server for the thin clients here?
thanx in advance ..
bye.
On Tue, 29 Jun 2004 Jeff Squyres wrote :
>On Mon, 28 Jun 2004, Narayan D. G wrote:
>
>> if we run simple programs with single process (mpirun pgm -np 1) on SIGNGLE m/c, the prg takes less time. If I run with MANY processeson SINGLE m/c, it will take more time as compared to 1 process. Actually, the time should be less if we run with multiple proceses. Why it is so ?
>
>I'm not sure what "m/c" is, but I think I understand the gist of your question. This has been discussed on the list several times in the past -- you might want to search the list archives for more complete answers.
>
>The short version is that you are overloading your CPU and the OS has to keep switching between all your different processes. This adds considerable overhead as compared to when you are only running one process. You're likely also thrashing your different levels of cache, which can kill performance.
>
>To get speedup, you need to spread those multiple processes across multiple CPUs (and potentially multiple nodes).
>
>-- {+} Jeff Squyres
>{+} jsquyres_at_[hidden]
>{+} http://www.lam-mpi.org/
>_______________________________________________
>This list is archived at http://www.lam-mpi.org/MailArchives/lam/
|
