[parsec-users] Question about # of threads for ferret-parallel

kishore kumar kishoreguptaos at gmail.com
Wed Feb 9 16:58:09 EST 2011


Hi Matt,

As far as I have seen, the *good range* (that gives max performance)
depends on the number of cores on your machine. Actually ferret is a
scalable application, I mean that it gives max performance when the
total number of threads created is more than the number of cores of
the machine (performance increases along with increasing thread counts).
According to the implementation, it creates (n * 4) + 3  threads with the
command line option value of n. On my 24-core AMD Opteron Solaris
machine, it gives best performance with a total of 63 threads
(corresponding command-line parameter value n is 15).

Best,
Kishore




On 9 February 2011 13:46, Matt Sinclair <msinclair at wisc.edu> wrote:

> Hi everyone,
>
> I have a question on the default run parameters for ferret.  I can see
> from the ferret-parallel.c program that several of the stages take
> their number of threads from the input parameters (parameter #6 to be
> specific).  However, in running the default tests:
>
> parsecmgmt -a run -p ferret -c gcc-pthreads -i [test, native]
>
> I see that parameter #6 is always set to 1 (for test and native at
> least), which implies to me that the number of threads in the default
> tests is always 1.  I realize that you can change this on the command
> line with the -n option, but I was wondering if this was intentional?
> If so, is it written somewhere what a good "range" of threads for
> ferret is?  Thanks again for the help!
>
> Regards,
> Matt Sinclair
> University of Wisconsin-Madison
> College of Engineering, ECE Dept.
> Vertical Research Group
> cs.wisc.edu/~sinclair <http://cs.wisc.edu/%7Esinclair>
> _______________________________________________
> parsec-users mailing list
> parsec-users at lists.cs.princeton.edu
> https://lists.cs.princeton.edu/mailman/listinfo/parsec-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.cs.princeton.edu/pipermail/parsec-users/attachments/20110209/1fab4794/attachment.html>


More information about the parsec-users mailing list