[parsec-users] the OpenMP version blackscholes has No speedup when has more than 4 cores

Eric Lu eirc.lew at gmail.com
Tue Oct 15 03:16:43 EDT 2013


I test the TBB version and the pthreads, both has normal speedup. The TBB
version data shown below:



              TBB

1            1.000481078

2            1.982362853

4            3.971750352

8            7.839389895

16          14.41142592

32          13.202854


On Tue, Oct 15, 2013 at 3:12 PM, Filippo Sironi <filippo.sironi at gmail.com>wrote:

> With the pthreads version I see much better speedups.
> Right now I cannot remove the serial part of the benchmark from the
> computation; however, if we consider the whole execution of the application
> what I get are the following numbers:
>
> cores   speedup (over the runtime)
> 1       1
> 2       1.71
> 4       2.87
> 8       4.29
> 16      5.75
> 32      6.77
>
> Filippo
>
> On Oct 15, 2013, at 6:58 AM, George Kurian <georgekurian.086 at gmail.com>
> wrote:
>
> > I would be interested to see if the pthreads version behaves differently
> from the OpenMP version. Do you observe the same behavior with the pthreads
> version? What confuses me is why the performance goes down from 2 to 4 to 8
> cores (considering blackscholes is embarrassingly parallel and even with
> initialization included).
> >
> > -George
> >
> >
> > On Mon, Oct 14, 2013 at 10:13 PM, Eric Lu <eirc.lew at gmail.com> wrote:
> > Yes, the performance means speedup, I run blackscholes with simlarge
> input not native inputs.
> >
> > I will read the paper you recommended.
> >
> > Thanks!
> >
> > Eric
> >
> >
> > On Tue, Oct 15, 2013 at 9:58 AM, kishore kumar <kishoreguptaos at gmail.com>
> wrote:
> > What do you mean by poor performance? Is it speedup? I guess that you
> run blackscholes with native inputs. With large inputs, blackscholes main
> thread does more work (serial execution -- reading input file, etc.,)
> compared to the worker threads (parallel execution). Around 95% of the
> entire execution time, you only see one thread, i.e., main thread. May be
> you want to read "Thread Reinforcer: Dynamically Determining Number of
> Threads" paper.
> >
> > Thanks.
> >
> > Best,
> > Kishore Kumar Pusukuri
> > http://www.cs.ucr.edu/~kishore
> >
> >
> > On Mon, Oct 14, 2013 at 6:34 PM, Eric Lu <eirc.lew at gmail.com> wrote:
> > Hi,
> >
> > I run PARSEC2.1 on Intel E7520 machine, the os is centos6.3, the gcc-4.7
> compiler. The machine has 32 cores in total, and the results seems
> blackscholes has too poor performance.
> > The data shows as follow:
> >       OpenMP
> > proc  blackscholes
> > 1     0.998269
> > 2     1.80499
> > 4     1.763301
> > 8     1.550916
> > 16    1.542871
> > 32    2.508076
> >
> >
> >
> > Best Regards!
> >
> >
> > Eric
> >
> > _______________________________________________
> > parsec-users mailing list
> > parsec-users at lists.cs.princeton.edu
> > https://lists.cs.princeton.edu/mailman/listinfo/parsec-users
> >
> >
> >
> > _______________________________________________
> > parsec-users mailing list
> > parsec-users at lists.cs.princeton.edu
> > https://lists.cs.princeton.edu/mailman/listinfo/parsec-users
> >
> >
> >
> > _______________________________________________
> > parsec-users mailing list
> > parsec-users at lists.cs.princeton.edu
> > https://lists.cs.princeton.edu/mailman/listinfo/parsec-users
> >
> >
> >
> >
> > --
> > George Kurian
> > 5th year Graduate Student,
> > Computer Science and Artificial Intelligence Laboratory (CSAIL),
> > Massachusetts Institute of Technology.
> > _______________________________________________
> > parsec-users mailing list
> > parsec-users at lists.cs.princeton.edu
> > https://lists.cs.princeton.edu/mailman/listinfo/parsec-users
>
> _______________________________________________
> 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/20131015/2afd588a/attachment.htm>


More information about the parsec-users mailing list