[Beowulf] MPI Implementations for SMP use

[Joe Landman]

Hi Andy:

  Short version:  Use shared memory based devices (--comm=shared, or use 
ch_shmem for MPI that will not go outside of this box) for MPI on a 
single SMP node.  OpenMP is (IMO, and no, this is not intended to be 
troll bait) very easy to use/program with.  MPI is (also IMO, and also 
this is not intended to be troll bait) a bit harder to program with, 
but, and this is quite important, it forces you to think in distributed 
mode, which means that you typically are more focused upon the job of 
spatial localization of data, and data layout in general.  That stricter 
discipline is very helpful. 

  There will be people who take issue with both of these statements, and 
it is worth while to have a healthy discussion of the relative merits of 
each.  <self_promotion> I should be covering a little of this in a class 
I am going to teach next semester (see 
http://www.cs.wayne.edu/winter2005courses.htm# ) </self_promotion>. 

Joe

Andrew D. Fant wrote:

> This may be opening up a can of worms, but does anyone have any 
> information about the relative merits of the various open 
> implementations of MPI in SMP systems?  I'm setting up a Linux server 
> with multiple CPUs and I wanted to know if one implementation is 
> significantly faster than others under these conditions.
>
> Thanks,
>     Andy
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit 
> http://www.beowulf.org/mailman/listinfo/beowulf


-- 
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: landman at scalableinformatics.com
web  : http://www.scalableinformatics.com
phone: +1 734 612 4615