> It does apply, however, many parallel algorithms used today are
> naturally blocking. Why? Well, complicating your algorithm to overlap
> communication and computation rarely gives a benefit in practice. So
> anyone who's tried has likely become discouraged, and most people
> haven't even tried.
Don't phrase that as if it is just because of inherent application
limitations. Properly double buffer your ghost cells and setup your
nonblocking receives and you have an opportunity to smooth out some of
your load imbalance. Well, you would if many NICs gave true independent
progress. Many people who have tried and got discouraged did so because
their network couldn't give them advantage. Of course, there was that
one code on BG/L that they presented at SC a year or two ago where they
inserted "strategically placed MPI_Test calls" to really boost
application performance... wonder if that had anything to do with
needing to be in the MPI library to make progress...
Keith
--
Keith D. Underwood Scalable Computing Systems
Senior Member of Technical Staff Sandia National Laboratories
[EMAIL PROTECTED]
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
