[Beowulf] MPI Implementations for SMP use

[Bogdan Costescu]

On Fri, 17 Dec 2004, Greg Lindahl wrote:

> You would think that, but the nice thing about pure MPI is that
> locality is perfect. So in most cases, a pure MPI code beats a
> hybrid code.

Would this still apply to multi-core CPUs ? The mighty forces have all
decided that multi-core is the way of the future. I'm especially
concerned about resources that will be shared between the cores, like
memory interface, on-chip cache, etc.

MPI, even with shmem, requires at least one copy of the data from one 
rank to the other. Then each rank would access different memory areas, 
meaning that the on-chip cache data cannot be shared and the main 
memory interface will need to be used "more often".
On the contrary, threads could share the same data - so there is no 
copy between ranks - and with the right access pattern they could even 
share cached data.

-- 
Bogdan Costescu

IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
E-mail: Bogdan.Costescu at IWR.Uni-Heidelberg.De