Re: MPICH-GM on heterogeneous clusters

Hi Valmor,

On Thu, 2003-07-24 at 15:02, Valmor de Almeida wrote:
>  How hard it is to develop MPICH-GM for a small 
>  heterogeneous environment; say 4 dual alpha nodes 
>  and 4 dual Xeon nodes connected by a Myrinet 2000
>  switch?

Not very hard, but not trivial either. It just requires a lot of
attention to the code and XDR. 

Actually, it depends of what you really need. If you want to translate
the payload, it would definitively affect performance and disable
zero-copy. If you just want to make the protocol works but you take care
of the payload endianness or structure packing, that's pretty easy and
the overhead would not be very high.

It's not on my TODO list, but it is on the MPICH(2)-MX list.

Patrick
-- 

Patrick Geoffray
Myricom, Inc.
http://www.myri.com