[Swift-devel] coaster status summary

[Mihael Hategan]

One unfortunate part there is the lack of a decent Java MPI
implementation.

Another unfortunate part may be that MPI may not work very well with
processes that come and go dynamically, but I guess that can be
addressed in a way or another.

On Mon, 2008-04-07 at 07:35 -0500, Ian Foster wrote:
> I wonder whether we should be making use of MPI on the BG/P where we
> can ... I suspect that is what is optimized, rather than the IP stack.
> 
> Mihael Hategan wrote: 
> > On Mon, 2008-04-07 at 07:03 +0000, Ben Clifford wrote:
> >   
> > > there's an interesting issue here of how much we should (in our general 
> > > codebase, rather than in special customisations such as plugging in 
> > > falkon) be supporting 'wierd' systems such as the BG/P which have, 
> > > apparently, neither a decent shared filesystem or IP layer network fabric, 
> > > vs support for more traditional clusters which seem to have both IP layer 
> > > interconnect between nodes and decent shared filesystems.
> > >     
> > 
> > That is a good point. However, UDP was chosen to support systems with a
> > very large number of CPUs in the first place. In other words if UDP
> > won't work on BG/P, I don't see much reason for going with it.
> > 
> >   
> > 
> > _______________________________________________
> > Swift-devel mailing list
> > Swift-devel at ci.uchicago.edu
> > http://mail.ci.uchicago.edu/mailman/listinfo/swift-devel
> > 
> >