On Tue, Dec 22, 2009 at 2:49 PM, Jon Harrop <jon@ffconsultancy.com> wrote:
> Sure but it is worth remembering that distributed parallelism across clusters
> is a tiny niche compared to multicores.

I think the balance is slightly different than this in
scientific/research computing (the original subject of this thread).
At least in the U.S., a ten-page proposal will garner at least a
million core-hours at a TeraGrid site for pretty much any non-crackpot
academic project, free of charge (the systems are funded by the
gov't). NSF graduate fellows get like 100k core-hours just to mess
around, no PI and no particular project proposal needed.

The infrastructure, documentation, and tech support at the TeraGrid
sites generally assume MPI-based jobs, and give you access to [tens
of] thousands of cores. That's a pretty compelling resource if the
alternative is the relatively measly theoretical speedup from a
multicore (8x, etc.).

Mike