(j3.2006) Nick's concerns about coarrays

[Aleksandar Donev]

On Wednesday 05 November 2008 11:17, Van Snyder wrote:

> Nick appears to have said "maybe only Cray can do it."
No, Nick is cautioning that it is very hard to implement coarrays, which rely 
on "true one-sided" communication, which I will call "good hardware" (e.g. 
communication-coprocessors, RMA, shared-memory), on what I call "bad 
hardware", which he calls "commodity clusters" (this includes some clusters 
with interconnects better than Ethernet). 

Cray is not the only vendor that makes and sells (tightly) integrated 
supercomputers, i.e., good hardware. Nick's concerns are not applicable to 
such machines. Most importantly, they are not applicable to shared memory 
machines, as most machines (multi-core, multi-cpu desktops) are and will be 
for a long time.

And yes, he is right that there are lots of commodity clusters out there 
administered for packs of wolves, running OS's that do not (really) target 
HPC architectures. I say, just don't run coarrays on those machines. No one 
is preventing anyone from running two-sided MPI and hand-tuning codes. It is 
well-outside of the compiler's job to make a code run fast on your hand-built 
machine, sorry!

Best,
Aleks

-- 
Aleksandar Donev, Ph.D.
Lawrence Postdoctoral Fellow @ Lawrence Livermore National Laboratory
High Performance Computational Materials Science and Chemistry
E-mail: donev1 at llnl.gov
Phone: (925) 424-6816  Fax: (925) 423-0785
Address: P.O.Box 808, L-367, Livermore, CA 94551-9900
Web: http://cherrypit.princeton.edu/donev