multi-cluster MPI support exists for the grid. mpich-g2 provides this
for pre-ws gram (gram2) installations. MPIg is the more recent
version and provides this functionality for both pre-ws gram and ws
gram (gram4).
Information on MPICH-G2 is here:
http://www3.niu.edu/mpi/
Information on MPIg is here:
ftp://ftp.cs.niu.edu/pub/karonis/MPIg/index.html
Work has been done in these implementations (but maybe only in MPIg,
not sure) to optimize the inter process message communication given
the network topology. e.g. use GEthernet when available between 2
processes. Process broadcast messages in an efficient manner.
Nick Karonis (cc'ed) could help you with any additional questions.
-Stu
On Feb 25, 2009, at Feb 25, 11:35 PM, Le Trung Kien wrote:
Hi,
I have a question about the ability of deploying parallel
application on a large grid system and
optimizing the use of available resources for sure.
We saw that a "classic" parallel application which worked well on
clustering systems, which have
their best communicating condition between cluster nodes (with speed
of Fabric or GEthernet).
And now, when the grid computing is widely developed and deployed,
the old clusters have been
joined in grids. For simple, we consider a cluster of computer as a
grid node; so, in general, I wonder if
one such classic parallel application could be executed on two or
more grid nodes concurrently
and give us the best performance in time consuming, assuming grid
nodes communicate with
others through DSL line (about 1 or 2 Mb/s of speed).
Thank you, any ideas please.
Best Regards.
--
Le Trung Kien.
