MPI on the SCC

We are certainly interested in having MPI on the SCC. There are some programs that we want to run, we actually use a modified version of RCCE to mock up some MPI like functions that we wanted and having a standard MPI library would really help us.

We would also love to have access to the MPI library.

We are interested in MPI on the SCC. Having a standard  MPI library for SCC and subcommunity will be grate.

Thanks.

We are also interested in having standard MPI on SCC and a related community. Indeed we have few MPI applications in which we are experts and we want to use it as benchmarks.

We would also appreciate the availability of a standard MPI library for the SCC. The MPI like wrapper functions we currently use are based on RCCE, RCCE_comm and iRCCE. A package with less dependencies would be nice.

We would also be interested in having access to the MPI .

The issue of MPI on SCC comes up from time to time. 

rckmpb (which will be described in detail in a soon to be published paper in ACM OS review) implements the data link layer of the TCP/IP stack.  The result is that anything that uses TCP/IP will just work on SCC.  That includes MPI.

So ... take you MPI of choice and rebuild it from source.  It will just work.  OpenMPI and MPIch just work if you configure and build them for SCC.

Now if I had lots of time, I would do this and put it in a public place where anyone could grab it.  I really would like to do that ... but I just don't have time.  Maybe I will in a few weeks when my current  batch of travel is done, but I can't make any promises.  But seriously ... anyone interested in MPI could do this on your own.

Completely orthogonally to the issue of building MPI is how rckmpb is configured.  I don't know if how to change the rckmpb confiuration is documented.  Maybe Ted could comment on that?

And if its not documented, I will go ahead and make my notes on rckmpb configuration available somewhere (I'm sure Ted can tell me where I could put these notes).

--Tim

Indeed, MPI has been possible on the SCC for a long time with the use of TCP/IP.  I have tested both MPICH2 and Open MPI, and I can confirm they work flawlessly.

However, Ted is referring to an MPI implementation that is customized for the SCC, and does away with the TCP/IP overhead.

Ted, we are also very interested in RCKMPI.

I'm currently installing the RCKMPI which can be found in the public svn repository. Does someone know where I can get the prereqs.tar.gz the README.RCKMPI is referring to? Thanks.

Hello,

you can donwload prereqs.tar.gz from: http://marcbug.scc-dc.com/svn/repository/tarballs/.

Tasos

Is this full MPI-1 or MPI-2?  Either way, I am very much interested in using it.

Jeff, it is de-facto MPICH2 with additional CH3 devices for the SCC, so you can (theoretically) use MPI-1 and MPI-2 functionality.

I'm looking forward to seeing compliance with MPI-2 validated to the extent possible with the MPICH2 test suite.  There is not a single implementation of MPI that is fully compliant with the MPI-2 standard.  Every implementation has bugs and there is no set of tests capable of verifying full compliance with the complete MPI-2 standard.