MCell & NPACi

MCell & NPACI -- Running in Parallel

    With support from NPACI centered at the San Diego Supercomputer Center, two means of parallelizing MCell are being pursued:

 
1. port MCell to the Cray T3E and IBM SP massively parallel supercomputer architecture.
2. develop an MCell/NetSolve client/server version of MCell so that heterogeneous clusters of hundreds of NetSolve machines may be used to run multiple MCell parameter space explorations in parallel.

    The port of MCell to the Cray T3E and IBM SP utilizes the MPI and PVM3 message passing protocols using the KeLp library and, when completed, will allow a single MCell job to split the heavy computational load of its diffusion algorithm as well as its large memory requirements evenly across any number of multiple processors on the machine.  In addition, using the MPI and PVM3 standards will make this parallel version of MCell highly portable to other parallel architectures, including heterogeneous clusters of machines.
    NetSolve is an alternative to MPI and PVM3 that makes it simple to turn a loosely associated collection of machines into a fault-tolerant client/server compute cluster.  We have begun working very closely with the authors of NetSolve (Jack Dongarra & Henri Casanova) to create an MCell/NetSolve client/server application.  This application will allow submission of perhaps thousands of MCell jobs to the cluster for automatic, fault-tolerant, concurrent execution.  During our initial tests jobs were run concurrently on 40 machines.  It is very inefficient to transmit large input files (e.g. 3D reconstruction information) repeatedly to the distributed compute cluster.  Future advancements in NetSolve will include a distributed file caching mechanism to solve this problem.  The MCell/NetSolve client/server will also be more closely integrated with the metacomputing tools AppLeS and NWS being developed by Francine Berman and Rich Wolski.