Parallel implementation of a molecular dynamics simulation program

We have taken a NIST molecular dynamics simulation program (md3), which was configured as a single sequential process running on a CRAY C90 vector supercomputer, and parallelized it to run in a distributed memory message passing environment. Since portability was a major concern during parallelization, we used the Message Passing Interface (MPI) standard. The features of MPI provide a basic set of interprocess communication primitives on many architectures. The parallel md3 program has two basic algorithms resulting in an MPMD (Multiple Program Multiple Data) structure, versus the more common SPMD (Single Program Multiple Data) structure, and has the potential to exploit heterogeneous processing. For any given number of nodes we have devised an equation to determine the initial node allocation among these multiple programs which yields near optimal load balance. We also dynamically manage the load balance between processes to correct for run time variations and to achieve better performance. We compare the performance of this MPMD parallel code run on a range of distributed memory machines (an IBM SP2, a cluster of Pentium Pros, and a cluster of SGI Indigo2s with the R10000 processor) against the original code performance on the Cray. In addition to better performance, the code on distributed memory machines offers an ability to scale the problem size based upon the combined memory size of the host systems.