next up previous
Next: Irregular Problems in Up: Parallel Cluster Algorithms Previous: ``Hybrid'' Parallelism

Message Passing vs Data Parallel

Coarse grain parallel algorithms using message passing on MIMD machines give excellent performance for regular algorithms such as Metropolis, and fairly good performance for irregular cluster algorithms. This is because the efficient sequential algorithms can be used on the sub-domain on each processor.

Data parallel languages such as High Performance Fortran handle regular algorithms very efficiently, and are much easier to program that explicit message passing languages. However they do not perform as well for irregular cluster algorithms. This is mainly because the slower SIMD algorithms must be used even for the domain within each processor. There are also load balancing problems.



Paul Coddington, Northeast Parallel Architectures Center at Syracuse University, paulc@npac.syr.edu