By William Fellows
Parallel programming pioneer Kuck & Associates Inc is working quietly on a set extensions to the OpenMP interface that it says will enable commercial software programmers to take advantage of what’s currently a mechanism for writing scientific and technical applications. OpenMP, a cross-industry programming interface specification, is designed to bring the performance advantages of traditional parallel programming techniques to shared memory systems. OpenMP was driven by the increased scaling of SMP systems, and the increasing use of standard SMP building-block system technology, such as Intel Corp’s SHV quads. Currently, OpenMP programs run in loops that can be decomposed. The Kuck work enables a succession of processes to be scheduled and queued. It’s going to take a couple of years to get the work to the OpenMP committee, Kuck thinks. Kuck, which is also supplying technology to Edinburgh Portable Compilers Ltd, one of the companies creating compilers that take advantage of the Intel IA- 64 architecture’s EPIC parallel execution functionality, says it’s now clear that EPIC optimization will take place after OpenMP parallelization has occurred. Meanwhile, the OpenMP bandwagon will gather more momentum at next week’s Supercomputing 98 show in Orlando, Florida this week when the first C and C++ implementations of the specification are made public. Until now it’s been a Fortran-only affair. The likes of Fluent, Ansys and LSTC already offer Fortran programming tools and compilers, and C and C++ work is expected to follow in short order. Although OpenMP is designed primarily to take advantage of processors within a shared memory system – MPI message passing and Posix threads are used to connect clustered or distributed systems – a group of academics are currently experimenting with layering OpenMP on top of clustering software protocols to link distributed systems, in a project known as Odin. Hewlett-Packard Co and IBM Corp are expected to demo OpenMP work at next week’s show.
