The POST Project - Industrial Relevance

Powerful multiprocessor workstations with shared memory (SMPs) are increasingly deployed in industry as stand-alone systems, nodes in workstation clusters, or as nodes of large-scale parallel computers. Some recent large-scale distributed memory systems, notably SGI's Origin 2000, have such low latency that moderately sized configurations may be programmed as if they were shared memory systems. The strong current growth of the SMP market is predicted to accelerate in the next years.

Partners of the consortium are committed to furthering the use of industry standards for high-performance programming, and have contributed to the development of such standards, in particular MPI and HPF , and have provided tools based upon them. They have been involved in a range of previous or on-going projects at national and international levels which develop and deploy MPI or HPF applications, including ESPRIT projects FITS, FRAME, FLOAT, PHAROS, PPPE and EUROPORT.

OpenMP

There have been many extensions to existing languages, notably FORTRAN and C, for shared memory parallel programming; they were, however, vendor-specific. An early attempt to define a standard in this area, PCF, was abandoned after several years of work without adoption, since at that time industry focused on distributed memory computing. OpenMP has been defined as a de facto standard for shared memory portable programming by an influential group of computer vendors in the US, including Compaq, DEC, IBM, Intel and SGI/Cray. OpenMP is suitable for application development on SMPs and low-latency NUMA architectures; it is similar to PCF, although it provides for a coarser granularity of parallelism. It has rapidly been endorsed by software vendors and tool providers alike.

A portable directive set has also been standardised by HPF, but experience shows that with current compilers, performance is only achieved on structured data applications, and parallelising just one part of an application is not easy. As OpenMP targets shared memory systems, incremental parallelisation and parallelising applications with irregular data access patterns (unstructured meshes) is much easier. This effectively broadens the applicability of the new standard.

It is expected that OpenMP will be quickly adopted by a broad group of ISVs, for whom it promises a reduction in development investment and the ability to satisfy a large number of end user demands.