Following the first related workshop some progress has been made in HPC system software and tools. Two important standards for programming languages for high performance computing systems have emerged through community led forums. MPI provides a unified message passing library that integrates ensembles of independent processing elements from clusters of heterogeneous systems to tightly coupled arrays of like processors. HPF provides a high level data parallel programming model suitable for many applications in the sciences and engineering while hiding much of the complexity of managing the underlying resources. Both programming systems are being evaluated by the HPC to determine their potential long term impact for portability and programmability.
Two successful multi-agency collaborations have been conducted in response to the findings of the first workshop. The Joint NSF-NASA Initiative in Evaluation or JNNIE engaged researchers from many parts of the country and across computational science disciplines to analyze the performance and behavior of HPC systems under diverse real-world workloads. Pablo is an important development in performance profiling of parallel computing which was co-sponsored by ARPA and the NSF.
Although not an accomplishment in the area of HPC system software, it is important to note the emergence of SMP (symmetric multiprocessor) systems as a major new platform type. These systems evolved from the workstation community and integrate a small number of processors in a tightly coupled shared memory configuration. The significance of this trend is that many ISVs are targeting for the first time a class of parallel systems on which to host their products. This introduces a rapidly growing parallel program development industry unprecedented in the history of computing, presenting an opportunity for parallel HPC systems to leverage this investment.
Another trend since the first workshop that will impact HPC system software is the increasing use of workstation clusters which are finding productive application by industry for real workloads and are being treated as a low cost entry point into very coarse grain parallel computation. Such systems have already been a major factor in the success of de facto standards in message passing such as MPI and PVM.
Heterogeneous computing is emerging as a way to utilize the NII and HPC. In practical terms, this emerging field is still in its infancy and considerable basic infrastructure remains to be developed.