Characteristics of HPC
Previous: Contents
The charter of Working Group 2 was centred on the characteristics of scientific and engineering applications and algorithms that require high-performance computer systems. In particular, the group tried to address the following points:
More broadly, the group's charter was to consider possible national initiatives in HPC software from an applications perspective and taking into account the findings of the 1st Pasadena Workshop [2]. The group's charter directed us to scientific and engineering applications although the group took a broader view and in fact a major recommendation of the group is to build HPC software from a broad end-user base with a viable commercial market. This implies that one needs to consider applications such as the National Challenges, as described in the HPCC National Coordination Office (NCO) ``Blue Books'' for 1994 and 1995 , to get an appropriate set of requirements and standards for HPC software.
Although the group focused on HPC, many of the issues and indeed some of the groups recommendations also apply to HPCC. The distinction between HPC and HPCC can be rather subtle depending upon the country and the market sector being considered. HPC is obviously a subset of HPCC and in the scientific and engineering market sector HPC is at least currently more important than the communications aspect of the second `C'.
Working Group 2 started its work with a set of short position papers presented by each of the (26) participants. These represented a broad spectrum of interests including academia (6), National Laboratories (9), Software and Systems Vendors (6), Government Agencies (5). Industrial research laboratories and the industrial user community were not directly represented, however. Applications expertise included: applied mathematics; astrophysics; biology; computer aided engineering; computational fluid dynamics; chemistry; medicine; real-time systems; physics; satellite data analysis and remote sensing and weather forecasting. Several Working Group 2 members had broad experience in a wide variety of applications outside this list. We have distilled comments from these position papers and the ensuing discussion into this present article.
The comments in the position papers of the individual Working Group members fell into two classes: relatively precise technical comments on the structure of applications such as their need for substantial I/O or or of adaptive irregular data structures; the second class of comments concerned the structure of the HPC enterprise and the nature of its evolution. This is illustrated by a concern that the HPC industry was widely perceived to be in financial difficulties and it was important to find ways to encourage progress in the field.
There was a lengthy discussion about HPC standards and software
models. Clearly standards such as HPF and MPI were deemed to be very
valuable but they have the characteristic of being motivated by the
parallel computing (HPC) world. They are consistent with mainstream
computing but not required by it. For example,a software developer on
a PC or workstation would use perhaps C, C++, Fortran 77 or even
Visual Basic and such software is not ``HPC compliant''. We
recommend that ``trickle-down'' strategies be complemented with a
``trickle-up'' standards and software model. For this, one starts
with a software model for an area of computing which appears to have a
solid fiscal and user base. We discuss some such areas in section
3.
Integration forums such as recent efforts for HPF and MPI are examples of the trickle-down phenomena. While they have had some impact, they have been driven largely by developers rather than end-users. Forums like these could use modern (Web based) collaborative technology to collect input from wider audiences from industry and from commercial users as well as from academia and developers. Other examples which could be the starting point of a ``trickle-up'' approach are the software for symmetric multiprocessors, distributed computing or the software used on the World Wide Web itself.
HPC Software models built on this broad base of user input may not be as optimal as something specialised for a narrow user base but this trickle-up philosophy may in the long-run have greater viability and vitality. This idea was embodied as a major recommendation in the group's presentations to the other workshop participants on both of the last two days. The material here represents an expansion of these two presentations.
Another important theme was the identification of viable enterprise models or ``Industry, Government & Academia business partnerships'' for HPC applications, tools and system software development. Our trickle-up strategy represents one such viable model whereas most of the others required varying amounts of significant government intervention and therefore resource investment. In particular, we also recommended that more attention be given to a pulse or seed funding activity similar to the EUROPORT model as another especially important enterprise model. Experience of this sort of matching-funding from the UK's Parallel Applications Programme (PAP) and from the European Union's EUROPORT programme indicates a relatively high uptake of HPC by industry and commercial organisations.
Characteristics of HPC
Previous: Contents