SC98 - Echos from SC '98: The future isn't what it used to be

The 1998 edition of the High Performance Computing & Networking Trade Fair, SC '98, took place from the 9th to the 13rd November in Orlando (Florida, United States). This was the tenth anniversary of the conference that brings together several thousand people involved in high performance computing and communications each year. According to all participants, SC '98 was of very high quality. Here are some impressions about it.

A weak European presence

Germany was truly the only European country to be represented, on the one hand with a large HLRS stand (Federal Centre for High Performance Calculation, a partnership between Stuttgart University and industrial groups such as Daimler-Benz), and on the other hand with software companies such as Pallas or Genias on individual or shared stands.

The HLRS-lead a European Networking Demo stand displayed the results of technology transfers achieved in particular within the framework of TTNs (Esprit-HPCN program).

HPC is no longer synonymous with calculation

At the 'industry' exhibition, computer manufacturers were represented on 17 stands whilst suppliers for storage solutions and products for I/O intensive applications held 13 stands and networks occupied nearly 10 stands (equipment, operators). During the conference itself a number of sessions were devoted to issues such as Parallel I/O, Networks, Data Intensive Applications, Digital Libraries, Digital Entertainment, ?

What does the future hold for HPC?

Everybody felt that the last decade had been dominated by the transition from vector to parallelism, from proprietary to CMOS technologies (and mainly to the great use of "Off the shelve components"), and by a strong integration of components. However two major technological problems still remain: memory bandwith and the lack of development environments for applications. The next few years may witness a continued increase in component density, an evolution in languages (C, Java ), an increase in the use of networks (metacomputing, Grid based computing), and a greater choice of applications (with a growing market of management, transactional and decision-making applications ) Scalable clusters of shared memory modules will lead the market within the next 5 to 10 years. Scalability will have two dimensions : parallelism of shared memory nodes, and the number of nodes in a cluster. However the advent of multithreading may bring new possibilities.

Evolutions will depend on the market. Given that production investments are higher and higher (the building of the new IBM plant for components manufacturing may cost several billions dollars!), the use of bulk produced components will be more and more essential. The hundreds of millions of PCs and the tens of billions of embedded systems ( smart cards, mobile phones) will have a major influence on the evolution of the technology.

'Architecture research has not died yet', insisted the roundtable participants. Nevertheless, it is often presented as short-term research, thus dependent on short-term changes. Thomas Sterling (Caltech) heavily criticized academic research, as research paralysed by the fear of failure and suspicious of non-conventional ideas ('good engineering creates sloppy science').

Recent years have seen U.S. government funding of the HPC spent, in the main, on the ASCI program. The PITAC report http://www.ccic.gov/ac/interim/ gives some recommendations for the re-distribution of these funds. A strengthening of basic research is necessary in the fields of architectures (without focusing solely on TeraFlops but by taking new applications and then TeraOctets/s into account), of systems and software, of development environments.

How will HPC look in 10 years time? None of the speakers dared to answer precisely given that estimates are so unreliable. For Gary Smaby, 'The Future isn't what it used to be'!