Urbana-Champaign, Ill. -- To further understanding of significant trends in high performance computing at the juncture of a new year, HPCwire interviewed Larry Smarr, director of NCSA (National Center for Supercomputing Applications) about his view of the most important developments that have taken place in the industry during 1996 and those which might be anticipated in the 12 months ahead. Following are selected excerpts from that discussion.

What did 1996 mean for HPC?

It was an incredible year. It marks the end of the market consolidation and shake-out of the high performance computing field. We're now in a situation where there are no longer stand-alone supercomputer companies. In their place, are three market-driven companies -- SGI, IBM, and HP -- that have desktop to teraflop offerings. The fact that they are market-driven -- and that the base of their business model is desktop -- means that the future of HPC will be radically different than the past.

In the past, HPC really wasn't market-driven. Yes, there were commercial customers. But it really emerged from a defense contractor portion of the economy. The companies that emerged from that arena just weren't appropriate business units for the kind of fast-paced, quarter-driven world we now live in. I feel that's the biggest change in the fifty year history of our industry.

What have recent competitions like PACI, ASCI and the DOD Modernization Program meant for the industry as a whole?

These are exciting -- and important for the future of the country. We're seeing the end of a phase of the way the federal government has organized its HPC program in the field and the start of a new world. The defense modernization program is really colossal. They've taken something as gigantic as the Defense Dept., consolidated their expertise, and upgraded their hardware into four major shared resource centers. That's radically different from the past.

Any DOD service member can utilize any one of those sites from anywhere in the country. NSF Supercomputer Centers, including SDSC and NCSA, are organic parts of the modernization program for the first time. So there will be a real flow from the university sector into the defense sector. This presents a great opportunity for those of us in the university sector to get some notion of the information infrastructure requirements of the defense side.

This has already led to early successes -- for example, linking the ImmersaDesk at the Army Corps of Engineers Center in Vicksburg, Mississippi through crosslinking the DOD Interim DREN with the NSF vBNS to NCSA. And we'll be doing collaborative visual supercomputing linked to the Chesapeake Bay, an Army Corps Mission and also an area of interest on a grand challenge scale to the university sector. After fifty years of separation between these two sectors, civilian and defense, I see the partnerships being formed by the DOD Mod program as bringing the country back together.

In the Dept. of Energy, there's the ASCI program that will definitely set the bar on what HPC means for the forseeable future: it will define the 'bleeding-edge.' Again, there's a major consolidation, with Sandia focused on Intel-based systems, Livermore on IBM, and Los Alamos on SGI/Cray. The DOE has just announced their 'ASCI Alliances' program, a competition for centers of excellence in the universities to link directly to the national labs: a pulling back together of the civilian sector and the Defense Program's mission-critical portion of DOE.

NASA has been going through a very complex examination of its HPC mission. Although this has a way to go before completion, it's possible to see major changes in the way Goddard, NAS, and JPL carry out their mission.

Finally, the same thing is going on in the NSF sector with the PACI competition, where we may end up with fewer leading-edge centers, but they will each be coupled to a much more broadly-based set of computer and application scientists and education outreach professionals in the university system.

There are important recurring themes here across the federal HPC agencies. Virtual institutes are replacing inside-the-walls centers. There is a linking across mission agency barriers with the university sector to effect an integration of leading-edge research. Trends toward virtual teaming, collaboration, putting national 'dream-teams' together mark the start of a new age. It's very different from what we've seen in the past and it will bodes well for the country's future.

What key technical developments do you anticipate will emerge in 1997?

Two technical trends will dominate the high performance arena in 1997. The first will be the emergence of distributed shared memory (DSM) as the de-facto standard for HPC. Machines from HP and SGI will compete, and there will be unprecedented opportunities to do portability studies across vector shared memory machines, microprocessor shared memory machines, distributed shared memory machines, and MPPs. You'll see a move on the users' part toward things like HPF (High Performance Fortran) and MPI that allow them to move across these platforms.

We're at the end of the company shake-out. But we still have two to three years of product shake-out to come within the big three companies. Look at the SGI/Cray roadmap publicized during SC'96, where they intend to gradually consolidate over a five-year period MPP, vector, SMP, and DSM products. For that sort of transition to be smooth from the user point-of- view, there will have to be a tremendous amount of research and work from computer scientists, computational scientists and the vendors themselves.

HP (Hewlett-Packard) is doing some fascinating work with Intel and Microsoft on moving toward what could be in five years a teraflop Wintel computer with DSM. Although this won't materialize in 1997, you can see its threads emerging. That's one area.

The other big trend will be distributed computing. By that, I mean things like the Wisconsin Condor project -- in which you link up lots of user-owned workstations on a net to form a collective compute-pool, or a Condor flock as they call it -- which can be used to attack large-scale problems. At the other end of the spectrum the vBNS and similar technologies will lead to a much more active hooking-together of high performance computers at very high speeds, so that single problems can be run on a national scale.

What's looming on HPC's horizon?

There is another very important area of integration and consolidation: the emergence of the visual supercomputer. For years, people who have done supercomputing also realized the necessity for high performance scientific visualization. But that was always done sequentially on different machines, and the process was characterized by bottlenecks in I/O, intermediate disc storage, software, etc.

One of the exciting things now coming from the SGI Origin is the ability to integrate InfiniteReality Engines inside a DSM cache-coherent environment. That leads to the potential for creation of real-time detailed 3D maps of the arrays being employed for supercomputing, and the ability to distribute by digital video over the new high-speed networks directly into multimedia desktop workstations. Thus, multiple investigators in a collaborative framework could simultaneously link into this visual supercomputer. Such a technology would dramatically transform the way we do computational science and engineering. I think you'll see the first prototypes of it in 1997.

Alan Beck is editor in chief of HPCwire. This article is reproduced with their permission, and appeared in the January 10, 1997 issue.

For HPCwire subscription information, email sub@hpcwire.tgc.com. For a free 4-week HPCwire trial subscription, email trial@hpcwire.tgc.com.