Features:

INTERVIEW WITH DAVE MORTON, TECHNICAL DIRECTOR, MHPCC
By Alan Beck, Editor-in-Chief, HPCwire

HPCwire: Thus far, the race between COTS-based cluster supercomputers and those based upon proprietary processors has resembled that of Achilles and the tortoise: the clusters approach ever nearer but never quite succeed in surpassing -- or indeed drawing even with -- their elite contenders. Will this situation ever change? Why or why not? And given the speeds involved, does it really matter?

DAVE MORTON: First off, I want you to know that these are my own views and opinions. They do not represent that of Northrop Grumman, MHPCC, Air Force Research Lab or the Air Force.

Good questions. In terms of COTS meeting or surpassing the more HPC custom designed processors, I don't think that it will happen in the immediate future. Designing and building a COTS microprocessor is such a compromise that there will be a niche that the custom designed hardware to fill. As chip feature size continues to fall, it is possible to implement almost any architectural and logic function on microprocessor. The limiting factor becomes chip I/O. Each pin you add on a microprocessor costs money and as COTS microprocessors are targeted at a large market that likely doesn't value bandwidth as much as the HPC world. In terms of whether or not it matters, it apparently does for many HPC customers. Several HPC vendors have been doing well financially as they exploit the high-end niche HPC market for custom hardware solutions. Their customers are voting with their pocketbooks that it does apparently matter to them.

HPC: I covered the immensely popular SC95 HPC Architecture Panel chaired by Bob Borchers. At that time, the fight was primarily between SMP, MPP and proprietary systems; yet, technological breakthroughs in networking soon catapulted cluster and Grid-based supercomputing into the spotlight. In your view, are there similar innovations, poised just over the horizon, that will galvanize an entirely new vision of supercomputing?

DM: Well, if I could really see into the future, I probably wouldn't be working for a living -- although my current job at MPHCC on Maui is a pretty nice gig. I do think that there are some real disruptive technologies on the horizon that will impact many of the companies and users in HPC today. The first of these is Dense Wavelength Division Multiplexing (DWDM) with fast tunable lasers. This technology is being driven by the networking world, but I think that it will soon cross over to the supercomputing world. With some of the designs you could envision a very wide fat pipe between nodes. With other designs you could have a non-blocking fiber optic ring network topology that has all the nodes only one "hop" away from any other node. The technology used is to use fast tunable lasers as transmitters and have each node's receiver only see a given "color" of the optical spectrum that goes through each node.

There also seems to be a burst forward in terms of system on a chip (SOC) technologies. Bandwidth on a chip is almost free compared to bandwidth off of a chip. IBM's Blue Gene product line is one case in point. A Blue Gene node consists of an ASIC and some memory parts. Processors in memory (PIM) also seem to be a technology whose time is coming. Lower costs and higher bandwidth are the driving forces and Moore's law is providing the logic and memory densities to make it happen.

None of these technologies are such that they will really galvanize the industry. They'll be disruptive and some companies will win and some will lose. Maybe something will come out of DARPA's supercomputing initiative that will really change the industry. Grid computing has the potential to really change the supercomputing landscape, but there are huge logistic problems that need to be resolved. Even with Grid computing, there are certain classes of problems that just won't run well in a distributed environment. Commercially successful optical or quantum computing could completely change the supercomputing world, but right now these seem to be more academic research projects as opposed to real working machines.

HPC: Will unconventional HPC systems, such as those built upon FPGAs, ever become popular supercomputing alternatives? Similarly, are some types of popular architectures doomed to the fate of the dinosaurs? Why or why not?

DM: I think that FPGA HPC systems will likely be always limited to certain applications such as signal processing and crypto work. While the potential performance improvements are impressive, the amount of work needed to implement a HPC algorithm in hardware is considerable. The compiler technology needed to automatically map generic HPC software to configurable hardware would require a ton of technical investments. In the meantime the COTS microprocessors and ASICS power and capabilities continue to march forward. I also don't see any multi vendor standards coming out of this effort so any solution you would buy locks you into a single vendor solution.

In terms of architectures dying off, the only constant is change. If there are any architectures out there that can't or won't change then they are dead. If an architecture is willing to migrate to embrace new technologies and change to keep up with the times then they will likely have some measure of success. The question is whether a given architecture can make the right choices in order to survive. Of coarse this all supposes that there won't be any major disruptive technologies like optical or quantum computing. If anything like this happens then all that I said above is wrong.

HPC: How has the growing importance of commercial HPC applications changed the face of supercomputing architecture? Do you see this as positive, negative or neutral -- and why?

DM: There have always been important commercial HPC applications, it is just that the focus has moved from FORTRAN compilers to analysis programs. I don't have a big opinion here, though I will say for any hardware purchase (whether it is a PC or supercomputer) that one should first decide what applications they want to run before they even start to look at hardware. In the sense that commercial application providers won't port their applications to a new architecture until it is proven and successful in the field, it can become a chicken and egg scenario for new architecture adoption.

HPC: Are there any other points regarding this topic that you would like our readers to understand or consider?

DM: I just want to say that I think that supercomputing is a great industry to be involved with. There are lots of smart and interesting people to work with, and it moves at a very rapid pace. It is also small enough industry that one organization (or person) can make a large impact.