There is a future for high-performance parallel computing out there

Although high-performance parallel computing needs to be redefined, Tony Hey, head of the Department of Electronics and Computer Science at the University of Southampton, is rather optimistic about the future. Some time ago, he gave a presentation at the Fortran Future's conference where there were a few hundred enthousiastic people. Also this EuroPar conference attracted a large number of attendees for a wide range of topis.

Hey sees however a few serious problems that have to be tackled. For instance, maintaining software is much too difficult. There are too many versions of software and operating systems variants. Application software providers must for instance have versions for HPF, Fortran 90, MPI, etc. The market is, however, much too small to support all that.

Programming high-end parallel machines, like for instance the ASCI, is very difficult. There is no one programming style for all machines with thousands of processors of that class and all the applications that works.

On the other hand, Hey noticed that there are high-level programming interfaces like Mathlab/Mathematica that do not require programmming in for instance Fortran, but that allow the user to specify and solve his mathematical problem in some high-level way. For many technical computing applications that is sufficient.

Hence, Hey claims, that you need parallel computing only for a small range of applications. There are the real Grand Challenge applications that require an enourmous amount of computing time, and there are problems that require a vaste amount of memory, that can only be supplied by a parallel computer.

The high-performance computing field also has to live with its second in the class status. Gone are the days when programming supercomputers and parallell computers were the hottest thing on IT earth. Distributed networked computing, Java, and JavaBeans, that is where the attention goes today. Supercomputers are no longer the flag ships of computer companies. In fact, it took Hey quite some time to find the SP supercomputer on the IBM web site. It is not on the front-page anymore but hidden deep-down.

But when we get used to the idea that the high-end field is a small area, but still interesting, we can look for realistic models that will help us in the future.

Hey sees a future for scientific programming, although, for instance, the HPF and Fortran compiler market never will be large.

As a good example of what can be done, Hey mentioned, Promenvir, an industrial metacomputing example. They distributed their problem to a number of systems all over Europe and were able to get a speed-up of over 100. He believes that for a limited number of applications these distributed metacomputing systems will be a good solution.

Another example is the Toolshed project. Industry has a whole design cycle of which simulation is only a small part. It is a bit like Amdahl's law when you reduce the simulation time to zero, you still have all the other stuff. Toolshed was linkeing the designer to the simulation. So it was all transparant for him. The whole system was thus "deskilled" and much easier to use by the designer.

So these interesting projects do show that high-performance computing has a future.

One problem you come across when you want to run on a number of machines all over Europe is licenses. You do need licenses of the software on all machines. But you do not want to buy licenses for say six months for all the machines. Hey would like to have the software providers to come up with some scheme for which they only charge for the time their licenses are used: "pay-as-you-run".

"So I am not all doom and gloom", said Hey, "but have to see whether we can come up with some vision of parallel processing. But if you use some combination of data parallelism and task parallelism as distributed objectives and wrap up the Fortran code in Java or Corba, then I believe there is a future in there. A future not only for the intellectual field but also for the companies. The market will have to address workstations, metacomputing clusters and parallel machines. We should aim at high level problem solving environments."