Europe can take the lead again in supercomputing - vectors can be the answer

Bristol, 03 December 98 Dutch scientists now happy working on their Cray C90 did not want to learn parallel processing during the past years. And for good reason. It is difficult, it is time consuming and for their class of problems does not work. They are not the only ones. In industry one finds that a lot of industrial codes just do not parallelize. Good old vectors could provide help here. Japanese vendors like NEC and Fujitsu have good machines out there that can bring a lot of performance with only a few processors. Europe could benefit from that. Since the US have effectively banned Japanese vector computers and do not have a real alternative of their own, Europe, which has long standing experience with these machines, could take of as the leading supercomputing continent. Instead of struggling to get thousands of processors to work as the poor Americans have to do in the Asci project, we could buy Japanese vector machines to do real application work with them. It may not be advanced HPC architectures but it can be used for advanced applications.

"The future is to parallel processing". This is the core adagium of the US government and the US hardware companies. This has lead to the Asci initiative and other highly parallel machines that populate the upper regions of the TOP500. To get a little more than one Tflop/s of Linpack performance Asci Red for instance needs more than 9,000 processors. There are, however, only few applications that really run efficiently on these type of massively parallel processors. Most day-to-day parallel applications scale well to 8, perhaps 16 or 32 processors. Especially large industrial codes, as those used for instance in structural analysis, do not parallelize well beyond that number of processors.

There is, however, an alternative in the parallel vector machines (PVP), such as the NEC SX-5, and the Fujitu's VPP series. The last large US vector machine, Cray T90 series was not a success and vectors were proclaimed dead, especially in the USA. Also in the TOP500, their share is steadily declining. So how could they be an alternative? PVP's were designed for number crunching. The vector processors are very well suited for the large HPC codes that are used in industry. More importantly, however, is the memory system. It is designed to get as much data as possible in and out of the processors in the shortest time. This results in an unmatched sustained performance. It is not so difficult to get 95%+ of the top speed out of a vector processor. With a PVP, three processors are enough to enter the TOP500. With an off-the shelf-processor, you need at least ten times as many. Most real world applications do not parallelize as well as the Linpack used in the TOP500, hence, there the use of PVPs is even more advantagous.

To get to the sustained Tflop/s range, a single CPU vector processsor will not do. You need something in the order of 128-256, depending on the application. Even these levels of parallelisation are already a tremendous task to achieve, but nothing compared to the effort one has to put in to keeping the thousands of processors of an MPP to work.

The PVPs are produced by large companies, NEC, Fujitsu and Hitachi, that are already in the business of supercomputing for over a decade. Hence, they probably will keep doing so for a while. This gives Europe the opportunity to buy these type of machines for demanding applications that really require a lot of computing power. Doing this will give it a competitive edge in applications, an area were it is still strong.

This does not mean there is no room for parallel processing, there is, especially in throughput computing it can be used efficiently. There is room for clusters of workstations, SMPs and even MPPs, but for production type high-end supercomputing, parallel vector processors provide a good option, especially in Europe were we are free to buy these types of machines.

How long Europe will have this advantage is not clear. The US Court of International Trade has recently overturned a finding by the International Trade Commission (ITC) that imports of supercomputers made by NEC and Fujitsu threaten to injure Cray Research. But the procedure will be lengthly, in any case.

Ad Emmen