There is something in the supercomputer airHeidelberg 23 jun 2001 Where is supercomputing heading? Well, to know that, you would go to the Mannheim Supercomputer conference in June and be completely up to date. This year the conference moved to Heidelberg, and the now completely international conference attracted more participants than ever before: over 380. So a lot of supercomputer experts gathered there. But not a clear picture, not a direction. There must be something in the air, but where is it what is it? Several presentations pointed at the limits that could prevent further developments in the supercomputer field. Physical limits for instance, that prevent miniaturisation to go on. With shrinking chip sizes, we soon reach the moment, when we would need half an electron to move which is not possible. Networks are increasing in bandwidth, but as for instance key-note speaker Kleinrock pointed out, the limit is the speed of light. This gives a limit for the latency: the time it takes for a signal to reach its destination. So although the bandwidth available for the high-end is now getting into the Gbit/s, latency is preventing further development.
What is the main job of a supercomputer director these days? Supervising building construction. Otherwise he cannot house the new huge, very huge supercomputers of today, as Horst Simon, director of NERSC pointed out. Supercomputers tend to get larger in size, despite the efforts of vendors to put more computing power on the square meter.
Software is reaching its limits. Despite lots of efforts in the past to come up with good parallel programming paradigms and tools, no one has really succeeded, as Brown pointed out. Yes there is MPI and yes there is MPP, but they are widely used standards, because the where en vogue at the time when there was an urgent need for standards, not because they are right for the job. MPI for instance was designed in the time of the hypercubes and the Connection Machines.
There are limits to what a single supercomputer company can do these days. No longer seems it possible for one company to design the processor, to design the architecture, to prescribe the operating system and the tools. Companies have to offer high-end supercomputers, where they can still use there own processors and OS, and they have to provide Intel clusters with NT, or Linux. Hence they are becoming more and more integrators.
Strange enough, it seems that there has also be reached a limit to how far the vector market can shrink. Despite it has been predicted many times these dinosaurs would appear from the scene, they now seem to have find a still considerably sized niceh market. NEC even has so much confidence in it, its road map now not only shows a SX-6 series of supercomputers, but also and SX-7 and even an SX-8 generation.
So where are we heading with supercomputers? Is it quantum computing as discussed in one of the Heidelberg lectures? No, at least not for decades to come. Does IBM know? No, not yet, that is why they started up their ambitious BlueGene project. Grid computing or Internet computing perhaps? No, they will make access more easy and have lots of other benefits, but are not a real solution for high-end supercomputing.
Something will happen to break the limits, there is something in the supercomputer air, but we just do not know what it is...
Ad Emmen |
 |