Are users disappointed? Yes they are.
Should they be disappointed? Probably not. The reason for the disappointment is that they had too high expectations caused by overhype from manufacturers - "Our system will run your existing codes" - the user wants this to mean that existing codes will work at peak performance and so assumes that is what was claimed. Overhype from the purchasers - "We have bought a 40Gigaflop machine and it is amongst the ten most powerful in the world." This mentality encourages the manufactures to design 40 GFlop Peak machines which produced 4 GFlop rather than 20 GFlop machines delivering 8 GFlop.
Is there a problem with attained efficiency levels? Not really. Properly written codes achieve the level of efficiency that should have been expected by informed people,
Does Efficiency Matter ? Not really - it is only one of many measures. The large memories, disc spaces etc. of the systems allow computations which are impossible on smaller apparently more efficient architecture. In private conversations with Chris Jones he told me that they were using the T3D for a class of problems which had just not been possible on the previous vector machines.
What can be done? The designers of systems should be encouraged in there endeavours to build systems with better balance between computation and communications. One way the user community can help is to drop asking about Peak Theoretical MFLOPS. If this is the glamour figure the marketing pressures will encourage the production of system with artificially high peak performance. The LinPack performance type benchmark has a similar effect. The HPC type machines are the sports cars of their world- they need a different set of skills to drive them fast One should not expect to use them without special training (i.e. reprogramming). Also they are only useful for a class of users with appropriate problems. Unfortunately one gets the impression that a job number on the T3D is seen as a sign of seniority in the EPSRC and so far too many people are trying to use it without adequate justification.
What should be done about it? Education is the key. The Community (Funding bodies, users etc.) needs to understand much more clearly that the Highest performance computer, like the largest telescopes and the large accelerators, a very specialised piece of equipment whose proper use is only for a group of trained specialists and is not a simple extension of a desk top PC.
Dennis Parkinson (dennisp@dcs.qmw.ac.uk)