At the beginning of SC08 we asked six critical questions about how prepared the industry is for multicore processors. To recap they were:

• Can the threads model really be made to work as core counts get into the tens and beyond?
• Which programming and memory models are going to work?
• How will processor architectures evolve?
• How will the processor arrays be managed?
• What languages are required to program these systems and where are they going to come from?
• How will software innovations track hardware developments?

So what state do we feel the industry is in in this regard as SC08 comes to an end?

As everyone heads off, the trucks are loaded up with millions of dollars worth of equipment, the stands - including The Ghastly Green Golfers - are collapsed for another few weeks, it is clear that the industry, while apprehensive about an impending recession, is in fairly good shape. There is undoubtedly going to be a shake-up, that much is inevitable in the present climate. There are those who won't make the next show and there will be mergers. It would not be right to identify any individuals, but there have been rumours around the halls, though most are without much foundation. We see a very complex industry with vast reservoirs of skills. The issue for all the companies is how deploy those skills successfully and to the advantage of everyone in an economic slowdown. With the downturn hitting many of the sectors on which the industry traditionally relies most strongly, some kind of contraction or re-alignment seems inevitable. And yet... there are still start-ups such as Wallach's Convey which have made a huge impact on the show. Whether that will be converted into sales remains to be seen, of course.

To return to our six points: we need to bear in mind that what we have seen is within the context of a show. While it showcases the best in traditional HPC, it has not shown much in the way of novel (processor) architectures. New devices, but not new architectures. To most at SC08, high core count means more chips with relatively low core counts per chip. There have been relatively few talks on high core count chips. There have been notable exceptions such as Patterson's and Mattson's talks, but these have been far from being in a majority. So...

• Will the threads model survive? Those who are threads-oriented at present seem to think that it will survive; but in the long term the wider software community doesn't offer consensus as to the model's future; there are simply too many options on offer. There are also a number of problems with threading at high core counts. The majority of the hardware community doubts the longer-term viability of the model.
• Which programming models and which memory models are going to work? The jury is out, see our comments above. But the industry has a dichotomy to face. The traditional methods may serve it well for the next few years, but they won't meet the needs of the architectures predicted over the next few years. Hardware models and memory models look set change and change to a more distributed model.
• How will processor architectures evolve? This seems to have been the most surprising outcome of the show. We anticipated that people would be happy to talk timescales for the evolution of hardware. Frankly that hasn't happened. There is overall agreement that chips with higher core counts going to happen. What the growth rate will be? No-one seems prepared to commit to a number beyond saying that it will roughly follow Moore's Law, which isn't saying much! The question of what architecture the cores are going to have does not seem to be being addressed clearly. People talk of x86 cores - but that seems to be more by default. Immediate plans focus on x86 architectures, with little evidence of that changing at least in the near future. Processors such as Intel's Terascale architecture are clearly designed as experimental platforms, and others must be looking at the issues; even if they don't have in-house silicon that they are prepared to talk about. Nobody is really sticking their heads above the parapet! Yet.
• How will multicore processor arrays be managed? The issues here are also poorly defined. While there has been a substantial amount of work on MPI discussed at the show and in the talks, this is mostly in the context of large arrays of low-core-count processors, although it has implications in the multicore world. It is not clear that MPI, which is a high overhead system, makes much sense for MCP systems where a much lighter-weight protocol is needed - OpenMP is often proposed for this situation but MPI is message passing and OpenMP is shared memory so there is a fundamental dichotomy here. SMPs appeared widely in talks at SC08 and also on the floor of the show. For the moment they appear to be the standard assumption for the next few years, faut de mieux. One or two speakers have tried to push their applicability out as far as they can, but that rapidly gets mired in communications and priority issues. However they are here and their evolution and market position will be interesting to watch over the next few years.
• What languages are required to program these systems and where are they going to come from? According to most people it is the usual suspects, namely C, C++, some Fortran, and some Java. We have seen evidence of dataflow models too on Pervasive's stand and one or two places elsewhere. We had talks on Haskell in the workshop about programming multicore hardware earlier in the week, but these have been exceptions rather than the rule.
• How will software innovations track hardware developments? There really has been very little evidence of agreement between hardware engineers and software engineers on how to address the move to multicore. The hardware guys know that they can meet a Moore's Law-like growth curve, but the software world hasn't begun to set out its timescales. If it can't deliver then either: a) the hardware industry will miss the curve and perform at very much sub-Moore's Law growth rate; or b) the hardware industry will take the initiative and enforce an approach. That has been done in the past, think IBM, AT&T/Bell Labs and a host of others. While many may argue that the world has changed, the imperatives of commercial growth, particularly in a time of recession, may be such that the hardware industry feels obliged to seize the bull by the horns in order to avoid damaging its growth prospects yet further.

At the end of the show, both hardware and software industries are clearly in a state of flux with regard to the multicore future and little seems to be being done to resolve the issues. There are both challenges and problems to be faced. Let's hope that there are more detailed plans in place by SC09 in Portland, Oregon.