Peter Dzwig, Concertant
As of Monday morning there were some 10,023 registrations for SC08, a record. What do they come for? There are a variety of reasons. There are tutorials for learning new things, posters and workshops for learning about research work in progress, paper presentations for learning about the more mature research and work, and, of course, it is a trade show.
The exhibition hall is packed with exhibitors from Intel and IBM on down; including a vast panoply of research institutions, hardware manufacturers, networking companies and software providers. Research institutions seem to be well-represented as they fight for scarce funds and probably even scarcer commercial clients. Given that we are in Austin, TX, you would expect that most would be US-based but while that is true, there are also significant numbers of Europeans, Chinese, Japanese and others. Machines for this sort of activity don't come cheap. Indeed as someone quipped earlier, we should probably redefine HPC as HCC for High Cost Computing.
Maybe that is true, but HPC is being deployed across an ever wider range of areas of research and development from the financial markets to weather forecasting. Someone has even parked a huge US truck on their stand in what is presumably to be interpreted as an example of applicability. Supercomputing does have real economic impact, particularly for the bigger corporations. Of those one of the major users is the automobile industry, for whom it has become an indispensable tool. At a time when the US automobile industry is in a parlous state, one has to ask how much longer they can continue to use it.
Away from the exhibition there is a vast array of workshops, seminars, discussion groups and lectures that range across all the activities above and then some. These attract a fair proportion of the participants. So far there doesn't seem to be a huge amount of interest in multicore processors per se, despite one workshop that ran yesterday. Most activity is focused on supercomputers with high parallelism true, but people are generally not concerned about the details of how the parallelism is realized.
When Tilera's CTO talked about their 64-core engine, which we have discussed in a previous article, a few people got up and left; it would seem that it had a few too many cores. Earlier talks had covered processors with up to eight cores. None had dealt with the issues dealing with high counts. Tilera, it would appear, are getting traction for the Tile64 and it is being widely deployed, with quite a lot of activity in Europe as well as in the US. They are using SMP Linux with C/C++ as their platform of choice but others are porting or have already ported, OpenMP, Erlang and even Smalltalk to it.
David Bader from Georgia Tech., gave a talk on “Accelerating Applications with Cell Broadband Engine, Graphics, and Multi-threaded Processors” which covered issues surounding parallelism for a wide range of processors from Cells to GPGPUs and SIMD arrays such as Clearspeed's machines. His opinion is that there is little hope of a common approach to parallelism, given that the issues that come with the reduction of a problem to a form associated with one system makes it almost inherently un-portable to another system. He is therefore looking to portable libraries and to standards as the way forward.
The show promises much. We expect a lively programme over the next few days. As a starter, the OpenCL session on Monday evening was very spirited. For more of that see our article.