Thursday, July 09, 2009

NSF EDA Workshop: Wrap up

Today morning was wrap up day. The theory group was "tasked" with coming up with a few slides suggested avenues for further cooperation. Dick talked about creating (or re-creating) an environment in academia where researchers are building chips and talking to other experts down the corridor (like theoreticians) rather than the 'planted' model where a theoretician is planted in a corporate design environment for 6 months or so.

I talked briefly about what I called "new" theory: the twin ideas of randomization and approximation that have been so powerful in algorithm design (even for intractable prolems), and how these play into the problems of dealing with massive high dimensional data.

Vijaya gave a synopsis of cache-obliviousness, multicore research, and her recent work in this area. There's a lot of interest in EDA and multicore work, and she made the argument that theoreticians are learning how to design efficient multicore algorithms and can be of great help in adapting EDA methods.

There were also four sub-panels that addressed various aspects of EDA. What to me was interesting that many of the panels brought up "EDA needs in theory" that matched some of the things we talked about. For example, there's a pressing need for methods that run linear (and even sublinear) time, tools that are incremental, in that they can progressively generate better and better solutions given more time, and tools that can parallelize well.

They also talked about providing benchmarks: for example, a 10,000 variable SAT instance and code that solves it. I thought (and said) that this would be an excellent way to get people dabbling in this area.

Randomization was a trickier matter. Although the EDA folks recognize the power of randomization, they are concerned about reproducibility. The DA pipelne is long and involved, and once you've fixed the output of a piece of the pipeline, you'd like to keep it in place and not change from iteration to iteration.

Of course this is something that can be fixed with seed control. For more portability, it's conceivable that you can cache the random coin tosses once you've settled on a reasonable solution to any piece in the pipeline.

Overall, it was quite interesting, although exhausting. The general idea with such workshops is that the findings make their way into the next call for proposals (sometime in October/November), so if you have EDA people you'd like to collaborate it, this might be a good opportunity.

It's time to go home.

Disqus for The Geomblog