Had fun presenting at HotPar. At first, I was very skeptical of the opportunistic computing talk, but then it ended up probably having the biggest impact on me. First, it reminded me of the whole issue of deciding how much compute to do for a problem -- instead of thinking about dropping frames in an scarce resource setting as I did in some Flapjax animation experiments, be more positive, and think about expressing the extra work you can do (whenever you have the time / energy to spare). Second, the proposal to run multiple instances of randomly terminating algorithms with good PDFs to achieve better speed can be adapted and tuned for energy efficiency (though I haven't encountered an opportunity for this in the browser yet). As Koushik remarked, it also finally motivates some fun statistical program analyses we talked about last year.
Overall, I left further convinced that I don't want to think about exposing web programmers to STM until the community performs a lot more semantics work and field studies (though there are a couple of places where I'm thinking of transactional implementations, but that's different from what you suggest end-user code should use). Finally, I want to take a stab at writing barrier-style code instead of task-parallel as I'm hitting a 24ms wall with TBB and Cilk++, which translates to a quarter second on a handheld.
(and Tim Harris is awesome.)
No comments:
Post a Comment