Saturday, October 10, 2009


Although I read this whole paper, I can’t find too much to say about it that wasn’t already covered in the blogs and discussion about Concurrencer. This is because I see a lot of similarity between what ReLooper does and how Concurrencer will convert an algorithm to use the Fork-Join framework automatically for you. What ReLooper does actually seems like an easier and more naïve change that what Concurrencer does with the fork-join framework, so I’m surprised that Relooper gets its own paper and Eclipse plug-in, while Concurrencer does the fork-join optimization as only one of three of its features.

I would be happy with a compiler making these optimizations for me automatically if there were an iron-clad guarantee that no potential problems could be introduced. However, I think that is a hard guarantee to make and would prefer to either have a tool like Relooper imperatively define what I’d like to happen, but I would prefer an option to tell the compiler declaratively, via an attribute in .NET for example (an annotation in Java?), that the marked code is safe to be automatically parallelized.

No comments:

Post a Comment