Friday, December 19, 2008

Preliminary second result

The web could get a bit faster. This is modeling a speedup on a different 10% of browser CPU time.

Likes: task parallelism via work-stealing from stacks, lexical scoping (eg., Cilk).

Dislikes: heavyweight task creation, crazy-wtf scoping (eg., OpenMP 3.0), heavyweight build processes with bad documentation (GCC), parallelism packages requiring manual lambda-lifting (TBB).

Hurray for Heidi methodically disentangling OpenMP scoping, task creation, and scheduling rules tonight :)

Anyways, our prototypes suggest we can decrease 20% of the time to maybe 3%.. hopefully we can find someone else to look at the components in the other 80% (which I'm *way* more optimistic about -- they didn't seem as much of a research challenge) and maybe even doing this with actual browser code.

No comments: