"Limits to implicit parallelism in functional application"
How much implicit parallelism is there in functional programs?
One answer is "lots (I'd imagine)." Functional languages (and especially non-strict functional languages) can place relatively few constraints on a program's order of evaluation, often allowing multiple evaluations to proceed in parallel--at least in principle. The resulting "implicit" parallelism may be easy and useful to locate and exploit.
Another answer is "not so much (in actual experience)." When you take an existing, non-trivial functional program and try to find practical ways to speed it up by judicious placement of `par` (or by strategies or whatever), a surprisingly large effort can produce a surprisingly small improvement. So maybe that implicit parallelism just isn't there.
Another is "well, maybe it's there but it's just hard to find." If people can't find much implicit parallelism in real software, maybe it's the fault of the people. Maybe `par` and strategies and such are just too hard for people to use. Maybe some automatic parallelization mechanisms, yet to be invented, could do much better. Or something.
One interesting new answer is "I don't know. Let's measure it." I've just written a short paper on measuring the implicit parallelism in real Haskell programs, such as the NHC98 compiler. These measurements are in the context of a limit study, so they let us say that there is _no more_ than some certain amount of implicit parallelism present; exploiting even that much might of course be impractical.
The paper is at http://www.detreville.org/papers/Limits.pdf. I'd appreciate comments, of course.
Active forum topics
New forum topics