Sean Kelly Wrote: > dsimcha Wrote: > > > == Quote from Sean Kelly (s...@invisibleduck.org)'s article > > > dsimcha Wrote: > > > > > > > > This is great for super-scalable concurrency, the kind you need for > > > > things like > > > > servers, but what about the case where you need concurrency mostly to > > > > exploit data > > > > parallelism in a multicore environment? Are we considering things like > > > > parallel > > > > foreach, map, reduce, etc. to be orthogonal to what's being discussed > > > > here, or do > > > > they fit together somehow? > > > I think it probably depends on the relative efficiency of a message > > > passing > > approach to one using a thread pool for the small-N case (particularly for > > very > > large datasets). If message passing can come close to the thread pool in > > performance then it's clearly preferable. It may come down to whether pass > > by > > reference is allowed in some instances. It's always possible to use casts > > to > > bypass checking and pass by reference anyway, but it would be nice if this > > weren't > > necessary. > > > > What about simplicity? Message passing is definitely safer. Parallel > > foreach > > (the kind that allows implicit sharing of basically everything, including > > stack > > variables) is basically a cowboy approach that leaves all safety concerns > > to the > > programmer. OTOH, parallel foreach is a very easy construct to understand > > and use > > in situations where you have data parallelism and you're doing things that > > are > > obviously (from the programmer's perspective) safe, even though they can't > > be > > statically proven safe (from the compiler's perspective). > > I'd like to think it isn't necessary to expose the internals of the algorithm > to the user. Parallel foreach (or map, since they're the same thing), could > as easily divide up the dataset and send slices to worker threads via > messages as with a more visible threading model. The only issue I can think > of is that for very large datasets you really have to pass references of one > kind or another. Scatter/gather with copying just isn't feasible when you're > at the limits of virtual memory.
Language extensions for message passing, such as Kilim for Java send messages giving away the ownership of data, not copying it. That's a reason for the need of compiler/runtime support. Also, parallel map/foreach is more feasible as a library-only solution, whether the message passing requires some support from the runtime environment. Please, correct me if I'm wrong.
