Victor Eijkhout <[email protected]> writes: > Ok, so I read the paper. I sounds very much like algorithms I've seen > before, except that before MPI 3 you had to spell out the non-blocking > barrier explicitly. But that's beside the point. What was your point > with this paper? I'll agree that there are applications where the > connectivity changes quite regularly, and so you need a Log(P) setup, > rather than the normal O(P). But it stays a Bulk Synchronous scheme > that is in no way hard to fit in the current IMP design; I just > haven't done so yet. So what was your reason for bringing it up?
I was asking whether you could *implement* this algorithm using IMP. It sounds like you are suggesting extending IMP to include it as a primitive. That's okay, but increasing the number of primitives makes the model more complicated and I think it is generally a sign of weakness when a base abstraction needs many primitives. It is typically not easy for users to extend primitives. If you just need a few and then the abstraction is rock solid for every purpose that comes along, great. But if the model frequently fails to express the desired method, it either needs work or serious thought to make it play well with others.
pgpaLEt2mDcs0.pgp
Description: PGP signature
