Messages in this thread | | | From | "Stephen C. Tweedie" <> | Date | Mon, 6 Mar 2000 18:05:10 +0000 (GMT) | Subject | Re: Help in DSM design |
| |
Hi,
On Sun, 5 Mar 2000 13:07:44 +0100, Jamie Lokier <lk@tantalophile.demon.co.uk> said:
> Not true for every iteration. You have N*N*N subcubes, one subcube per > node, and the relations are such that it takes N/T iterations for input > on one side of the cube to affect output on the other.
> A cube transfers its current state to a surface depth of D (where D < > N) to all its neighbours. It's now possible for every neighbour to > compute for D/T iterations before requiring the next state update.
Yep, at the cost of increasing the total amount of computation required: you now have to compute the outer D layers of each of your neighbours, too. Total work on each node goes up from N**3 to (N+2D)**3. Ouch, that looks nasty.
This might be a workaround, but it's not a generic solution. For special cases where the domain density is extremely high and so D<<N, it can improve performance somewhat.
--Stephen
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |