On Thu, Jul 03, 2008 at 01:45:44PM -0400, Robert G. Brown wrote: > This is one reason that people use PVM and MPI and so on. It can be > argued (and has been argued on this list IIRC) that raw networking code > will always result in a faster parallel program, all things being equal, > because encapsulating it in higher level abstractions always comes at a > cost
Yes, but there is plenty of proof that MPI can be extremely low overhead -- see InfiniPath MPI. Now if you insist on using TCP for your MPI and your "bare metal" code, you may beat your MPI because it's not so good at TCP. But it's probably cheaper to fix the MPI than re-invent the wheel many times. There's plenty of good code you can study for writing good network servers -- the original irc server is a pretty good event-driven non-blocking program. strace is your friend, too, that's the way you can see that MPICH-1's TCP driver isn't so hot. -- greg _______________________________________________ Beowulf mailing list, Beowulf@beowulf.org To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
