[Beowulf] New HPCC results, and an MX question
lindahl at pathscale.com
Wed Jul 20 14:28:41 PDT 2005
> Unless you want a very selected group to use your software, you'll have
> to take care that multithreaded software also works fast for clusters.
Funny, in all my years of activity in the high performance computing
niche, it's fairly rare that people run multi-threaded programs
accessing MPI. Pure MPI is the most widely used programming model,
followed as a distant second by OpenMP programs where only the main
thread does MPI (so a thread-safe MPI is not required), followed by a
few, rare very codes which are multi-threaded with either OpenMP or
Posix threads, and all the threads call MPI. Admittedly, this last
choice is can be a good one if you have a multi-threaded program and
want to add MPI. But there is a speed hit with low latency networks
due to the extra locking needed. And that was the small point I was
attemping to make.
Perhaps you can point me at some huge existing market of threaded MPI
programs? I'm sure you can speak with authority about your chess
program, but I didn't realize you were an expert on the general MPI
marketplace. With the advent of dual-core cpus, I figured the most
likely future was that most people were going to continue to run their
pure MPI programs as pure MPI programs.
More information about the Beowulf