[Beowulf] problem : mpi dynamic scheduling ??
Many of your questions may have already been answered in earlier discussions or in the FAQ. The search results page will indicate current discussions as well as past list serves, articles, and papers.
Mark Hahn hahn at physics.mcmaster.caTue Jul 5 10:04:03 PDT 2005
- Previous message: [Beowulf] problem : mpi dynamic scheduling ??
- Next message: [Beowulf] problem : mpi dynamic scheduling ??
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
> i m using mpiJava and i need to know ,Do MPI 0ffers Api's / Functions for > dynamic scheduling , Process migration , Load and balancing?? . not generally, and I think there are good reasons, not just laziness. to make any large, tightly-coupled application run efficiently, each proc needs to avoid being preempted. that means that you really can't do any load-balancing, since you have to set aside a whole cpu for a proc, and hope that the kernel and random daemons don't interfere too much. (indeed, there are systems which try to micro-kernel-ize the environment. it's also traditional to gang-schedule, so that if you really need to run a daemon, you do it on all nodes at once.) in other words, large-tight apps are incompatible with meaningful load-balancing. not all the world consists of large, tightly coupled MPI applications, of course. but I'm pretty convinced that for looser applications, especially in the presence of dynamic parallelism, and migration, I'd try to build a sort of peer-to-peer version of Linda, rather than use MPI. regards, mark hahn.
- Previous message: [Beowulf] problem : mpi dynamic scheduling ??
- Next message: [Beowulf] problem : mpi dynamic scheduling ??
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the Beowulf mailing list
