[Beowulf] problem : mpi dynamic scheduling ??
hahn at physics.mcmaster.ca
Thu Jul 7 14:38:38 PDT 2005
> > in other words, large-tight apps are incompatible with meaningful
> > load-balancing.
> There are a couple of examples of tightly coupled apps that do either
> static or dynamic load balancing, but in most cases this load
> balancing was programmed in by the programmer above the MPI level. For
> example, ocean codes don't have to compute on land, so you can static
> load balance by using different-sized rectangles for your data
absolutely - I meant "load balance" in the context of the original poster,
who was talking about openmosix moving programs to less-loaded nodes.
that is, process load-balancing, rather than domain-decomp-type balancing.
> A system that does this invisibly is Charm++, used by
> namd, a molecular dynamics code. In namd's case the workload at a
> given point varies over time.
indeed - my experience is that any *serious* simulation-type code
needs to do this kind of data-load balancing (adaptive domain decomposition).
often, the picture gets really interesting when you make timesteps
adaptive as well as spatial dimensions ;)
really, my comment is mainly to emphasize that for intensive MPI apps
(which means significant in size and fairly tightly-coupled)
you have to avoid jitter at all costs. here's a BEAUTIFUL paper:
regards, mark hahn.
More information about the Beowulf