[Simgrid-user] Scalability of PlatformGenerator improved...
arnaud.legrand at imag.fr
Thu Jul 24 16:15:14 CEST 2008
Kurt Vanmechelen wrote:
> Well the main factor in the slow generation of the platform file was the
> string handling so I wouldn't necessarily exclude this for performance
That's what I also had understood from your previous explanations. And
that's why I was suggesting making it an option.
> In fact, we (and I assume others as well) would be interested in a platform
> description of the network topology that is graph-based (as opposed to giving
> info on all static routes). For frameworks that do not use static routing
> (e.g. GridSim), this info is indeed necessary and also a lot more space-
> friendly. I therefore wonder if you have versions of the platform files in the
> description archive with the router info included?
I'll let Martin or Frederic answer this question as I'm not completely
sure of the answer.
> In any case, such a description would also help scaling SimGrid up because, if
> I understand correctly, you now have the cluster tags in the new DTD but for
> instantiating the simulation you still need a routing table that scales
> quadratically in the number of hosts (CPUs) in the grid right?
Actually, the scaling issue can be circumvented in (at least) two ways.
The first one is simply to have a model that does not require any
routing information like the one I've been implementing a few months
ago. The second option is indeed to change the SURF internal
representation of the routing table by making use of the cluster
information. Doing so, we should gain an order of magnitude with respect
to the memory footprint while retaining the same accuracy. We're
planning on doing that in a few months.
By the way, our proposal on the SimGrid scalability has been accepted a
few days ago and we should thus get some fundings to hire more manpower
and transfer new interesting research ideas to SimGrid. We'll keep you
Optimization hinders evolution.
-- "Epigrams in Programming", by Alan J. Perlis of Yale University.
More information about the Simgrid-user