On Fri, 16 Apr 2004, Gordon, JC (John) wrote:
> > .. You can't , GLUE assumes all hardware is the same.
> > Select some middle ground perhaps with the exception of
> > memory where you should publish the smallest memory available.
>
> GLUE has subclusters which should be heterogeneous. A cluster can then
> consist of a number of clusters. That was the plan anyway. Whether CEs
> actually map onto clusters I cannot confirm.
Unfortunately it doesn't work that way. The RBs don't have a mechanism
for dealing with multiple subclusters so we have the restriction that
a CE can only map to a single subcluster, which is defined by GLUE to
be completely homogeneous (in principle down to things like exact
processor type and clock speed). Even if the RB could direct jobs
to a particular subcluster among many the schema is faulty in that
it doesn't tell you how many nodes are in a subcluster, and things
like the number of running jobs, free cpus, estimated traversal time
etc are all defined at the CE (queue) level so you couldn't do
scheduling at a subcluster level.
Stephen
|