LHC Computer Grid - Rollout [mailto:[log in to unmask]]
> On Behalf Of Gonçalo Borges said:
> - One single LRMS queue per VO including 2 sets of (heterogeneous) resources
One point here, it should usually no longer be necessary to have one queue per VO, assuming that queue parameters like cpu time limits are the same. That was needed when we were still using the old LCG RB, but the WMS understands the VOViews which publish separate attributes for each VO within a single CE.
> - Install a single gLite-Cluster node and configure a single Cluster
> with two subclusters inside
> - Point the two CEs to the Cluster.
>
> Is this assumption correct?
Not quite. You shouldn't configure more than one Cluster per SubCluster - the glue schema allows it but the WMS doesn't know how to match against it. So the mapping needs to be one CE => one Cluster => one SubCluster (where by CE I mean a GlueCE, so equivalent to a batch queue). You can however have multiple Clusters published from one glite-CLUSTER node, and of course multiple CEs can submit to one cluster.
> If my assumption is true, it is not clear to me how in scenario B the
> two CEs can still execute jobs under the 2 sets of (heterogeneous)
> resources using the same LRMS queue. I guess this will not be possible
> and one will have to configure one LRMS queue per each subcluster. Is
> this right?
Yes; the only way jobs can be targeted at a particular subcluster is to have a batch queue that submits only to those WNs.
So, the overall process is that you start by deciding how many subclusters you need, and create a set of LRMS queues which submit only to WNs in each subcluster. Then in the info system you set up a glite-cluster node with one cluster per subcluster, and with all the subcluster attributes set appropriately for the WNs they represent. Usually it will be easiest to give the Cluster and SubCluster the same UniqueID (and note that the UniqueIDs need to be globally unique in the Grid, so generally they should be based on DNS names).
Finally, you need to connect each GlueCE to the Cluster that matches its batch queue configuration. For the LCG CE that's supported by yaim, for CREAM you currently need to do it by hand as described in the release notes.
Stephen
|