Survey of cluster manager scheduling capabilities from the point of view of resource utilization.

Ganeti

Evaluation of Ganeti found that a) there is only a commandline interface and b) it is focused on instance failure mechanisms (using LVM and DRBD).  There are no concepts of backfill, reservations, or running jobs.

SGE

SGE does not have seem to have direct VM scheduling support.  There are currently some prototype projects that use SGE inside VMs but we are looking for a scheduler to directly control the cluster's resources.  A system such as the workspace pilot could be used to reserve resources using a "fake" job.  In such a setup, Nimbus could conceivably submit with a coarse strategy: "low priority" VM jobs for backfilling and "high priority" jobs for IaaS requests into the SGE queue.

Condor

Condor excels at resource utilization because of its focus on preemption.  It looks possible to configure a Condor based VMM node to do one of two things: a) start "background" VMs that can be preempted by higher priority requests (the background VM would have a Condor overlay inside of it for general jobs) or b) accept jobs in dom0 directly which are themselves preempted by VM requests from the Nimbus service.  In both cases the Nimbus service is a client to Condor that, because of the configuration of each VMM node, is empowered with the ability to make high priority requests.