[Haizea] Coscheduling multiple jobs

[Borja Sotomayor]

Hi,

> How we can co-schedule a set of jobs on a physical host?

Are you referring to jobs or leases? If the former, I would argue this
is not Haizea's responsibility. Once a lease is granted (e.g., to set
up a cluster), Haizea doesn't really care what goes on inside the
lease: it would be up to the cluster scheduler running inside the
lease to co-schedule jobs on the same hosts.

If the latter, the way you describe could work, although it seems like
it would in essence be a host selection policy that packs as many
leases in each host before moving on to the next one. This wouldn't
necessarily guarantee that several leases get scheduled in the same
physical host (e.g., Haizea could start checking a host that is almost
full, and only schedule one lease on it, and the rest on a separate
host, when it might have been possible to schedule all the leases on
the same host if it had first searched for a host that was big
enough).

A better way could be to add a new "affinity" attribute to leases,
where you indicate if this lease has an affinity for other leases (or,
alternatively, just a way to designate that several leases are part of
the same group or lease). When scheduling a lease, the scheduler would
have to check if it has an affinity for other leases and, if it does,
it would attempt to schedule it on the same physical host (which might
not be possible all the times). To avoid the scenario described
earlier, when the first such lease comes along, the scheduler could
schedule it on an empty host so that, when the other leases come
along, there is a good chance they'll be scheduled on the same host.

Cheers!
-- 
Borja Sotomayor
-Scientific Writer at UChicago's Computation Institute
http://people.cs.uchicago.edu/~borja/
-Community Manager for the OpenNebula project
http://www.opennebula.org/