Haizea uses several scheduling algorithms internally to determine what resources to allocate to a lease. For the most part, modifying these algorithms requires digging deep into the Haizea code. However, several scheduling decisions that depend on an organizations own resource allocation policies are factored out of the main scheduling code into pluggable policy decision module. In particular, the following decisions are factored out:
9 Writing your own policies
- Lease admission
- : Should a lease request be accepted or rejected? Take into account that this decision takes place before Haizea determines if the request is even feasible. For example, an organization may require that all AR leases must be requested at least one hour in advance, regardless of whether there would be enough resources to satisfy the request before that time. However, being accepted doesn't guarantee the lease will get resources (although this could factor into the decision too); an AR lease could meet the ``one hour advance warning'' requirement, but still end up being rejected because there are no resources available at the requested time.
- Lease preemptability
- : How preemptable is a lease? Not all leases are created equal and, if the scheduler determines that a lease request can only be satisfied by preempting other leases, it may have to determine what leases are better candidates for preemption. For example, given a choice of preempting a lease that's been running for a week and another that's been running for five minutes, an organization might prefer to not interrupt the long-running lease.
- Host selection
- : What hosts should a lease be scheduled in? When the scheduler has a choice of several physical hosts on which to deploy VMs, some might be preferable than others. For example, an organization might want to pack as many VMs into the same hosts, to shut down those that are not running VMs, while another might want to spread those VMs across several hosts, leaving some free resources available in each host in case the VMs need extra capacity further down the road.
As you can see, these are all policy decisions that are driven by an organization's own goals for its resources. Thus, Haizea makes it simple to write your own policy decision code without having to modify Haizea's code. All you have to do is write a simple Python module, and then ``plug it'' into Haizea by instructing it (through the configuration file) to use that module. This chapter describes how this is done.
This documentation refers to Haizea objects, such as Lease and SlotTable that are not yet documented in this manual. For now, you will need to read the Haizea Pydoc documentation (linked from the Documentation section of the Haizea website) to see what attributes and methods these classes have. A more complete documentation will be included in the final 1.0 release.