Bug 912076 (external-scheduler)

Red Hat Enterprise Virtualization Manager includes a new scheduler to handle the placement of virtual machines, allowing users to create new scheduling policies and write their own logic in Python and include that logic in a policy. The new oVirt scheduler serves virtual machine scheduling requests during migration or while virtual machine are running. Scheduling is performed by applying hard constraints and soft constraints to get the optimal host for that request at a given point of time. Scheduling Policy Elements * Filter: a basic logic unit that filters out hypervisors that do not satisfy the hard constraints for placing a given virtual machine. * Weight function: a function that calculates a score for a given host based on the internal logic of the function. Weight functions are a method for implementing soft constraints in the scheduling process. Since these are weights, a lower score is considered to be better. * Load balancing module: code implementing logic for distributing loads. To date, the definition of a load was mostly CPU related, which could be resolved by migrating a virtual machine. The new scheduler allows users to write their own logic to handle other load types such as network and I/O by other means such as integrating with third-party systems. Scheduling Process Outline Every cluster has a scheduling policy. To date, there were three main policies - None, Even distribution, and Power saving - and now administrators can create their own policies or use the built-in policies. Each policy contains a list of filters, one or more weight functions and a single load-balancing module. The scheduling process takes all relevant hosts and runs them through the relevant filters of a specific policy. Note that filter order is meaningless. The filtered host list is then used as an input to the relevant weight functions of that policy, which creates a cost table. The cost table indicates the host with the lowest weight (cost), which is the optimal solution for the given request. Multiple weight functions may be prioritized using a factor. Adding User Code See Bug #912059 for more information on adding code. Important Notes - New scheduling policies created by administrators are not validated by the system. This may end up with unexpected results, so it is highly important to verify that a new policy is not introducing issues or instability to the system. - User-provided code is unsupported. - Using user-provided code may affect performance, so administrators are advised to carefully test their code and the general performance changes.