Smart Scheduler (SolverScheduler) - An efficient and richer scheduling driver
OpenStack scheduler has nice options like filters which allow resource selection based on simple constraints e.g. Don't put instances in a set on the same host. However for complex constraints, building a filter would be as complex as building a real constraint solver e.g. place VMs while minimizing average (VM-storage-
Initial performance benchmarks have shown this driver to perform better than the default Filter Scheduler.
This new driver will provide a better alternative to the default scheduler driver.
Collaborators: Xinyuan Huang, Yathi Udupi, Debo Dutta
Blueprint information
- Status:
- Started
- Approver:
- None
- Priority:
- Undefined
- Drafter:
- Yathiraj Udupi
- Direction:
- Needs approval
- Assignee:
- Xinyuan Huang
- Definition:
- Pending Approval
- Series goal:
- None
- Implementation:
-
Beta Available
- Milestone target:
- None
- Started by
- Yathiraj Udupi
- Completed by
Whiteboard
Added the nova-spec: https:/
Deferred to icehouse-3 as the blueprint was not approved by the icehouse-2 blueprint approval deadline. --russellb
Gerrit topic: https:/
Addressed by: https:/
Add a pluggable solver scheduler
This provides a solver scheduler, where complex constraints can be specified using pluggable solvers. An illustrative implementation of a solver is included and is based on PULP. Here the constraints are based on the disk space and memory requirements, and optimization is done using the host's free ram as a cost metric.
Addressed by: https:/
Add a pluggable solver scheduler
I'm not sure we want to add this right now since we're in the middle of making plans to split the scheduler out into a new repo. --russellb
russellb, why should the splitting out the scheduler effort block this? -- jogo
jogo, Can you please review this blueprint and get it approved for the next icehouse milestone target? The latest code patch now passes the JENKINS build and provides the base solver_scheduler driver and a reference implementation of a solver that is based on PULP. This patch is now ready for review. --Thanks, Yathi.
Addressed by: https:/
Add a pluggable LP solver to the solver scheduler
Addressed by: https:/
Add a pluggable LP solver to the solver scheduler
Addressed by: https:/
Adding ram, disk, vcpu resource linear constraints
Addressed by: https:/
Adding the Volume Affinity Cost function
Addressed by: https:/
Add Active-Host Constraint to the Solver Scheduler
Addressed by: https:/
Add Max-Instances-
Addressed by: https:/
Implements: blueprint solver-scheduler
Addressed by: https:/
Add One-Host-
Addressed by: https:/
Adding the Availability Zone constraint
Addressed by: https:/
Adding the IO Ops constraint
Apologies, this missed the deadline for Feature Freeze. Please rebase patches as soon as Juno opens, and we will try to get this in during that period. --johnthetubaguy (5th March 2014)
However seems like some of the approaches need to be discussed more, and we need a much clearer blueprint, with steps on what is being done, so we can try and stop blocking these changes after so much work has been done. Un-approving this one. Need more details on what configuration values will be added or altered, and how exisiting filters and constraints will be consolidated. --johnthetubaguy
-------
[Yathi - March 5 2014]
We would like to make a request for FFE for the Solver Scheduler work. A lot of work has gone into it since Sep’13, and the first patch has gone through several iteration after some reviews. The first patch - https:/
First Patch: https:/
Second patch with enhanced support for pluggable constraints and costs: - https:/
Subsequent patches add the constraints and the costs.
John,
Regarding the concerns you raised above to make this BP "pending Approval", the plan is not many new configuration values will be added, just the ones to specify the solver to use, and the pluggable constraints, and costs to use, with the weights for the costs. (these are mainly part of the second patch - https:/
The plan is to gradually support the concepts for the existing filters as the constraints that are accepted by our Solver Scheduler. Depending on the constraints and the costs chosen, the final scheduling will be done by solving the problem as an optimization problem.
Please reconsider this blueprint, and allow a FFE.
---------
Gerrit topic: https:/
Addressed by: https:/
Add a pluggable solver scheduler
Kicking out of juno-1, this needs a spec approved first. --johnthetubaguy 10th June 2014
