Hyper-V Cluster

Hyper-V Cluster

https://blueprints.launchpad.net/nova/+spec/hyper-v-cluster

Hyper-V Clustering has been introduced since Windows / Hyper-V Server 2008 and it introduced several benefits such as highly available VMs, better performance, faster live migrations and other features. [1][2][3]

Problem description

Hyper-V Clustering can bring a set of advantages to advantages that are not available otherwise and also improve the performance of existing features. A few examples would be highly available VMs, faster live migrations, network health detection, etc. A more detailed list of features can be found in the References section [1][2][3].

Currently, there is no support for Hyper-V Clusters in OpenStack. This blueprint is addressing this issue and adds an implementation.

Use Cases

This feature is particularly useful for its increased performance, highly available VMs and virtual machine and virtual machine network health detection.

Proposed change

There are two methods for creating and deploying a Hyper-V Cluster, each with their own advantages and disadvantages:

  • Option A. Hyper-V Cluster controlled by a single nova-compute service. This means that the nova-compute service will run on a single Hyper-V Node in a Cluster and can manipulate WMI objects remotely on all the Cluster Nodes.

    Advantages:

    • Consistent disk resource tracking. The Cluster Shared Storage is only tracked by a single compute service.

    • Smaller overhead, as only one nova-compute service will necessary, as oposed to one nova-compute service / node.

    Disadvantages:

    • neutron-hyperv-agent are still mandatory on every Node. Even though its performance has been enhanced over the past release cycles, it won’t be able to handle port binding efficiently, VLAN tagging and creating security group rules for each new port (up to thousands of ports in some scenarios).

    • ceilometer-agent-compute will have to run on each Node or implementing a Hyper-V Cluster Inspector is necessary, in order to poll the metrics of all the resources.

    • Free memory tracking issue. Consider this example: 16 x Nodes Cluster, each having 1 GB free memory => ResourceTracker will report 16 GB free memory. Deploying a 2 GB instance in the Cluster fails, as there is no viable host for it.

    • Free vCPU tracking issue. Same as above.

    • nova-compute service might perform poorly, as it will spawn threads for console logging for a considerably larger number of instances, which will cause the serial console access to be less responsive.

    • When performing actions on an instance, extra queries will be necessary in the Hyper-V Cluster Driver to determine on which Node the instance resides, in order to properly manipulate it.

    • The Hyper-V Cluster will act as a scheduler in choosing a node for a new instance, resulting in poor allocation choices.

    • The underlying cluster infrastructure will be opaque and the user won’t be be able to know on which physical node the instance resides usinf Nova API.

    • Users cannot choose to live-migrate in the same cluster. As there is only one compute node reported in nova, all the ‘foo’ instances will be deployed on the host ‘bar’ and running the command:

      nova live-migration foo bar

      will result in a UnableToMigrateToSelf exception. This will negate one of the Hyper-V Cluster’s advantages: faster live migrations within the same Cluster.

  • Option B. nova-compute service on each Hyper-V Cluster Node.

    Advantages:

    • Correct memory and vCPU tracking.

    • nova-scheduler will properly schedule the instances in the Cluster.

    • No decrease in nova-compute service’s performance.

    • Live migrations within the same cluster are faster.

    Disadvantages:

    • Free disk resource tracking. Since all the nova-compute services will report on the same Cluster Shared Storage, each ResourceTracker will report different amount of storage used. For example, having a 500 GB shared storage and 2 instances with 200 GB used storage each on a single node in the cluster, that node will report having 100 GB free storage space, while other nodes, with no instances, will report as having 500 GB free. Trying to deploy another 200 GB instance would fail. (WIP)

This blueprint will address Option B, as its value far outweighs Option A.

Almost all the existing Hyper-V code in nova is reusable for the purpose of creating the Hyper-V Cluster Driver, though a few changes are necessary for Option B:

  • Instances will have to added to be clustered when they are spawned.

  • Need to check before live migration if the new host is in the same Cluster. If it is in the same Cluster, cluster live migration will have to be performed, otherwise, the instance will have to unclustered before doing a classic live migration.

  • Cold migrations are still possible in Hyper-V Clusters, the same conditions as live migration apply.

  • The instance must be unclustered before it is destroyed.

  • When new instance is added to the Cluster via live migration or cold migration from a non-clustered Hyper-V Server or from another Cluster, the instance will have to be clustered.

  • Develop method to query free / available disk space for a Cluster Shared Storage, which will be reported to the Resource Tracker.

  • Develop method to ensure that only one Hyper-V compute node will fetch a certain glance image.

Alternatives

None, in order take advantage of the benefits offered by the Hyper-V Cluster, the instances have to be clustered.

Data model impact

None

REST API impact

None

Security impact

nova-compute service will have to run with an Active Directory user which has Hyper-V Management priviledges on all the Hyper-V nodes.

Notifications impact

None

Other end user impact

None

Performance Impact

  • Because of the cluster shared storage, the images will have to cached only once per cluster, instead of once per node, resulting in less storage used for caching and less time spent doing it.

  • Because of the cluster shared storage, live migration and cold migration duration is greatly reduced.

  • Host evacuation takes place automatically when a clustered compute node is put into maintenance mode or is taken down. The instances are live-migrated, assuring high availability.

Other deployer impact

  • Hyper-V Cluster requirements: [4]

  • Creating Hyper-V Cluster: [5]

  • Hyper-V nodes will have to be joined in an Active Directory.

  • Hyper-V nodes will have to be joined in a Failover Cluster and the setup has to be validated.[6][7]

  • Only nodes with the same version can be joined in the same cluster. For example, clusters can contain only Windows / Hyper-V Server 2012, Windows / Hyper-V Server 2012 R2 or Windows / Hyper-V Server 2008 R2.

  • All Hyper-V nodes in the cluster must have access to the same shared cluster storage.

  • The path to the shared storage will have to be set in the compute nodes’ nova.conf file as such: instances_path=\SHARED_STORAGEOpenStackInstances

  • The compute_driver in compute nodes’ nova.conf file will have to be set as such: compute_driver=nova.virt.hyperv.cluster.driver.HyperVClusterDriver

  • The WMI namespace for the Hyper-V Cluster is ‘/root/MSCluster’. When using that namespace, the driver will fail to start due to stack overflow exception while instantiating the namespace. This is happens because of a missing magic method in the WMI module (__nonzero__). This happens in python wmi module, for versions 1.4.9 or older.

  • Hyper-V nodes in the same Cluster should be added to the same host aggregate. This will ensure that the scheduler will opt for a host in the same aggregate for cold migration.

Developer impact

None

Implementation

Assignee(s)

Primary assignee:

Claudiu Belu <cbelu@cloudbasesolutions.com>

Work Items

As described in the Proposed change section.

Dependencies

None

Testing

  • Unit tests.

  • Tempest tests will be able to validate this feature and they will run as part of the Hyper-V CI.

Documentation Impact

Documentation about HyperVClusterDriver will be added.

History

Creative Commons Attribution 3.0 License

Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.

nova-specs