Watcher Overload standard deviation algorithm

https://blueprints.launchpad.net/watcher/+spec/watcher-overload-sd

Problem description

Cluster overload is a complex problem, that can be solved by migrating VMs from one node to another. The decision to migrate is based on VM parameters such as CPU utilization or memory utilization. We propose a strategy for Watcher, that monitors if there is a higher load on some hosts compared to other hosts in the cluster and re-balances the work across hosts to minimize the standard deviation of the loads in a cluster.

The main purpose of the strategy is to choose the pair VM:dest_host that minimizes the standard deviation in a cluster best.

This specification relates to blueprint: https://blueprints.launchpad.net/watcher/+spec/watcher-overload-sd

The original code implementing this algorithm (not in the context of Watcher) was published here: https://github.com/joker946/nova/commits/drs

Use Cases

As the Administrator I want to reduce load on cluster by triggering a Watcher optimization with the Workload Stabilization goal to create action plan with the list of recommended actions. This strategy can be configured by using watcher input parameters (e.g. thresholds, weights, metrics). After then I approve proposed action plan, which contains set of VM live-migration actions.

Project Priority

Not relevant because Watcher is not in the big tent so far.

Proposed change

The new Workload Stabilization Strategy allows to determine if there is an outlier in the load for the hosts in a cluster and respond to it by migrating VMs to average the load across the cluster. Input: It is possible to get information about workload of every VM using Ceilometer. In order to do that the means of CPU and RAM utilization for the last 10 minutes are used (configurable). Then the hypervisor workload of each node can be calculated by summing up workload of all VMs on each node. Knowing total amount of CPU and RAM on every node allows to normalise CPU and RAM load. After that the means of CPU and RAM loads for the cluster are calculated. All this information makes it possible to calculate the standard deviation of CPU or RAM. After standard deviation is calclulated for CPU and RAM, the strategy compares sd values to cpu_threshold and ram_threshold (these are provided by watcher input parameters). If one of the values is greater than the threshold value, the strategy runs cluster stabilization process to reduce load on cluster by performing live-migration of VMs. Pseudo code:

load_for_node = {'compute1': {'cpu', 'ram'}, 'compute2': ...}
for all vm in node:
    load_for_vm = {'uuid', 'cpu_util', 'memory.resident', 'vcpus'}
    load_for_vm[cpu_util] = transform_vm_cpu_util_to_host_cpu_util()
    load_for_node += load_for_vm
load_for_cluster = normalize load_for_node

for all node in cluster:
    compute deviation for each node against load_for_cluster.
    Then choose the src vm and dest host.

The implementation of transform_vm_cpu_util_to_host_cpu_util function is described in a Cluster stabilization process paragraph.

Cluster stabilization process: For each active VM we simulate migration to another node of the cluster by recalculating workload of source node and destination node. The following code shows a general case of simulating:

new_hosts[src_hp_id][metric] -= vm_load[metric]
new_hosts[dst_hp_id][metric] += vm_load[metric]

Meanwhile, the shown code isn’t acceptable for cpu_util metric because of probable difference in the number of node vcpus. Therefore we transform VM CPU utilization to overall host CPU utilization. The following pseudocode shows the right way of transforming:

\(cpu_{impact} = cpu_{util} * {vm_vcpus}/{host_vcpus}\)

After that the load values of nodes are normalised and the new standard deviation is calculated with them. The total standard deviation is calculated as weighted arithmetic mean:

\(\sum weight_{metric}*sd_{metric}\), where:

  • weight_metric is the weight of metric. The value of weight is to be in range(0, 1). Each weight of metric is set by watcher input parameters.

  • sd_metric is a calculated standard deviation of metric in cluster.

From all options of moving VM to other nodes we choose the one that minimizes the standard deviation in a cluster best. That option is added to the map of migration options, that afterwards gets sorted by standard deviation in ascending order. From the resulting list we iteratively take a host/vm pair and place it into solution with action “MIGRATION”. We compare the resulting deviations of CPU and RAM to the threshold values. If the resulting deviations are less than the threshold values, the cycle stops and the resulting action plan is sent to Watcher Applier.

Alternatives

The alternatives to this approach are to use different Goals and associated Strategies defined in Watcher.

Data model impact

None expected.

REST API impact

There is no impact on the REST API.

Security impact

None expected.

Notifications impact

No specific notifications associated with executing a specific Strategy are envisaged. (Notifications could arise from the resulting actions, but these are presumably handled in other parts of Watcher).

Other end user impact

This capability will not have any specific impact on either the API or python-watcherclient.

Performance Impact

No specific performance impact is expected.

Other deployer impact

No specific deployer impact is envisaged.

Developer impact

This will not impact other developers working on OpenStack.

Implementation

Assignee(s)

Primary assignee:

Alexander Chadin <alexchadin>

Other contributors:

Alexander Stavitskiy <alexstav>

Work Items

This task can be considered atomic. It just requires the development and test of a single class. The main WorkloadStabilization class of this strategy is inherited from BaseStrategy class.

Dependencies

There is a dependency with BP optimization threshold.

Testing

Several unit tests will be provided to test various scenarios.

Documentation Impact

It will be necessary to add new content relating to this new Goal and Strategy to the documentation.

References

No references.

History

No history.