Concurrency

Introduction

Modern processors typically contain multiple cores all capable of executing instructions in parallel. Ensuring applications can fully utilize modern underlying hardware requires developing with these concepts in mind. The OpenStack foundation maintains a number of libraries to facilitate this utilization, combined with constructs like CPython’s GIL the proper use of these concepts becomes more straightforward compared to other programming languages.

The primary libraries maintained by OpenStack to facilitate concurrency are futurist and taskflow. Here futurist is a more straightforward and lightweight library while taskflow is more advanced supporting features like rollback mechanisms. Within Watcher both libraries are used to facilitate concurrency.

Threadpool

A threadpool is a collection of one or more threads typically called workers to which tasks can be submitted. These submitted tasks will be scheduled by a threadpool and subsequently executed. In the case of Python tasks typically are bounded or unbounded methods while other programming languages like Java require implementing an interface.

The order and amount of concurrency with which these tasks are executed is up to the threadpool to decide. Some libraries like taskflow allow for either strong or loose ordering of tasks while others like futurist might only support loose ordering. Taskflow supports building tree-based hierarchies of dependent tasks for example.

Upon submission of a task to a threadpool a so called future is returned. These objects allow to determine information about the task such as if it is currently being executed or if it has finished execution. When the task has finished execution the future can also be used to retrieve what was returned by the method.

Some libraries like futurist provide synchronization primitives for collections of futures such as wait_for_any. The following sections will cover different types of concurrency used in various services of Watcher.

Decision engine concurrency

The concurrency in the decision engine is governed by two independent threadpools. Both of these threadpools are GreenThreadPoolExecutor from the futurist library. One of these is used automatically and most contributors will not interact with it while developing new features. The other threadpool can frequently be used while developing new features or updating existing ones. It is known as the DecisionEngineThreadpool and allows to achieve performance improvements in network or I/O bound operations.

AuditEndpoint

The first threadpool is used to allow multiple audits to be run in parallel. In practice, however, only one audit can be run in parallel. This is due to the data model used by audits being a singleton. To prevent audits destroying each others data model one must wait for the other to complete before being allowed to access this data model. A performance improvement could be achieved by being more intelligent in the use, caching and construction of these data models.

DecisionEngineThreadPool

The second threadpool is used for generic tasks, typically networking and I/O could benefit the most of this threadpool. Upon execution of an audit this threadpool can be utilized to retrieve information from the Nova compute service for instance. This second threadpool is a singleton and is shared amongst concurrently running audits as a result the amount of workers is static and independent from the amount of workers in the first threadpool. The use of the DecisionEngineThreadpool while building the Nova compute data model is demonstrated to show how it can effectively be used.

In the following example a reference to the DecisionEngineThreadpool is stored in self.executor. Here two tasks are submitted one with function self._collect_aggregates and the other function self._collect_zones. With both self.executor.submit calls subsequent arguments are passed to the function. All subsequent arguments are passed to the function being submitted as task following the common (fn, *args, **kwargs) signature. One of the original signatures would be def _collect_aggregates(host_aggregates, compute_nodes) for example.

zone_aggregate_futures = {
    self.executor.submit(
        self._collect_aggregates, host_aggregates, compute_nodes),
    self.executor.submit(
        self._collect_zones, availability_zones, compute_nodes)
}
waiters.wait_for_all(zone_aggregate_futures)

The last statement of the example above waits on all futures to complete. Similarly, waiters.wait_for_any will wait for any future of the specified collection to complete. To simplify the usage of wait_for_any the DecisiongEngineThreadpool defines a do_while_futures method. This method will iterate in a do_while loop over a collection of futures until all of them have completed. The advantage of do_while_futures is that it allows to immediately call a method as soon as a future finishes. The arguments for this callback method can be supplied when calling do_while_futures, however, the first argument to the callback is always the future itself! If the collection of futures can safely be modified do_while_futures_modify can be used and should have slightly better performance. The following example will show how do_while_futures is used in the decision engine.

# For every compute node from compute_nodes submit a task to gather the node it's information.
# List comprehension is used to store all the futures of the submitted tasks in node_futures.
node_futures = [self.executor.submit(
    self.nova_helper.get_compute_node_by_name,
    node, servers=True, detailed=True)
    for node in compute_nodes]
LOG.debug("submitted {0} jobs".format(len(compute_nodes)))

future_instances = []
# do_while iterate over node_futures and upon completion of a future call
# self._compute_node_future with the future and future_instances as arguments.
self.executor.do_while_futures_modify(
    node_futures, self._compute_node_future, future_instances)

# Wait for all instance jobs to finish
waiters.wait_for_all(future_instances)

Finally, let’s demonstrate how powerful this do_while_futures can be by showing what the compute_node_future callback does. First, it retrieves the result from the future and adds the compute node to the data model. Afterwards, it checks if the compute node has any associated instances and if so it submits an additional task to the DecisionEngineThreadpool. The future is appended to the future_instances so waiters.wait_for_all can be called on this list. This is important as otherwise the building of the data model might return before all tasks for instances have finished.

# Get the result from the future.
node_info = future.result()[0]

# Filter out baremetal nodes.
if node_info.hypervisor_type == 'ironic':
    LOG.debug("filtering out baremetal node: %s", node_info)
    return

# Add the compute node to the data model.
self.add_compute_node(node_info)
# Get the instances from the compute node.
instances = getattr(node_info, "servers", None)
# Do not submit job if there are no instances on compute node.
if instances is None:
    LOG.info("No instances on compute_node: {0}".format(node_info))
    return
# Submit a job to retrieve detailed information about the instances.
future_instances.append(
    self.executor.submit(
        self.add_instance_node, node_info, instances)
)

Without do_while_futures an additional waiters.wait_for_all would be required in between the compute node tasks and the instance tasks. This would cause the progress of the decision engine to stall as less and less tasks remain active before the instance tasks could be submitted. This demonstrates how do_while_futures can be used to achieve more constant utilization of the underlying hardware.

Applier concurrency

The applier does not use the futurist GreenThreadPoolExecutor directly but instead uses taskflow. However, taskflow still utilizes a greenthreadpool. This threadpool is initialized in the workflow engine called DefaultWorkFlowEngine. Currently Watcher supports one workflow engine but the base class allows contributors to develop other workflow engines as well. In taskflow tasks are created using different types of flows such as a linear, unordered or a graph flow. The linear and graph flow allow for strong ordering between individual tasks and it is for this reason that the workflow engine utilizes a graph flow. The creation of tasks, subsequently linking them into a graph like structure and submitting them is shown below.

self.execution_rule = self.get_execution_rule(actions)
flow = gf.Flow("watcher_flow")
actions_uuid = {}
for a in actions:
    task = TaskFlowActionContainer(a, self)
    flow.add(task)
    actions_uuid[a.uuid] = task

for a in actions:
    for parent_id in a.parents:
        flow.link(actions_uuid[parent_id], actions_uuid[a.uuid],
                  decider=self.decider)

e = engines.load(
    flow, executor='greenthreaded', engine='parallel',
    max_workers=self.config.max_workers)
e.run()

return flow

In the applier tasks are contained in a TaskFlowActionContainer which allows them to trigger events in the workflow engine. This way the workflow engine can halt or take other actions while the action plan is being executed based on the success or failure of individual actions. However, the base workflow engine simply uses these notifies to store the result of individual actions in the database. Additionally, since taskflow uses a graph flow if any of the tasks would fail all childs of this tasks not be executed while do_revert will be triggered for all parents.

class TaskFlowActionContainer(...):
    ...
    def do_execute(self, *args, **kwargs):
        ...
        result = self.action.execute()
        if result is True:
            return self.engine.notify(self._db_action,
                                      objects.action.State.SUCCEEDED)
        else:
            self.engine.notify(self._db_action,
                               objects.action.State.FAILED)

class BaseWorkFlowEngine(...):
    ...
    def notify(self, action, state):
        db_action = objects.Action.get_by_uuid(self.context, action.uuid,
                                               eager=True)
        db_action.state = state
        db_action.save()
        return db_action