Oversubscription

High-priority user-facing services are typically provisioned on large clusters for peak load and unexpected load spikes. Hence, for most of time, the provisioned resources remain underutilized. Oversubscription takes advantage of temporarily unused resources to execute best-effort tasks such as background analytics, video/image processing, chip simulations, and other low priority jobs.

How does it work?

Oversubscription was introduced in Mesos 0.23.0 and adds two new agent components: a Resource Estimator and a Quality of Service (QoS) Controller, alongside extending the existing resource allocator, resource monitor, and Mesos agent. The new components and their interactions are illustrated below.

Oversubscription overview

Resource estimation

  • (1) The first step is to identify the amount of oversubscribed resources. The resource estimator taps into the resource monitor and periodically gets usage statistics via ResourceStatistic messages. The resource estimator applies logic based on the collected resource statistics to determine the amount of oversubscribed resources. This can be a series of control algorithms based on measured resource usage slack (allocated but unused resources) and allocation slack.

  • (2) The agent keeps polling estimates from the resource estimator and tracks the latest estimate.

  • (3) The agent will send the total amount of oversubscribed resources to the master when the latest estimate is different from the previous estimate.

Resource tracking & scheduling algorithm

  • (4) The allocator keeps track of the oversubscribed resources separately from regular resources and annotate those resources as revocable. It is up to the resource estimator to determine which types of resources can be oversubscribed. It is recommended only to oversubscribe compressible resources such as cpu shares, bandwidth, etc.

Frameworks

  • (5) Frameworks can choose to launch tasks on revocable resources by using the regular launchTasks() API. To safe-guard frameworks that are not designed to deal with preemption, only frameworks registering with the REVOCABLE_RESOURCES capability set in its framework info will receive offers with revocable resources. Further more, revocable resources cannot be dynamically reserved and persistent volumes should not be created on revocable disk resources.

Task launch

  • The revocable task is launched as usual when the runTask request is received on the agent. The resources will still be marked as revocable and isolators can take appropriate actions, if certain resources need to be setup differently for revocable and regular tasks.

NOTE: If any resource used by a task or executor is revocable, the whole container is treated as a revocable container and can therefore be killed or throttled by the QoS Controller.

Interference detection

  • (6) When the revocable task is running, it is important to constantly monitor the original task running on those resources and guarantee performance based on an SLA. In order to react to detected interference, the QoS controller needs to be able to kill or throttle running revocable tasks.

Enabling frameworks to use oversubscribed resources

Frameworks planning to use oversubscribed resources need to register with the REVOCABLE_RESOURCES capability set:

FrameworkInfo framework;
framework.set_name("Revocable framework");

framework.add_capabilities()->set_type(
    FrameworkInfo::Capability::REVOCABLE_RESOURCES);

From that point on, the framework will start to receive revocable resources in offers.

NOTE: That there is no guarantee that the Mesos cluster has oversubscription enabled. If not, no revocable resources will be offered. See below for instructions how to configure Mesos for oversubscription.

Launching tasks using revocable resources

Launching tasks using revocable resources is done through the existing launchTasks API. Revocable resources will have the revocable field set. See below for an example offer with regular and revocable resources.

{
  "id": "20150618-112946-201330860-5050-2210-0000",
  "framework_id": "20141119-101031-201330860-5050-3757-0000",
  "agent_id": "20150618-112946-201330860-5050-2210-S1",
  "hostname": "foobar",
  "resources": [
    {
      "name": "cpus",
      "type": "SCALAR",
      "scalar": {
        "value": 2.0
      },
      "role": "*"
    }, {
      "name": "mem",
      "type": "SCALAR",
      "scalar": {
        "value": 512.0
      },
      "role": "*"
    },
    {
      "name": "cpus",
      "type": "SCALAR",
      "scalar": {
        "value": 0.45
      },
      "role": "*",
      "revocable": {}
    }
  ]
}

Writing a custom resource estimator

The resource estimator estimates and predicts the total resources used on the agent and informs the master about resources that can be oversubscribed. By default, Mesos comes with a noop and a fixed resource estimator. The noop estimator only provides an empty estimate to the agent and stalls, effectively disabling oversubscription. The fixed estimator doesn't use the actual measured slack, but oversubscribes the node with fixed resource amount (defined via a command line flag).

The interface is defined below:

class ResourceEstimator
{
public:
  // Initializes this resource estimator. This method needs to be
  // called before any other member method is called. It registers
  // a callback in the resource estimator. The callback allows the
  // resource estimator to fetch the current resource usage for each
  // executor on agent.
  virtual Try<Nothing> initialize(
      const lambda::function<process::Future<ResourceUsage>()>& usage) = 0;

  // Returns the current estimation about the *maximum* amount of
  // resources that can be oversubscribed on the agent. A new
  // estimation will invalidate all the previously returned
  // estimations. The agent will be calling this method periodically
  // to forward it to the master. As a result, the estimator should
  // respond with an estimate every time this method is called.
  virtual process::Future<Resources> oversubscribable() = 0;
};

Writing a custom QoS controller

The interface for implementing custom QoS Controllers is defined below:

class QoSController
{
public:
  // Initializes this QoS Controller. This method needs to be
  // called before any other member method is called. It registers
  // a callback in the QoS Controller. The callback allows the
  // QoS Controller to fetch the current resource usage for each
  // executor on agent.
  virtual Try<Nothing> initialize(
      const lambda::function<process::Future<ResourceUsage>()>& usage) = 0;

  // A QoS Controller informs the agent about corrections to carry
  // out, but returning futures to QoSCorrection objects. For more
  // information, please refer to mesos.proto.
  virtual process::Future<std::list<QoSCorrection>> corrections() = 0;
};

NOTE The QoS Controller must not block corrections(). Back the QoS Controller with its own libprocess actor instead.

The QoS Controller informs the agent that particular corrective actions need to be made. Each corrective action contains information about executor or task and the type of action to perform.

Mesos comes with a noop and a load qos controller. The noop controller does not provide any corrections, thus does not assure any quality of service for regular tasks. The load controller is ensuring the total system load doesn't exceed a configurable thresholds and as a result try to avoid the cpu congestion on the node. If the load is above the thresholds controller evicts all the revocable executors. These thresholds are configurable via two module parameters load_threshold_5min and load_threshold_15min. They represent standard unix load averages in the system. 1 minute system load is ignored, since for oversubscription use case it can be a misleading signal.

message QoSCorrection {
  enum Type {
    KILL = 1; // Terminate an executor.
  }

  message Kill {
    optional FrameworkID framework_id = 1;
    optional ExecutorID executor_id = 2;
  }

  required Type type = 1;
  optional Kill kill = 2;
}

Configuring oversubscription

Five new flags has been added to the agent:

Flag Explanation
--oversubscribed_resources_interval=VALUE The agent periodically updates the master with the current estimation about the total amount of oversubscribed resources that are allocated and available. The interval between updates is controlled by this flag. (default: 15secs)
--qos_controller=VALUE The name of the QoS Controller to use for oversubscription.
--qos_correction_interval_min=VALUE The agent polls and carries out QoS corrections from the QoS Controller based on its observed performance of running tasks. The smallest interval between these corrections is controlled by this flag. (default: 0ns)
--resource_estimator=VALUE The name of the resource estimator to use for oversubscription.

The fixed resource estimator is enabled as follows:

--resource_estimator="org_apache_mesos_FixedResourceEstimator"

--modules='{
  "libraries": {
    "file": "/usr/local/lib64/libfixed_resource_estimator.so",
    "modules": {
      "name": "org_apache_mesos_FixedResourceEstimator",
      "parameters": {
        "key": "resources",
        "value": "cpus:14"
      }
    }
  }
}'

In the example above, a fixed amount of 14 cpus will be offered as revocable resources.

The load qos controller is enabled as follows:

--qos_controller="org_apache_mesos_LoadQoSController"

--qos_correction_interval_min="20secs"

--modules='{
  "libraries": {
    "file": "/usr/local/lib64/libload_qos_controller.so",
    "modules": {
      "name": "org_apache_mesos_LoadQoSController",
      "parameters": [
        {
          "key": "load_threshold_5min",
          "value": "6"
        },
        {
      "key": "load_threshold_15min",
      "value": "4"
        }
      ]
    }
  }
}'

In the example above, when standard unix system load average for 5 minutes will be above 6, or for 15 minutes will be above 4 then agent will evict all the revocable executors. LoadQoSController will be effectively run every 20 seconds.

To install a custom resource estimator and QoS controller, please refer to the modules documentation.