The Apache open source community introduces the unified resource management framework Yarn to share Hadoop clusters, improve their scalability and reliability, and eliminate a performance bottleneck of JobTracker in the early MapReduce framework.
The fundamental idea of Yarn is to split up the two major functionalities of the JobTracker, resource management and job scheduling/monitoring, into separate daemons. The idea is to have a global ResourceManager (RM) and per-application ApplicationMaster (AM).
An application is either a single job in the classical sense of MapReduce jobs or a Directed Acyclic Graph (DAG) of jobs.
Architecture
ResourceManager is the essence of the layered structure of Yarn. This entity controls an entire cluster and manages the allocation of applications to underlying compute resources. The ResourceManager carefully allocates various resources (compute, memory, bandwidth, and so on) to underlying NodeManagers (Yarn's per-node agents). The ResourceManager also works with ApplicationMasters to allocate resources, and works with the NodeManagers to start and monitor their underlying applications. In this context, the ApplicationMaster has taken some of the role of the prior TaskTracker, and the ResourceManager has taken the role of the JobTracker.
ApplicationMaster manages each instance of an application running in Yarn. The ApplicationMaster negotiates resources from the ResourceManager and works with the NodeManagers to monitor container execution and resource usage (CPU and memory resource allocation).
The NodeManager manages each node in a Yarn cluster. The NodeManager provides per-node services in a cluster, from overseeing the management of a container over its lifecycle to monitoring resources and tracking the health of its nodes. MRv1 manages execution of the Map and Reduce tasks through slots, whereas the NodeManager manages abstract containers, which represent per-node resources available for a particular application.
Table 1 describes the components shown in Figure 1.
Name |
Description |
|---|---|
Client |
Client of a Yarn application. You can submit a task to ResourceManager and query the operating status of an application using the client. |
ResourceManager(RM) |
RM centrally manages and allocates all resources in the cluster. It receives resource reporting information from each node (NodeManager) and allocates resources to applications on the basis of the collected resources according a specified policy. |
NodeManager(NM) |
NM is the agent on each node of Yarn. It manages the computing node in Hadoop cluster, establishes communication with ResourceManger, monitors the lifecycle of containers, monitors the usage of resources such as memory and CPU of each container, traces node health status, and manages logs and auxiliary services used by different applications. |
ApplicationMaster(AM) |
AM (App Mstr in the figure above) is responsible for all tasks through the lifcycle of in an application. The tasks include the following: Negotiate with an RM scheduler to obtain a resource; further allocate the obtained resources to internal tasks (secondary allocation of resources); communicate with the NM to start or stop tasks; monitor the running status of all tasks; and apply for resources for tasks again to restart the tasks when the tasks fail to be executed. |
Container |
A resource abstraction in Yarn. It encapsulates multi-dimensional resources (including only memory and CPU) on a certain node. When ApplicationMaster applies for resources from ResourceManager, the ResourceManager returns resources to the ApplicationMaster in a container. Yarn allocates one container for each task and the task can only use the resources encapsulated in the container. |
In Yarn, resource schedulers organize resources through hierarchical queues. This ensures that resources are allocated and shared among queues, thereby improving the usage of cluster resources. The core resource allocation model of Superior Scheduler is the same as that of Capacity Scheduler, as shown in the following figure.
A scheduler maintains queue information. You can submit applications to one or more queues. During each NM heartbeat, the scheduler selects a queue according to a specific scheduling rule, selects an application in the queue, and then allocates resources to the application. If resources fail to be allocated to the application due to the limit of some parameters, the scheduler will select another application. After the selection, the scheduler processes the resource request of this application. The scheduler gives priority to the requests for local resources first, and then for resources on the same rack, and finally for resources from any machine.
Principle
The new Hadoop MapReduce framework is named MRv2 or Yarn. Yarn consists of ResourceManager, ApplicationMaster, and NodeManager.
- ResourceManager is a global resource manager that manages and allocates resources in the system. ResourceManager consists of Scheduler and Applications Manager.
- Scheduler allocates system resources to all running applications based on the restrictions such as capacity and queue (for example, allocates a certain amount of resources for a queue and executes a specific number of jobs). It allocates resources based on the demand of applications, with container being used as the resource allocation unit. Functioning as a dynamic resource allocation unit, Container encapsulates memory, CPU, disk, and network resources, thereby limiting the resource consumed by each task. In addition, the Scheduler is a pluggable component. You can design new schedulers as required. Yarn provides multiple directly available schedulers, such as Fair Scheduler and Capacity Scheduler.
- Applications Manager manages all applications in the system and involves submitting applications, negotiating with schedulers about resources, enabling and monitoring ApplicationMaster, and restarting ApplicationMaster upon the startup failure.
- NodeManager is the resource and task manager of each node. On one hand, NodeManager periodically reports resource usage of the local node and the running status of each Container to ResourceManager. On the other hand, NodeManager receives and processes requests from ApplicationMaster for starting or stopping Containers.
- ApplicationMaster is responsible for all tasks through the lifecycle of an application, these channels include the following:
- Negotiate with the RM scheduler to obtain resources.
- Assign resources to internal components (secondary allocation of resources).
- Communicates with NodeManager to start or stop tasks.
- Monitor the running status of all tasks, and applies for resources again for tasks when tasks fail to run to restart the tasks.
Capacity Scheduler Principle
Capacity Scheduler is a multi-user scheduler. It allocates resources by queue and sets the minimum/maximum resources that can be used for each queue. In addition, the upper limit of resource usage is set for each user to prevent resource abuse. Remaining resources of a queue can be temporarily shared with other queues.
Capacity Scheduler supports multiple queues. It configures a certain amount of resources for each queue and adopts the first-in-first-out queuing (FIFO) scheduling policy. To prevent one user's applications from exclusively using the resources in a queue, Capacity Scheduler sets a limit on the number of resources used by jobs submitted by one user. During scheduling, Capacity Scheduler first calculates the number of resources required for each queue, and selects the queue that requires the least resources. Then, it allocates resources based on the job priority and time that jobs are submitted as well as the limit on resources and memory. Capacity Scheduler supports the following features:
- Guaranteed capacity: As the administrator, you can set the lower and upper limits of resource usage for each queue. All applications submitted to this queue share the resources.
- High flexibility: Temporarily, the remaining resources of a queue can be shared with other queues. However, such resources must be released in case of new application submission to the queue. Such flexible resource allocation helps notably improve resource usage.
- Multi-tenancy: Multiple users can share a cluster, and multiple applications can run concurrently. To avoid exclusive resource usage by a single application, user, or queue, the administrator can add multiple constraints (for example, limit on concurrent tasks of a single application).
- Assured protection: An ACL list is provided for each queue to strictly limit user access. You can specify the users who can view your application status or control the applications. Additionally, the administrator can specify a queue administrator and a cluster system administrator.
- Dynamic update of configuration files: Administrators can dynamically modify configuration parameters to manage clusters online.
Each queue in Capacity Scheduler can limit the resource usage. However, the resource usage of a queue determines its priority when resources are allocated to queues, indicating that queues with smaller capacity are competitive. If the throughput of a cluster is big, delay scheduling enables an application to give up cross-machine or cross-rack scheduling, and to request local scheduling.