Hadoop: YARN Federation

Purpose

YARN is known to scale to thousands of nodes. The scalability of YARN is determined by the Resource Manager, and is proportional to number of nodes, active applications, active containers, and frequency of heartbeat (of both nodes and applications). Lowering heartbeat can provide scalability increase, but is detrimental to utilization (see old Hadoop 1.x experience). This document described a federation-based approach to scale a single YARN cluster to tens of thousands of nodes, by federating multiple YARN sub-clusters. The proposed approach is to divide a large (10-100k nodes) cluster into smaller units called sub-clusters, each with its own YARN RM and compute nodes. The federation system will stitch these sub-clusters together and make them appear as one large YARN cluster to the applications. The applications running in this federated environment will see a single massive YARN cluster and will be able to schedule tasks on any node of the federated cluster. Under the hood, the federation system will negotiate with sub-clusters resource managers and provide resources to the application. The goal is to allow an individual job to “span” sub-clusters seamlessly.

This design is structurally scalable, as we bound the number of nodes each RM is responsible for, and appropriate policies, will try to ensure that the majority of applications will reside within a single sub-cluster, thus the number of applications each RM will see is also bounded. This means we could almost linearly scale, by simply adding sub-clusters (as very little coordination is needed across them). This architecture can provide very tight enforcement of scheduling invariants within each sub-cluster (simply inherits from YARN), while continuous rebalancing across subcluster will enforce (less strictly) that these properties are also respected at a global level (e.g., if a sub-cluster loses a large number of nodes, we could re-map queues to other sub-clusters to ensure users running on the impaired sub-cluster are not unfairly affected).

Federation is designed as a “layer” atop of existing YARN codebase, with limited changes in the core YARN mechanisms.

Assumptions:

• We assume reasonably good connectivity across sub-clusters (e.g., we are not looking to federate across DC yet, though future investigations of this are not excluded).
• We rely on HDFS federation (or equivalently scalable DFS solutions) to take care of scalability of the store side.

Architecture

OSS YARN has been known to scale up to about few thousand nodes. The proposed architecture leverages the notion of federating a number of such smaller YARN clusters, referred to as sub-clusters, into a larger federated YARN cluster comprising of tens of thousands of nodes. The applications running in this federated environment see a unified large YARN cluster and will be able to schedule tasks on any nodes in the cluster. Under the hood, the federation system will negotiate with sub-clusters RMs and provide resources to the application. The logical architecture in Figure 1 shows the main components that comprise the federated cluster, which are described below.

Running Applications across Sub-Clusters

When an application is submitted, the system will determine the most appropriate sub-cluster to run the application, which we call as the application’s home sub-cluster. All the communications from the AM to the RM will be proxied via the AMRMProxy running locally on the AM machine. AMRMProxy exposes the same ApplicationMasterService protocol endpoint as the YARN RM. The AM can request containers using the locality information exposed by the storage layer. In ideal case, the application will be placed on a sub-cluster where all the resources and data required by the application will be available, but if it does need containers on nodes in other sub-clusters, AMRMProxy will negotiate with the RMs of those sub-clusters transparently and provide the resources to the application, thereby enabling the application to view the entire federated environment as one massive YARN cluster. AMRMProxy, Global Policy Generator (GPG) and Router work together to make this happen seamlessly.

The figure shows a sequence diagram for the following job execution flow:

1. The Router receives an application submission request that is complaint to the YARN Application Client Protocol.
2. The router interrogates a routing table / policy to choose the “home RM” for the job (the policy configuration is received from the state-store on heartbeat).
3. The router queries the membership state to determine the endpoint of the home RM.
4. The router then redirects the application submission request to the home RM.
5. The router updates the application state with the home sub-cluster identifier.
6. Once the application is submitted to the home RM, the stock YARN flow is triggered, i.e. the application is added to the scheduler queue and its AM started in the home sub-cluster, on the first NodeManager that has available resources. a. During this process, the AM environment is modified by indicating that the address of the AMRMProxy as the YARN RM to talk to. b. The security tokens are also modified by the NM when launching the AM, so that the AM can only talk with the AMRMProxy. Any future communication from AM to the YARN RM is mediated by the AMRMProxy.
7. The AM will then request containers using the locality information exposed by HDFS.
8. Based on a policy the AMRMProxy can impersonate the AM on other sub-clusters, by submitting an Unmanaged AM, and by forwarding the AM heartbeats to relevant sub-clusters. a. Federation supports multiple application attempts with AMRMProxy HA. AM containers will have different attempt id in home sub-cluster, but the same Unmanaged AM in secondaries will be used across attempts. b. When AMRMProxy HA is enabled, UAM token will be stored in Yarn Registry. In the registerApplicationMaster call of each application attempt, AMRMProxy will go fetch existing UAM tokens from registry (if any) and re-attached to the existing UAMs.
9. The AMRMProxy will use both locality information and a pluggable policy configured in the state-store to decide whether to forward the resource requests received by the AM to the Home RM or to one (or more) Secondary RMs. In Figure 1, we show the case in which the AMRMProxy decides to forward the request to the secondary RM.
10. The secondary RM will provide the AMRMProxy with valid container tokens to start a new container on some node in its sub-cluster. This mechanism ensures that each sub-cluster uses its own security tokens and avoids the need for a cluster wide shared secret to create tokens.
11. The AMRMProxy forwards the allocation response back to the AM.
12. The AM starts the container on the target NodeManager (on sub-cluster 2) using the standard YARN protocols.

Configuration

To configure the YARN to use the Federation, set the following property in the conf/yarn-site.xml:

Running a Sample Job

In order to submit jobs to a Federation cluster one must create a seperate set of configs for the client from which jobs will be submitted. In these, the conf/yarn-site.xml should have the following additional configurations:

Any YARN jobs for the cluster can be submitted from the client configurations described above. In order to launch a job through federation, first start up all the clusters involved in the federation as described here. Next, start up the router on the router machine with the following command:

  $HADOOP_HOME/bin/yarn --daemon start router

Now with $HADOOP_CONF_DIR pointing to the client configurations folder that is described above, run your job the usual way. The configurations in the client configurations folder described above will direct the job to the router’s client RM port where the router should be listening after being started. Here is an example run of a Pi job on a federation cluster from the client:

  $HADOOP_HOME/bin/yarn jar hadoop-mapreduce-examples-3.0.0.jar pi 16 1000

This job is submitted to the router which as described above, uses a generated policy from the GPG to pick a home RM for the job to which it is submitted.

The output from this particular example job should be something like:

  2017-07-13 16:29:25,055 INFO mapreduce.Job: Job job_1499988226739_0001 running in uber mode : false
  2017-07-13 16:29:25,056 INFO mapreduce.Job:  map 0% reduce 0%
  2017-07-13 16:29:33,131 INFO mapreduce.Job:  map 38% reduce 0%
  2017-07-13 16:29:39,176 INFO mapreduce.Job:  map 75% reduce 0%
  2017-07-13 16:29:45,217 INFO mapreduce.Job:  map 94% reduce 0%
  2017-07-13 16:29:46,228 INFO mapreduce.Job:  map 100% reduce 100%
  2017-07-13 16:29:46,235 INFO mapreduce.Job: Job job_1499988226739_0001 completed successfully
  .
  .
  .
  Job Finished in 30.586 seconds
  Estimated value of Pi is 3.14250000......

The state of the job can also be tracked on the Router Web UI at routerhost:8089. Note that no change in the code or recompilation of the input jar was required to use federation. Also, the output of this job is the exact same as it would be when run without federation. Also, in order to get the full benefit of federation, use a large enough number of mappers such that more than one cluster is required. That number happens to be 16 in the case of the above example.