Category Archives: JobTracker

Yarn provides greater scalability than MapReduce

MapReduce could have resource utilization problems because an arbitrary process could have allocated all map slots, while some reduce slots are empty (and vise-versa). Yarn (Yet Another Resource Negotiator) splits the JobTracker into a global ResourceManager (RM) and a per-application ApplicationMaster (AM) which works with the per-machine NodeManagers to execute and monitor tasks.  The ResourceManager has a Scheduler which only schedules (does not monitor).

The ApplicationMaster is far more scalable in 2.0 than 1.0. HortonWorks has successful simulated a 10k node cluster. This is possible due to the ApplicationMaster not being global to the entire cluster but rather has an instance per application so is no longer a bottleneck through which all applications must pass.

The ResourceManager is also more scalable in 2.0 since it’s scope is reduced to scheduling and no longer is responsible for fault tolerance of the entire cluster.


What benefit does Yarn bring to the existing MapReduce?

Within the classic MapReduce is the Job Tracker component. Yarn splits Job Tracker into two further components: Resource Manager (aka RM) (allocating cpu, ram, etc) and Node Manager (aka NM) (which operates at the level of a single node/machine). The Application Manager (aka AsM) negotiates resources from the Resource Manger and with the Node Manager to execute tasks. Job Tracker is already an ancient architecture — five years old!!

Yarn is sometimes referred to as MapReduce 2.0 or MRv2.

Resource Manager supports hierarchical application queues to guarantee allocation ratios of cluster resources. However, it does not enable recovery from application or hardware failures. It does not monitor. It only schedules. Scheduling methods include FIFO (default) and Capacity. Fair is not currently supported.

ZookKeeper monitors Resource Manager in order to switch to a secondary if Resource Manager itself fails. In a failover scenario, running applications are restarted and the queue continues. Preservation of state within currently running applications is handled by checkpoints stored by the Application Master within HDFS.

Rather than having specific containers to execute Map jobs and Reduce jobs, Yarn enables containers for more generic jobs, which enables developers to write other applications that run on the cluster.

It’s unclear whether Yarn will make the system run faster or slower. Generalization and modularization usually comes at a cost. However, Yarn allows for more complete utilization of CPU and RAM resources so in theory can squeeze every last bit of capacity out of a cluster, whereas the fixed size containers in MapReduce 1.0 could have left some resources idle. Yarn does not mange I/O which is typically a bigger bottleneck than RAM. There’s also no management of network bandwidth in Yarn. (Note to self, got to figure this out: I saw another article that says that Yarn does manage cpu, disk and network, yet didn’t mention RAM).

Another benefit of a more modularized architecture is that it makes the system easier to maintain. Any updates to MapReduce 1.0 requires the replacement of a pretty big chunk of software. Being able to run multiple versions of MapReduce within a cluster of thousands of nodes is important. Significant downtime would otherwise be required for upgrades.


Selecting the right hardware for a Hadoop cluster

I’m summarizing this article. For specifics (such as how to configure split machines across racks to better configure the network switches) see the article. None of this content is operating system or hardware vendor specific, but generally the discussions assume Linux.

Goal is to minimize data movement and process on the same machine that stores the data. Therefore each machine in the cluster needs appropriate CPU and disk. Problem is that when building the cluster the nature of the queries and the resulting bottlenecks may not yet be known. If a business is building its first Hadoop cluster, it may not yet fully understand the types of business problems that will eventually be solved by it. That’s in contrast to a business deploying it’s Nth Oracle server.

Types of bottlenecks:

  1. IO: reading from disk (or a network location)
    • indexing
    • data import/export
    • data transformation
  2. CPU: processing the Map query
    • clustering
    • text mining
    • natural language processing

Other issues, since a cluster could eventually scale to hundreds or thousands off machines

  1. Power
  2. Cooling

The Cloudera Manager can provide realtime statistics about how a currently running MapReduce job impacts the CPU, disk, and network load.

Roles of the components within a Hadoop cluster:

  1. Name Node (and Standby Name Node): coordinating data storage on the cluster
  2. Job Tracker: coordinating data processing
  3. Task Tracker
  4. Data Node

Data Node and Task Tracker

  • The vast majority of the machines in a cluster will only peform the roles of Data Node and Task Tracker, which should not be run on the same nodes as Name and Job.
  • Other components (such as HBase) should only be run on the Data Nodes if they operate on data. You want to keep data local as much as possible. HBase needs about 16 GB Heap to avoid garbarge collection timeouts. Impala will consume up to 80% of available RAM.
  • Assumed to be lower performance machines than the Name Node and Job Tracker

Name Node and Job Tracker

  • Standby Name Node should (obviously) not be on the same machine as the Name Node.
  • Name Node (and Standby Name Node) and Job Tracker should be enterprise class machines (redundant power supplies, enterprise class raid’ed disks)
  • Name Node should have RAM in proportion to number of data blocks in the cluster. 1GB RAM for every 1 million blocks in HDFS. With 100 Data Node cluster, 64 GB RAM is fine. Since the machine’s tasks will be disk intensive, you’ll want enough RAM to minimize virtual memory swapping to disk.
  • 4 – 6 TB of disk, not raid’ed (JBOD configuration)
  • 2 CPUs (at least quad code). Recommend more CPUs and/or cores as opposed to faster CPU speed, since in a large cluster the higher speed will draw more power and generate more heat, yet not scale as well as if there were simply more CPUs or better yet nodes.

Cloudera has defined four standard configurations

  1. Light Processing (I’m not sure what the use case is for this. Prototype? Sandbox?)
  2. Balanced Compute (recommeded for your 1st cluster, since it’s not likely you’ll properly identify which configuration is best suited for your use case)
  3. Storage Heavy
  4. Compute Heavy


Apache Hadoop and its components

Hadoop consists of two components

  1. MapReduce -
    • programming framework
    • Map
      • distributes work to different Hadoop nodes
    • Reduce
      • gathers results from multiple nodes and resolves them into a single value
      • the source come from HDFS, and the output is typically written back to HDFS
    • Job Tracker: manages nodes
    • Task Tracking: takes orders from Job Traker
    • MapReduce originally developed by Google.
    • Apache MapReduce is built on top of Apache YARN which is a framework for job scheduling and cluster resource management.
  2. HDFS (Hadoop Distributed File System) – file store
    • It is neither a file system nor a database, it’s neither yet it’s both.
    • Within HDFS are two components
      • Data Nodes:
        • data repository
      • Name Nodes:
        • where to find the data; maps blocks of data on slave nodes (where job and task trackers are running)
        • Open, Close, Rename files
    • On top of HDFS you can run HBase
      • Super scalable (billions of rows, millions of columns) repository for key-value pairs
      • This is not a database, cannot have multiple indices