Page History

Introduction

HAMA is a distributed computing framework based on BSP (Bulk Synchronous Parallel) computing techniques for massive scientific computations (e.g., matrix, graph, network, ..., etc), Currently being incubated as one of the incubator project by the Apache Software Foundation.

HAMA Architecture

Hama is consists of three major components: BSPMaster, GroomServers and Zookeeper. The below diagram displays a overview of the overall architecture and shows which components communicate with each other. The components will be explained in detail in the course of this document.

Components

Apache Hama, based on Bulk Synchronous Parallel model\[1\], comprises three major components: 

• BSPMaster
• GroomServer
• Zookeeper.
  

It is very similar with Hadoop architecture, only except the portion of communication and synchronization mechanisms.

In a normal usecase the user submits a so called "Job" which is a definition of how to run a computation. A job once submitted will have multiple tasks that are launched across the cluster.Image Removed

BSPMaster

BSPMaster is responsible to do for the following:

• Maintaining its own state.
• Maintaining groom server status.
• Controlling super steps Maintaining supersteps and other counters in a cluster.
• Maintaining job progress information jobs and tasks.
• Scheduling Jobs and Assigning assigning tasks to groom servers
• Disseminating Distributing execution class classes and configuration across groom servers.Controlling fault.
• Providing users with the cluster control interface (web and console based).
  

A BSP Master and multiple grooms are started by the script. Then, the bsp master starts up with a RPC server for to which groom servers can dynamically register itself. Groom servers starts up with a BSPPeer instance - later, BSPPeer needs to be integrated with GroomServer - and a RPC proxy to contact the bsp master. After started, each groom periodically sends a heartbeat message that encloses its groom server status, including maximum task capacity, unused memory, and so on.

Each time the bsp master receives a heartbeat message, it brings up-to-date groom server status - the bsp master makes use of groom servers' status in order to effectively assign tasks to idle groom servers - and returns a heartbeat response that contains assigned tasks and others actions that a groom server has to do. For now, we have a FIFO job scheduler and very simple task assignment algorithms.

GroomServer

A [Groom Server|GroomServer] (shortly referred to as groom) is a process

 that manages life cycle of bsp tasks assigned by BSPMaster. Each groom contacts the BSPMaster, and reports task statuses by means of periodical piggybacks with BSPMaster. Each groom is designed to run with HDFS or other distributed storages. Basically, a groom server and a data node should run on one physical node to get the best performance for data-locality. Note that in a massive parallel environment, the benefit of data locality is lost when large amount of virtual processes must be multiplexed onto physical processes\[2\].

Zookeeper

A Zookeeper is used to manage the efficient barrier synchronisation synchronization of the BSPPeers. Later, it will also be used for the area of a fault tolerance system.

BSP Programming Model

A BSP program consists of a sequence of supersteps. Each superstep consists of the three phases below:

• Local computation
• Process communication
• Barrier synchronization
  

These phases will be defined by user in a bsp() method. Let's see the below example code:

public class BSPEaxmple {

  public static class MyBSP extends BSP {

    @Override
    public void bsp(BSPPeer bspPeer) throws IOException, KeeperException,
        InterruptedException {
      // 1. Do something locally
      
      // 2. Sends/receives data to/from neighbor nodes
      bspPeer.send(peerName, msg);

      while ((message = bspPeer.getCurrentMessage()) != null) {
         byte[] data = message.getData();
      }

      // 3. Barrier synchronization
      bspPeer.sync();
    }

    @Override
    public Configuration getConf() {
      return conf;
    }

    @Override
    public void setConf(Configuration conf) {
      this.conf = conf;
    }    

  }
  
  // BSP job configuration 
  public void main(String[] args) throws Exception {
    BSPJob bsp = new BSPJob(new HamaConfiguration(), BSPEaxmple.class);
    // Set the job name
    bsp.setJobName("My BSP Job");
    bsp.setBspClass(MyBSP.class);

    // Submit job
    BSPJobClient.runJob(bsp);
  }
}  

Communication

Barrier Synchronization

An BSP examples

Communication and Synchronization Process

Each BSP task has a set of Outgoing Message Manager and Incoming Queue.

Outgoing Message Manager collects the message to be sent, serializes it, compresses it and puts it in a bundles. At barrier synchronization phase, each BSP task exchanges the bundles, deserializes it, decompresses it and puts it into the Incoming Queue.

System Diagram

Image Added

1. BSPMaster starts up
2. GroomServer starts up
3. ZooKeeper cluster starts up
4. GroomServer dynamically registers itself to BSPMaster
5. GroomServer forks/ manages BSPPeer(s)
6. BSPPeers communicate/ perform barrier synchronization through ZooKeeper cluster.

Reference

\[1\]. Valiant, Leslie G., A bridging model for parallel computation. 

\[2\]. David B. Skillicorn, Jonathan M. D. Hill, and W. F. [McColl]. Questions and Answers about BSP. Scientific Programming, 6(3):249-274, Fall 1997.

...