Java in High Performance computing: Cluster computing using Java

Java in High Performance computing: Cluster computing using Java.

this page compiled by YoonJung Ha and Kavan Mehta for the Underdog Group Project
DS420 - Fall 2000, Professor Clark Elliott

List of general reference for this section.

Due to Java’s platform-independence execution model, its support for networking, multithreading and mobile code, it seen the next logical step would be High Performance computing in Java.

In this section we try to present brief introduction to various approaches and associated tools to implement Cluster Computing with Java Virtual Machine.

We contrast these approaches form the viewpoint on using cluster computing to implement a scalable Java application like Java Servers.

In order to implement cluster computing, the Java application must have the Illusion that the cluster is a single computing resource, even though the application will execute on the multiple, independent, nodes composing the cluster.

There are three different approaches to enable Java Application to see single cluster.

Above JVM: Using standard libraries such as RMI, Corba.

Remote Method Invocation allows the distributed applications to be written in the way that the methods of remote Java objects can be invoked from other Java Virtual Machines on different hosts on the network [Qusay H.Mahmoud. Distributed Programming with Java. Manning, 1999]. These libraries provide support for implementation of distributed system with supporting seamless remote invocations on objects in different Java virtual machine and callbacks from servers to clients, integrating the distributed object model into the Java language, making differences between the distributed object model and the local Java object model apparent and preserving the safety provided by the Java run-time environment. Some libraries have an approach for implementing RMI, based on native compilation, and allow for better optimization eliminating the need for processing of type information at run time and making a light weight communication protocol possible.

However intent of these libraries is to facilitate implementation of distributed system and not so much as to implement dedicated clusters. This next section we discuss Java Proactive Library to implement cluster on JVM.

Typical RMI Implementation

Advantages of Above JVM approach:

Standard libraries, therefore maintain portability and interpretability of the application.

Drawback of Above JVM approach:

In order to implement cluster, programmers have to explicit spell out every bit of details. This approach error prone and inherently inefficient.
Modification to Java application code is required to implement it in a Distributed fashion.
Restricts the use of Java semantics, there by preventing optimizations of Java application.

At JVM: Example cJVM Cluster Java Virtual Machine

This approach provides an implementation of a Java virtual machine (Java Platform), which itself cluster enabled.

cJVM (Cluster JVM) is a cluster-enabled implementation of a Java Virtual Machine. It provides a single system image of the cluster to a Java application.

Each cJVM process holds certain portion of application objects and thread.

Collection of All cJVM processes as a whole forms a cJVM (Cluster JVM). (Individual cJVM process not cJVM)

Benefits of At JVM approach:

Provides SSI (single system image) to Java application. In other word any Java application could be implemented on cJVM without any modification to the code. Extends " Code Once run every where" to cluster computation.
By working at the level of the virtual machine, it provides a venue to exploit opportunities for optimizations based upon the semantics of Java. (i.e. optimization is not restricted by use external libraries like RMI etc )
Provides scalable the architecture.

Drawback of At JVM approach:

As the standard JVM is made cluster enabled (i.e. modified or replaced by cJVM), it is hard to implement.

Below JVM: Using cluster enabled infrastructure: This approach uses a cluster-enabled platform below the JVM to implement Cluster. Attached diagrams show the architecture of Sun HPC ClusterTools 3.1^TM, one the many commercial platform that are cluster enabled.

Benefits of Below JVM

Provides SSI (single system image) that any pure Java application can use. In other word any pure Java application can be implemented on cluster enabled platform without any modification to the code. Extends " Code Once run every where " to cluster computation.

Completely hides from the application the management and formation of the Clusters.

Provides scalable the architecture.

Drawback:

It require dedicated platform. Therefore harder to implement.
Java application program has no control formation and management of cluster. Therefore restricts optimizations based upon the semantics of Java.

this page compiled by YoonJung Ha and Kavan Mehta for the Underdog Group Project DS420 - Fall 2000, Professor Clark Elliott

this page compiled by YoonJung Ha and Kavan Mehta for the Underdog Group Project
DS420 - Fall 2000, Professor Clark Elliott