MPJ Express on InfiniBand

Abstract

Computational science [1] requires vast amounts of computing power to perform huge number of calculations in short span of time [1]. Such a number of calculations mean huge amount of input data and a corresponding amount of output data or information. All this data needs to be catered for. For this very purpose, computational science employs the use of supercomputers [2]. Supercomputers are a vast array of compute nodes connected to each other via a high performance interconnect such as Gigabit Ethernet (GigE) [3], Myrinet eXpress (MX) [4], Quadrics [5] and InfiniBand (IB) [6]. These interconnects are responsible for carrying data from one node to the other for processing. The software technique most commonly used to carry out these tasks is the Message Passing Interface (MPI). Various MPI libraries exist to serve the purpose. Earlier libraries were implemented using FORTRAN programming language. Later as the need arose for more stable and powerful programming language to program bigger machines, C fulfilled the need. And even later with the advent of multi-core compute clusters there was an urgent need for a language inherently capable of threading, platform-independence and portability: Java came into the picture, becoming one of the mainstream HPC programming languages. [7] MPJ Express is one such Java based MPI library [8] [9]. MPJ Express currently runs using Fast Ethernet (FE), GigE and MX. However with the advent of even faster next generation interconnects such as Quadrics and IB, there was a need for MPJ Express to fill the gap and implement a communication device driver for these interconnects. Thus the aim of this research work was to improve MPJ Express so that it can run on InfiniBand. The chapters ahead provide Introduction to the problem domain in detail, Literature Review carried out to find the best possible scenario, an explanation to the Methodology followed, Discussion of Issues encountered, an explanation of Results gathered and lastly the Concluding remarks for this research work along-with the possible Future Work scenarios.