LLgrid: Enabling On-Demand Grid Computing with gridMatlab and pMatlab
Albert Reuther, Tim Currie, Jeremy Kepner, Hahn G. Kim, Andrew P. McCabe, Michael P. Moore, Nadya Travinin · 2004
The concept of grid computing -- back-room computational resources that are as accessible and available as the electric grid -- has gained momentum recently. Numerous Grid computing projects such as NetSolve and Legion have provided infrastructure to enable the launching and monitoring of mostly parameter sweep applications. These Grid computing projects demand that users endure a steep learning curve to program and use the system. Also, these systems draw a strong distinction between the users' computers and the grid computing resources: users' work is done on their own computers, while grid jobs are executed on grid resources. A computational power grid should have characteristics similar to the electric power grids: always available, ubiquitous throughout the organization, and easy to use. The goal of the MIT Lincoln Laboratory Grid (LLgrid) project is to develop an On-Demand Grid Computing capability to address these characteristics and use MATLAB -- the dominant programming language for implementing numerical computations, widely used for algorithm development, simulation, data reduction, testing, and system evaluation -- as its initial target application. MIT Lincoln Laboratory has over one thousand MATLAB users; nearly two hundred users run very long jobs that could benefit from parallel processing. The LLgrid project has developed three technologies that allow these users to run parallel MATLAB jobs transparently on the LLgrid computational resources: MatlabMPI for point-to-point messaging; pMatlab for global array semantics (similar to High Performance Fortran); and gridMatlab for integrating users' computers into the LLgrid and automatically allocating grid computing resources.