A high productivity framework for parallel data intensive computing in Matlab

Programmer's productivity is recognized as one of the most significant challenges in the effective use of high-performance computing. Higher level languages like Matlab are being increasingly adopted. However, it has significant shortcomings when used for large-scale computationally intensive applications that require very high performance and/or significant amounts of memory. Developing efficient runtime frameworks to aid the high-level languages to make them scalable to larger problem sizes is an effective solution to this problem. Our solutions, mexMPI, GAMMA and LA enable parallel computing directly in Matlab for high-performance while retaining its productivity aspects. mexMPI provides message passing semantics to enable parallel computing within Matlab environment using high-performance networks. GAMMA presents a distributed shared memory programming model wherein the programmer developes his/her parallel algorithms using a 'Get-Compute-Put model' and 'LA' is a runtime framework that enable users to develop large scale applications directly in Matlab. We demonstrate the effectiveness of our frameworks using NAS benchmarks. The experimental evaluation of these frameworks indicate effectiveness of our approach.