An Acceleration Toolkit of MATLAB Based on Hybrid CPU/GPU Clusters

The resource demand of MATLAB increases as rapidly as the problem size of user applications. To resolve this problem, many acceleration toolkits based on CPU or GPU clusters have been proposed for MATLAB. However, they cannot support users to simultaneously use both of CPU and GPU for resolving the same problem, and also cannot adapt the type of used resources according to the properties of user applications. As a result, the computational power of available resources is not completely explored, and user applications cannot obtain the best performance. On the other hand, they usually use all of the CPUs or GPUs available in a cluster for executing any MATLAB instructions without considering if it is the most helpful for user applications or not. Consequently, it is possible to waste resources while users cannot obtain a better performance. To resolve these problems, we propose a novel acceleration toolkit based on hybrid CPU/GPU clusters for MATLAB in this paper. This toolkit can automatically adapt the types and numbers of resources used for different MATLAB instructions by resource pruning, and balance the workload of used resources. Consequently, it can effectively exploit both of CPUs and GPUs for improving the performance of MATLAB.