Processing dependent tasks on a Heterogeneous GPU resource architecture

The development of General Purpose GPUs (GPGPU) provides multiple cores of computing power and high memory bandwidth that can meet the high performance requirements of researchers in diverse areas such as weather forecasting, bioinformatics, and even nuclear simulation. In this work we present a Heterogeneous GPU computing system (HG) which can utilize GPUs with different capabilities. Firstly, the HG quantifies the capability of each GPU, and then assigns an acceptable workload to each GPU. The HG is different from other distributed GPU systems because HG can process dependent tasks which indicates the tasks in HG can communicate with each other. Using this architecture HG can deal with more complex task dependent applications. To validate our approach, in the first set of experiments we show that multiple GPU performs better in complex computing applications. Next we use 2D heat equation application to compare HG with a MPI program running on CPUs. With increasing work set size, computing power of HG results in a much better performance than the CPUs system.