Performance Evaluation and Analysis for Conjugate Gradient Solver on Heterogeneous (Multi-GPUs/Multi-CPUs) platforms

High performance computing (HPC) presents a technology that allows solving high intensive problems in a reasonable period of time, and can offer many advantages for large applications in various fields of science and industry. Current multi-core processors, especially graphic processing units (GPUs), have quickly evolved to become efficient accelerators for data parallel computing. They can maintain parallel programmability and provide high computing throughput. In this paper, the authors present an implementation and performance analysis of sparse iterative linear solver on heterogeneous multi-CPUs/multi-GPUs architectures using PARALUTION and StarPU libraries. More particularly, the authors compare the performance of parallel preconditioned conjugate gradient (PCG) solver on different platforms. Experimental results have been conducted using GPU platforms and show a significant speed up compared to central processing units CPUs implementations. In order to provide the highest performance, the system supports Multi-CPU/Multi-GPU architectures, where it scales up very high.