Comparing SpMV for solver applications

In this paper, we propose a new re-ordering technique for improving the performance of Sparse Matrix Vector Multiplication (SpMV) for systems supported with Graphics Processing Units (GPUs). We conducted the test by applying SpMV on solver based applications which are widely used in the domain of engineering and science. We studied and analyzed the existing representations and storage structures of SpMV to identify the areas of improvement. We compare the performance of two solver based applications, Page-rank and Conjugate Gradient Solver (CGS) using various SpMV representations. We obtained a speedup in the range - over the existing implementations. We concluded that our approach proves to be better not only for power law characteristics data sets but also for an general class of characteristics data sets.