A simulation study of heuristic node reordering for collective I/O

As the scale of parallel computing systems grows, the importance of collective I/O also increases because it provides the function of single file based parallel I/O. Collective I/O follows two-phase I/O scheme that consists of a data exchange phase and an I/O phase. The previous studies of collective I/O were usually related with improving the performance of each phase. Especially, Heuristic Node Reordering (HNR) is the technique to improve the data exchange performance where a different number of cores per node are used for a parallel application. In this study, we implement the HNR simulator and analyze the effectiveness of HNR under various cases. As a result, we find out the number of nodes, the maximum number of cores and the shape of I/O data array are the important factors related to improving the HNR performance.