high performance computing on graphics processing units: hgpu.org

Nowadays computer applications are becoming heavier and require, at the same time, real-time results. The Heterogeneous clusters with their computing power represent a good solution to this request. However, it is possible that during the execution, a computing element of the cluster becomes defaulting, needs maintenance, or that the load needs to be re-balanced. In this paper, we propose a migration strategy for relocating the execution of a task to another computing element. In particular, we are interested in remap nodes of Data Flow Graph (DFG), representing Digital Signal Processing (DSP) application, onto heterogeneous (CPU-GPU) clusters while keeping up the flow of data and minimizing the temporal perturbation. For our approach, we give a lower bound for the flow of data after the migration and, validate it by the real-time construction of visual saliency map from video input.