A scalable method of parallel tasks after the extension of machine systems based on equal change rate

It is a challenging problem to adapt a parallel computing task to the extended machine systems aiming at making full use of the extended computing resources in reconfigurable computing environments. In this paper, firstly, we focus on the main performance parameters of parallel computing, and then construct a change rate vector of architecture and a change rate vector of parallel task. In addition, using the historical data generated in the process of extending parallel tasks to make it match machine architectures, we employ the regression prediction method to mine the intrinsic characteristics of and relationships among parallel tasks with different scales. Under the assumption that the parallel machine architecture is upgraded or extended in advance, based on the idea of the equal change rate, some strategies of extension and reconstruction for a DAG task are given, and a set of equations about elastic extension for the parallel task is established. Thus, an equal change rate scalable method is proposed in order to adjust the workload of a parallel task and to optimize the algorithm structure according to the change of the host architecture. Our proposed scalable method can reconstruct a parallel task to better match the new machine system in many aspects after the machine system has been upgraded or extended, for example, the computational tasks in the fields of earth simulation, weather forecasting. Simulation experiments show that our method is effective.