Performance Improvement of Large-Scale Interconnection Network Simulator by Using GPU

In a parallel computing system that uses a lot of computing nodes, the performance of interconnection network is important since communication among computing nodes has a large effect on the performance of whole systems. Therefore performance evaluation of interconnection network prior to the implementation of the actual interconnection network is necessary. In order to evaluate the performance of interconnection network, a performance simulation by software is useful in terms of easily changeable conditions and at low cost. However, a software simulation of large-scale interconnection network requires enormous computation and takes a lot of time. This paper introduces a simulator which can simulate large-scale interconnection network, and we accelerate the simulator using GPU. To implement the parallelized simulator, this paper adopts cellular automata in modeling the behavior of interconnection networks and a table-based simulation method. The simulation principle naturally exploits parallelism and well matches the GPGPU acceleration. Furthermore, this paper describes some tuning techniques to exploit the maximum performance of GPU devices. For example, the tuning techniques include the modification of data structure for coalesced accesses and the reduction in the number of global memory access. The parallelized simulator that is applied all tuning techniques achieved the speed up ratio of maximum 37.8 times compared to a sequential simulator, and simulation speed achieved 7.0M[cycles*nodes/sec].