Design and Evaluation of a Distributed Shared Memory Network for Application-Specific PC Cluster Systems

摘要

Recently, several gigabit Ethernet network interfaces that support remote direct memory access (DMA) data transfer and TCP/IP offload engines (TOE) have been released. However, a convenient distributed shared memory (DSM) for a cluster computer is not readily available, despite proposals for cluster network interface cards (NICs) offering a variety of functions to suit high-performance computers. In particular, a PC cluster comprising general PCs was proposed for scientific computing. We focus on the development of a reconfigurable DSM suitable for applications involving scientific computations. In this paper, we propose a reconfigurable DSM, wherein the connections between the nodes can be reconfigured according to the problem. The reconfigurable DSM generates connection routes according to the problem and the network switch connects the computer resources at each node. To demonstrate an implementation example of the proposed network switch, we implemented a network switch using an FPGA development board kit. We found that an 18-port reconfigurable network switch can be realized. In addition, we simulated the network switch as a reconfigurable DSM. Simulation results for a sparse matrix reveal that the frequency of collisions in the communication route is low.