Grid Memory Service Architecture for High Performance Computing

Abstract

Increasing network and data intensive applications demand more to computing and capacity. But tight coupling between CPU and makes it hard to share components, resulting in difficulty to improve system performance and reduce system cost. Furthermore, nowadays architecture cannot adapt small computing, big memory feature of application. Supporting by optical interconnect and reconfigurable technology, can be separated with CPU and gathered into server to provide grid service. This paper proposed a grid service (GMS) architecture, which is implemented by box. Design of GMS architecture focuses on high speed, high volume, dynamically shared, high scalability, flexibility and autonomy. Programming model and uniform access interface is designed to satisfy vary clients and application. A serial, packet switching based optical interconnect GMIP is designed to obtain high speed and high scalability. With rich and cheap embedded processor and control logic in box, distributed intelligent is implemented, including distributed processing, multi-level management and active memory. A prototype is implemented as verification platform and the test result illustrates low access latency of box, improved performance than disk, and good effect of active memory.