Dedicated distributed memory servers

Availability of very high bandwidth in local area networks justifies consideration of network resources as potential components of a progressively tightly coupled distributed processing system. Without approaching the high performance limit where a distributed system would be considered completely equivalent to a parallel multiprocessor, this research explores the implications of the availability of high speed networks on memory organization at the network level and the introduction of dedicated distributed memory servers in the storage hierarchy as an approach to improve overall loosely coupled process performance. Simulation is used to study the behavior of this class of servers and investigate their effect on high performance distributed computations. This work especially focuses on utilization of remote memory for applications with memory size requirements far in excess of the configuration offered by any single node in the network. Results indicate that dedicated distributed memory servers are a viable solution to problems posed by workstation memory limitations for applications with very high memory demands. Furthermore, simulations show that availability of a high-bandwidth network not only speeds up processing, it also makes the dedicated distributed memory servers more scalable and allows maximum server utilization for a given application. Finally, implemented simulations also demonstrate that mixture distributions provide a superior representation for computational service times than representations based on any standard single analytical distribution