Research On Cluster-NUMA System

Distributed parallel computer system is a complex system, which has many co-operative, interactive, sub-systems. Such a system is also an open system, opened to the remote node inside the system, and opened to administration and the end user outside. The main problem related to the performance of the parallel system is how to develop, config and get use of the resources in the system.Cluster and NUMA system are developed very fast in the past few years. Such a system can offer a higher comprehensive performance. The advantages of NUMA is its manageability and usability. Well cluster is a union of independent PCs, workstations or SMPs, acting as a single computing or processing resource, and its availability, scalability and re-usability of the cluster system is better.In this dissertation, based on the fully understanding the organization of the present parallel machines, integrating the developing trend of computer technique, advance research is down on several key problems of the parallel machine, including the structure of node, communication interface, single system image. Also, Cluster-NUMA, a mew system structure which views the usability as the main focus of system design, is present, the aim of the system is to its scalability, high availability, and flexibility to configure and re-construction. Cluster-NUMA combines both the advantages of cluster and tradition NUMA machine. Another contribution of this dissertation is making a thorough inquiry on implementing such a system on the multiple layered structure, especially on some fundamental problems such as hardware communication mechanism, information exchange between system, and resource and load sharing, and partial system function is implemented to spur and verify the research result.