This thesis describes a virtual memory management mechanism for large scale MIMD multiprocessor systems with proper performance. In the next decade, the prices of the computer will drop so low that 10, 20, or perhaps 100 powerful microprocessors per user will be feasible. Although there have been many virtual memory management systems since the Atlas system, they are not appropriate to such multiprocessor systems.
In this thesis, we describe the design and implementation of a virtual memory management mechanism for No Remote Memory Access architecture system that has several hundreds nodes. We discuss the issues to provide user with the virtual memory space on that system, design a virtual memory management mechanism which will be a solution of those issues, and implement a prototype on the basis of that design.
We propose the virtual swap server for swapping services with proper bandwidth in that scale multiprocessor system. And we exploit the trading system for naming services of that paging manager - for dynamic binding between paging manager(eg, loader server, virtual swap server, and static swap server) and processes that use swapping services for paging activities. The virtual swap server is made with idle remote memory that exists in remote node in the NoRMA system, provides and withdraws swapping services to all system through trading services.
From the simulation of performance model, we found that the virtual swap server is useful for proper performance of virtual memory and the trading services help the virtual memory management system to be scalable; the trading doesn't affect the scalability of the system, and the utilization of the interconnection network can be managed by trader. As the results, we provide a virtual memory management mechanism with proper performance for the NoRMA architecture system that has several hundreds nodes.
