Distributed pacing of shared memory

Abstract

A network of workstations is an attractive platform for highly parallel computations. This is mainly due to the cost and the scalability of such systems. The main problem that is associated with a network of workstations is that the workstations do not physically share resources. As a result some workstations may be running out of memory, even though there may be a large amount of free memory on the network. One solution to this problem is network RAM which allows the workstations to share the available free RAM to be used as a swapping medium by the other workstations. In this thesis, we present NEVIM, a kernel level implementation of a network RAM system for a network of workstations running Linux. NEVIM is integrated into the virtual memory manager of Linux and is totally transparent to the user. NEVIM supports a large number of workstations and automatically does backups of the memory pages.