DSM under mach: a quantitative approach

The Mach operating system allows the user to provide an external virtual memory pager. This facility is well suited to implement a distributed shared memory (DSM), allowing parallel programs to run on a network of workstations. The DSM-software has to maintain page coherency between the memories of the cooperating workstations. It is well known that the network traffic caused by page faults plays an important role in the overall DSM performance. Therefore a straightforward implementation of a DSM does not yield the desired performance. Much work has already been done to reduce the amount of data transported over the network, e.g. using a release consistent memory protocol. This paper concentrates on minimizing the time of an individual page fault. The time spent in each of the communication layers involved is measured by instrumenting the DSM-software and the operating system. These times show what can be done to improve performance: replacing the two topmost Mach communication layers and using a tailored UDP protocol reduces the page fault time by 40%. Applications such as matrix multiplication and quicksort show the significant impact of the reduced page fault time on applications