NSF Award
9261543
Large, complex software applications are pushing the performance limitations of today's engineering workstations. Such applications frequently require virtual memory greatly in excess of available physical memory, and performance becomes dominated by paging time. Paging performance is ultimately dependent on the speed of disk I/O, which runs orders of magnitude slower than the maximum performance possible with physical memory and CPU speed alone. Paging performance is directly related to the number of different pages the application touches (i.e., read or writes) per unit time. The inverse of this measure can be thought of as a measure of "locality of reference." When the number of pages touched per unit time significantly exceeds the amount of free physical memory, performance begins to degrade, sometimes dramatically. This research aims to provide a mechanism for increasing locality of reference in an existing Common Lisp system by clustering objects in virtual memory with the goal of significantly decreasing the number of pages touched per unit time. Specifically, the work is in two areas: (1) prototyping tools for reporting useful information on paging and locality of reference, and (2) to investigating methods for automatically increasing locality of reference during application execution.
