Optimal management of dynamic buffer caches

Abstract

Optimal memory management strategies such as VMIN are generally considered unrealizable in view of the impracticality of obtaining a computation's reference string prior to execution. Addressing such strategies, this paper focuses on the dynamic management of variable-size buffer caches in the framework of the locality-set model, a memory management model that characterizes the reference behavior of a computation in terms of locality sets rather than reference strings. Several cost measures—the number of page faults, the space-time product, and one that combines them—are considered and conditions under which they are equivalent are derived. We define two novel strategies, PSETVMIN and SETVMIN, which manage buffer caches with and without prepaging, respectively, and prove that they minimize a cost measure that takes both page faults and the space-time product into account. The two strategies are of theoretical interest in view of their optimal behavior, but—more importantly—they are also realizable since only a limited amount of information about the reference behavior of a computation, the locality-set sequence, is required in advance. We demonstrate the use of these strategies for join processing in relational database management systems. The performance benefits of this technique are discussed and illustrated by simulation results.