Memoization of methods using software transactional memory to track internal state dependencies

Abstract

Memoization is a well-known technique for improving the performance of a program, but it has been confined mostly to functional programming, where no mutable state or side-effects exist. Most object-oriented programs, however, are built around objects with an internal state that is mutable over the course of the program. Therefore, the execution of methods often depends on the internal state of some objects or produces side-effects, thus making the application of memoization impractical for object-oriented programs in general. In this paper, we propose an extended memoization approach that builds on the support provided by a Software Transactional Memory (STM) to identify both internal state dependencies and side-effects, hence removing many of the limitations of traditional memoization. We describe the Automatic Transaction-Oriented Memoization (ATOM) system, a thread-safe implementation of our memoization model that requires minimal learning effort from programmers, while offering a simple and customizable interface. Additionally, we describe a memoization advisory system that collects per-method performance statistics with the ultimate goal of aiding programmers in their task of choosing which methods are profitable to memoize. We argue that ATOM is the first memoization system adequate to the unique characteristics of object-oriented programs and we show how memoization can be implemented almost for free in systems that use an STM, presenting the reasons why this synergy can be particularly useful in transactional contexts. We show the usefulness of memoizing object-oriented programs by applying memoization to the STMBench7 benchmark, a standard benchmark developed for evaluating STM> implementations. The memoized version of the benchmark shows up to a 14-fold increase in the throughput for a read-dominated workload.