Empirical Studies of Distributed Garbage Collection

Garbage collection has been a necessary evil of computer languages that employ dynamic data structures. The 1990’s has seen some significant technology shifts that present new challenges for automatic garbage collection. In particular, distributed systems are providing an economical approach to parallel processing. Many distributed garbage collection algorithms have been proposed but few have been analysed or evaluated. To take advantage of distributed systems, precise evaluation of new distributed garbage collection algorithms is prudent. This thesis presents an empirical evaluation of the several important distributed garbage collection algorithms. All proposals for distributed garbage collectors stem from uniprocessor collectors. For this reason, uniprocessor garbage collectors are first surveyed and classified. Then we give a survey of distributed garbage collection schemes and develop a taxonomy in terms of the issues of distribution. In the light of the surveys, we describe the goals of evaluation by simulation and the performance metrics of interest. We simulate and evaluate nine important distributed garbage collection schemes giving empirical measurements of their performances. As a result of the evaluation, we describe and evaluate two improved distributed collectors. The first is an improved version of the Lermen and Maurer scheme that can reduce communication overhead by up to 3.8478%. The second is an improvement of Weighted Reference Counting that reduces communication overhead by up to 26.5656%. The improved Weighted Reference Counter performs better than original Lermen and Maurer scheme by as much as 11.0076%. It performs better than improved Lermen and Maurer scheme by up to 7.5396%. Of all the distributed garbage collection schemes simulated, Shapiro et al scheme came out as the best in terms of all the factors considered. However, the implementation does not collect inter-node cycles of garbage. The only collector that reclaims inter-node cycles of garbage has up to 34.3454% additional communication overhead. The cost of collecting inter-node cycles is no more than the cost of using the improved Weighted Reference Counting scheme.