Sketch based Distributed Garbage Collection, Theory and Empirical Evaluation
Résumé
Distributed Garbage Collection (DGC) algorithms are fundamental components of modern object-oriented distributed systems. The design of a DGC raises numerous issues in term of performance and scalability. DGC object-migration approaches are known to be inefficient compared to graph cycle detection approaches because of the tremendous costs associated with object migration compared to graph-fragment migration where only the object identifiers are carried over the network. Yet, even object identifiers have a large memory footprint in practice. In this paper, we introduce the idea of sketched Cycle Detection Message (CDM). We show that explicit graph-fragment representations are very costly compared to the sketch-based ones. We prove, under reasonable assumptions that sketch-based messages are smaller (up to one order of magnitude) than their explicit counterparts. Those results apply to most of state-of-art DGC methods. The improvements brought by this approach is discussed in details in the particular case of the Veiga and Ferreira DGC algorithm. In this case, the amount of improvement is more than a factor~3 under very limited assumptions.