Performance evaluation of garbage collection policies

Abstract

Many modern programming languages allow the programmer to allocate the memory in a simple and transparent way, without the need of the explicit deallocation of the memory when an object is not necessary any more. This is achieved by means of the Garbage Collectors, i.e., special threads that use some algorithm to mark unused objects and then reclaim their space. However, the drawback of this approach is twofold. First, the response time of an application using the garbage collection is generally worse than that of an equivalent that explicitly allocates and deallocates the memory because of the process time required by the garbage collection itself. Second, the algorithms used by the collectors are usually CPU-intensive and hence cause a high consumption of CPU-cycles and a consequent waste of energy.

The goal of this thesis is to tatistically characterise the memory allocation requirements of some classes of applications, to provide numerically tractable models to predict some performance indices of the system, i.e., throughput and average response time, given different garbage collection policies, and to validate those models through a comparison with experimental results.