Towards intelligent analysis techniques for object pretenuring

Object pretenuring involves the identification of long-lived objects at or before their instantiation. It is a key optimization for generational garbage collection systems, which are standard in most high performance Java virtual machines. This paper presents a new study of factors that are used to indicate object lifespans. We adopt the information theory measurement of normalized mutual information to compare these various different factors in a common framework. A study of garbage collection traces from four standard Java benchmark programs shows that there is high dependence on some of these factors such as allocation site and object type. We also identify and measure new factors based on object-oriented metrics.