Object Lifetime Prediction in Java

Accurately predicting the lifetimes of objects in object-oriented and functional languages is important because such predictions can be used to improve memory management performance at run-time. A typical approach to this prediction problem is first to observe object lifetimes by tracing a sample program execution, then to construct a predictor function based on these observations, and finally to test the predictions on reference program executions. Four quantitative measures characterize this approach: coverage, the proportion of objects in the reference program execution for which the predictor function is defined; accuracy, the fraction of predicted lifetimes that are correct; precision, the granularity of the predictions; and size, the size of the predictor itself. These four properties are not independent; for example, increased precision often leads to less coverage and accuracy. We describe a fully precise prediction method and report experimental results on its performance. By “fully precise” we mean that the granularity of predictions is equal to the smallest unit of allocation. We show that for a number of benchmark programs in the Java programming language, fully precise prediction can be achieved, together with high coverage and accuracy. Our results also show that a significant proportion of objects have a measured lifetime of zero, a fact which a dynamic compiler could use to avoid explicit allocation. The method described here is the first to combine high-precision and efficiency in a single lifetime predictor.