Using Class-Level Static Properties to Predict Object Lifetimes

Today, most modern programming languages such as C # or Java use an automatic memory management system also known as a Garbage Collector (GC). Over the course of program execution, new objects are allocated in memory, and some older objects become unreachable (die). In order for the program to keep running, it becomes necessary to free the memory of dead objects; this task is performed periodically by the GC.

Research has shown that most objects die young and as a result, generational collectors have become very popular over the years. Yet, these algorithms are not good at handling long-lived objects. Typically, long-lived objects would first be allocated in the nursery space and be promoted (copied) to an older generation after surviving a garbage collection, hence wasting precious time.

By allocating long-lived and immortal objects directly into infrequently or never collected regions, pretenuring can reduce garbage collection costs significantly. Current state of the art methodology to predict object lifetime involves off-line profiling combined with a simple, heuristic classification. Profiling is slow (can take days), requires gathering gigabytes of data that need to be analysed (can take hours), and needs to be repeated for every previously unseen program.

This thesis explores the space of lifetime predictions and shows how object lifetimes can be predicted accurately and quickly using simple program characteristics gathered within minutes. Following an innovative methodology introduced in this thesis, object lifetime predictions are fed into a specifically modified Java virtual machine. Performance tests show gains in GC times of as much as 77% for the “SPEC jvm98” benchmarks, against a generational copying collector.