Page-Level Cooperative Garbage Collection

Programs written in garbage-collected languages like Java often have large working sets and poor locality. Worse, a full garbage collection must visit all live data and is likely to touch pages that are no longer resident, triggering paging. The result is a pronounced drop in throughput and a spike in latency. We show that just a slight reduction in available memory causes the throughput of the SPECjbb benchmark to drop by 66% and results in garbage collection pauses lasting for several seconds. This paper introduces page-level cooperative garbage collection, an approach in which the garbage collector works with the virtual memory manager to limit paging. We present a novel, cooperative garbage collection algorithm called Hippocratic collection. By communicating with the virtual memory manager and “bookmarking” objects, the Hippocratic collector eliminates paging caused by garbage collection. Our approach requires only modest extensions to existing virtual memory management algorithms. We present empirical results using our modified Linux kernel. We show that the Hippocratic collector runs in smaller footprints than traditional collectors while providing competitive throughput. Under memory pressure, Hippocratic collection improves the throughput of SPECjbb by a factor of two over the next best garbage collector.