Replication-based incremental copying collection

Abstract

We introduce a new replication-based copying garbage collection technique. We have implemented one simple variation of this method to provide incremental garbage collection on stock hardware with no special operating system or virtual memory support. The performance of the prototype implementation is excellent: major garbage collection pauses are completely eliminated with only a slight increase in minor collection pause times. Unlike the standard copying algorithm, the replication-based method does not destroy the original replica when a copy is created. Instead, multiple copies may exist, and various standard strategies for maintaining consistency may be applied. In our implementation for Standard ML of New Jersey, the mutator continues to use the from-space replicas until the collector has achieved a consistent replica of all live data in to-space. We present a design for a concurrent garbage collector using the replication-based technique. We also expect replication-based ge methods to be useful in providing services for persistence and distribution, and briefly discuss these possibilities.