Garbage collection for mobile and replicated objects

This paper describes the Hidra garbage collection system fur non-migratable singleton objects, migratable objects and replicated objects. Garbage detection is based on a distributed reference counting protocol and on the use of forwarders when migration occurs. The algorithm is triggered by object reference operations and is completely asynchronous, enabling techniques such as piggy-backing and batching of messages. Once some object is computed as garbage, the ORE sends an unreferenced notification to its implementation. In the case of replicated objects, the unreferenced notification is asynchronously delivered to every replica implementation. The algorithm is safe and live in the absence of failures. When failures arise a reference count reconstruction protocol reestablishes the garbage detection algorithm invariants.