Distributed garbage collection by timeouts and backward inquiry

We present a practical and efficient garbage collection mechanism for large scale distributed systems. The mechanism collects all garbage including distributed cyclic garbage without global synchronization or backward links. The primary method used for local and remote garbage collection is timeouts: each object has a time-to-live, and clients which have a link to an object must refresh the target object within the time-to-live to guarantee that the link will remain valid. For cyclic garbage collection: objects suspected to be garbage are detected by last referenceable timestamp propagation; and cyclic garbage is reclaimed by backward inquiry (back-tracing). Since, without additional overhead, the information about backward references can be obtained during the refreshing process, and since messages necessary for cyclic garbage collection are bundled with the messages used for the refreshing, communication, computation and storage overhead is minimized. This mechanism has been implemented and evaluated on Prospero directory service, and the performance results show that it works well for large scale distributed systems.