Collecting a Distributed Transactional Object Heap

Distributed systems with a large number of nodes use internode reference counting for timely and fault-tolerant garbage collection. However, this fails to collect cyclic garbage distributed across nodes. One fix is to migrate all objects on a garbage cycle to a single node, where they can be collected by the tracing-based local collector. Existing proposals based on this technique have practical problems due to unnecessary migration of objects. We propose a scheme that avoids migration of live objects, batches objects to avoid a cascade of migration messages, and short-cuts the migration path to avoid multiple migrations. We use simple estimates to detect objects that are highly likely to be cyclic garbage and to select a node to which such objects are migrated. The scheme collects all distributed cyclic garbage, has low overhead, and preserves the decentralized and fault-tolerant nature of distributed reference counting and migration.

We present a scalable garbage collection scheme for systems that store objects at multiple servers while clients run transactions on locally cached copies of objects. It is the first scheme that provides fault tolerance for such a system: Servers recover from failures and retrieve information needed for safe garbage collection; clients do not recover from failures, yet the scheme is able to reclaim objects referenced only from failed clients. The scheme is optimized to reduce overhead on common client operations, and it provides fault tolerance by doing work in the background and during client operations that are infrequent.

Thor is a distributed object-oriented database where objects are stored persistently at highly-available servers called object repositories (ORs). In a large Thor system, performance tuning and system reconfiguration dictate that objects must be able to migrate among ORs. The paper describes two schemes for object references that support object migration, one using location-independent names and the other, location-dependent names. The paper analyzes the performance of the two schemes and concludes that location-dependent names are the right choice for systems like Thor, where we want fast access to objects that have migrated.