1986 Reports

Replication and Nested Transactions in the Eden Distributed System

Pu, Calton

Hardware redundancy in distributed systems offers the potential for increased availability and performance, but this requires software support if the full potential is to be realized. We have designed and implemented two mechanisms for such support. The first provides crash-resistant resources, replicated transparently and consistently to increase the availability of distributed data. To update multiple copies despite down nodes, we have introduced the Regeneration method used in the implementation of a replicated system directory. Regeneration restores inaccessible copies elsewhere in the network, maintains the availability of resources, and adapts to configuration changes. The second mechanism is a systell1 supporting nested transactions, which can manage the complex failure modes in a distributed system, synchronize concurrent resource access internal to applications, and facilitate safe module composition. In the tree-structured nesting, each transaction has a Transaction Manager (TM), responsible for the concurrency control and crash recovery of its subtransactions. Many concurrency control and recovery techniques can be combined in this TM Tree design framework. We chose locking and versions for the first implementation. Using Eden objects and the replicated directory, our nested transactions provide consistent concurrent access 10 location-independent, crash- resistant resources. In summary, the principal contributions of this research are the Regeneration method and the TM Tree framework. Regeneration uses the separation of hardware repair from data restoration to increase replicated data availability. TM tree composes existing techniques to derive many difficult designs for nested transaction. Both have been proven in the design and implementation of actual systems.