Contention management for distributed data replication

Optimistic replication schemes provide distributed applications with access to shared data at lower latencies and greater availability. This is achieved by allowing clients to replicate shared data and execute actions locally. A consequence of this scheme raises issues regarding shared data consistency. Sometimes an action executed by a client may result in shared data that may conflict and, as a consequence, may conflict with subsequent actions that are caused by the conflicting action. This requires a client to rollback to the action that caused the conflicting data, and to execute some exception handling. This can be achieved by relying on the application layer to either ignore or handle shared data inconsistencies when they are discovered during the reconciliation phase of an optimistic protocol. Inconsistency of shared data has an impact on the causality relationship across client actions. In protocol design, it is desirable to preserve the property of causality between different actions occurring across a distributed application. Without application level knowledge, we assume an action causes all the subsequent actions at the same client. With application knowledge, we can significantly ease the protocol burden of provisioning causal ordering, as we can identify which actions do not cause other actions (even if they precede them). This, in turn, makes possible the client’s ability to rollback to past actions and to change them, without having to alter subsequent actions. Unfortunately, increased instances of application level causal relations between actions lead to a significant overhead in protocol. Therefore, minimizing the rollback associated with conflicting actions, while preserving causality, is seen as desirable for lower exception handling in the application layer. In this thesis, we present a framework that utilizes causality to create a scheduler that can inform a contention management scheme to reduce the rollback associated with the conflicting access of shared data. Our framework uses a backoff contention management scheme to provide causality preserving for those optimistic replication systems with high causality requirements, without the need for application layer knowledge. We present experiments which demonstrate that our framework reduces clients’ rollback and, more importantly, that the overall throughput of the system is improved when the contention management is used with applications that require causality to be preserved across all actions.