Dynamic hybrid active -passive replication.

Abstract

Traditional active and passive replication schemes are widely used to provide fault tolerant distributed data services. However, neither approach directly solves the problem of accessing data in a real-time environment. Many embedded real-time applications, such as computer-aided manufacturing and process control, require timely execution of tasks, and their own processing needs should not be compromised by fault tolerant access to data repositories. In many real-time environments, the schemes employed in conventional replication systems may prove inadequate for the needs of applications. When time is scarce and the overhead for managing redundancy is too high, an alternative solution is required to provide both timing predictability and fault tolerance. Additionally, both schemes are incapable of handling data accesses with hybrid or dynamically changing patterns, which are often required by wide-area applications. If a system requires a mix of active and passive replications or the ratio of read to write operations in the system changes dynamically, then appropriate techniques must be developed to handle these access patterns. This dissertation presents the development of a dynamic hybrid active-passive framework for replica management for real-time systems. The framework combines the advantages of both the active and passive approaches with temporal consistency model, temporal consensus protocol, and probabilistic replication. It addresses the four requirements concerning replica management for real-time systems: timing predictability, data consistency, dynamic data access, and hybrid data access/replication. The key contribution of this dissertation are the introduction of two alternative data consistency models: controlled inconsistency and probabilistic consistency, the development of a suite of replication protocols: real-time primary-backup replication, real-time active replication, and probabilistic replication, and the conceptualization of the mechanism of metamorphosis that makes the dynamic hybrid replication framework possible. We implemented several of the key proposed replication schemes on Open Group's Real-Time Mach operating system within the <italic>x</italic>-kernel architecture as well as on the common UNIX platform with UNIX socket communication support. Significant performance data was collected, extensive evaluation was conducted, and the results show that our models are feasible and highly efficient.