Fault-tolerant Clock Synchronization for Distributed Systems with High Message Delay Variation

Fault-tolerant clock synchronization is an important requirement in many distributed systems , especially in time-critical and safety-critical applications. Frequently, interactive convergence algorithms are used for fault-tolerant clock synchronization, providing advantages such as fully distributed operation, low message exchange overhead, and simplicity of implementation. This paper presents the measured performance of three interactive convergence clock synchronization algorithms. Our experiments were conducted in a distributed UNIX environment featuring high message delay variation, which poses severe constraints on the clock synchronization tightness that may be achieved. The algorithms that were tested are: FTMA (fault-tolerant midpoint algorithm) 1], AEFTMA (adaptive exponential averaging fault-tolerant midpoint algorithm) 2], and SWA (sliding window algorithm) 3]. Our experimental results indicate that SWA outperforms the other algorithms in this environment, being able to achieve tighter synchronization under diierent simulated fault conditions. The superiority of SWA can be attributed to its high degree of fault tolerance, combined with its ability to treat messages with much longer than expected delays as faults. In distributed systems, computers cooperate to provide the expected functionality to a given application. Some tasks that are often found in such systems are: synchronizing activities that occur at diierent points of the system, ordering events in time, enforcing deadlines, and measuring elapsed time. A system with one or more of these requirements must use proper synchronization mechanisms to establish an agreed-upon global time scale among its components. Particularly in the case of safety-critical applications, synchrony must be maintained in spite of the presence of faults in the system. Frequently, fault-tolerant clock synchronization is achieved via interactive convergence algorithms in which nodes exchange their clock values and determine clock correction terms at regular intervals. This paper presents the measured performance of three interactive convergence algorithms: the sliding window algorithm (SWA) 3], the fault-tolerant midpoint algorithm (FTMA) 1], and the adaptive exponential averaging fault-tolerant midpoint algorithm (AEFTMA) 2]. The measurements were carried out with an application-level implementation running in a distributed UNIX environment 2]. This environment poses some practical constraints to clock synchronization, in particular a high variation in the