Adaptive Deadlock Detection and Resolution in Real-Time Distributed Environments

In real-time distributed transaction processing, deadlocks must be detected and resolved. Timeouts are not a viable option because they lead to lost work and missed deadlines. We have proposed a suite of deadlock detection protocols and a resolution protocol which carry a varying degree of (generally low) overhead. The protocols behave differently under varying load conditions and transaction characteristics. Further, the invocation period of these protocols can be controlled to improve performance when the overhead tends to become large. The performance of these protocols has been demonstrated using a distributed real-time transaction processing simulator which provides an interactive interface to set parameters, and also has provisions to select from a variety of concurrency control, priority assignment, workload distribution and other protocols. The impact of transaction workload, underlying system configuration, resource availability, detection rates, and congestion on each of the proposed protocols is observed and presented. In general, the multi-cycle detection protocol demonstrated the most superior performance over a broad range of parameters. The results presented in this paper were obtained from over 147,000 simulations.