A Family of Distributed Deadlock Avoidance Protocols and Their Reachable State Spaces

We study resource management in distributed systems. Incorrect handling of resources may lead to deadlocks, missed deadlines, priority inversions, and other forms of incorrect behavior or degraded performance. While in centralized systems deadlock avoidance is commonly used to ensure correct and efficient resource allocation, distributed deadlock avoidance is harder, and general solutions are considered impractical due to the high communication overhead. However, solutions that use only operations on local data exist if some static information about the possible sequences of remote invocations is known. We present a family of efficient distributed deadlock avoidance algorithms that subsumes previously known solutions as special instances. Even though different protocols within the family allow different levels of concurrency and consequently fewer or more executions, we prove that they all have the same set of reachable states, expressed by a global invariant. This result enables: (1) a design principle: the use of different protocols at different sites does not compromise deadlock avoidance; (2) a proof principle: any resource allocation protocol that preserves the global invariant and whose allocation decisions are at least as liberal as those of the least liberal in the family, guarantees absence of deadlock.