Fast detection and resolution of generalized distributed deadlocks

This paper presents an algorithm for detecting generalized deadlocks in distributed systems. Most of the existing algorithms use a diffusing computation technique for deadlock detection in which deadlock detection messages called probes are propagated through the wait-for graph and replies are sent backwards to the initiator of the algorithm carrying the information on wait-for relations between nodes. Based on the collected information, the initiator determines a deadlock. In order to expedite the deadlock detection activity, the proposed algorithm exploits the probes rather than the replies to carry the information required for deadlock detection. As the probes are propagated, the unblocking function of each node is merged and carried onto the message. As a result, time and communication costs are reduced by up to half of those of the existing algorithms. Furthermore, deadlock resolution is simplified in our scheme with no additional message rounds as in other algorithms