A Wait-Free Probe Algorithm for Distributed Deadlock Detection and Resolution

Edge-chasing is the basis of many deadlock detection algorithms. This method detects a deadlock by propagating special messages called probes along dependency edges. When the initiator of a probe receives the probe back, it knows the existence of a deadlock. Once a deadlock is detected, a special message called token is sent to clean up those probes in the deadlock cycle which, if not removed, may later lead to phantom deadlock detections. Only after the token has traversed the entire deadlock cycle and returned to its initiator, the deadlock is resolved by aborting a so-called victim in the deadlock cycle. In a deadlock, all involved transactions are held waiting and all involved resources are locked up. It is thus desirable to resolve a deadlock as soon as it is detected, without waiting for the token message to go around the deadlock cycle. This paper proposes an algorithm that achieves this and thereby reduces the average deadlock persistence time by as much as two thirds.