Deadlock in Process Networks: a Dynamic Detection and Resolution

A Process Network (PN) is an ideal model for streaming-based multimedia and signal-processing applications because of unbounded channel capacities. Mechanisms to implement this model have been proposed but deadlock problems appear. In this paper, we introduce an algorithm with O(n) complexity to solve artificial deadlocks, even when modifications to the evolution of the computation in the PN are considered. The communication cost and the latency depend on the number of candidates for being deadlock detector. In order to reduce the number of messages, the algorithm takes advantage of the collected information while the system evolves to a deadlock situation. Moreover, our algorithm verifies both a safety property and a liveness property. If a deadlock exists, the former property implies that only one node will detect it and the latter involves that one node detects the deadlock eventually.