An algorithmic approach on deadlock detection for enhanced parallelism in multiprocessing systems

This paper presents a non-blocking deadlock detection scheme with immediate knot detection in multiprocessing systems. We assume an expedient state and a special case where each request is limited to one resource unit at a time. Unlike the previous deadlock detection schemes, this new method, using some different data structures takes O(1) time for detecting a knot and O(nm) time for blocking or handling resource release where n and m are the number of processes and that of resources in the system. The deadlock detection latency is thus minimized and is constant regardless of n and m. However, in a multiprocessing system, the operating system can handle the blocking or release on-the-fly running on a separate processor, thus not interfering with user process execution. To some applications where deadlock is concerned, a predictable and zero-latency deadlock detection scheme could be very useful.