High performance cycle detection scheme for multiprocessing systems

Abstract

This paper presents a non-blocking deadlock detection scheme with immediate cycle detection in multiprocessing systems. It assumes an expedient state and a special case where each type of resource has one unit and each request is limited to one resource unit at a time. Unlike the previous deadlock detection schemes, this new method takes O(l) time for detecting a cycle and O(n+m) 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 in parallel on a separate processor, thus not interfering with user process execution. To some applications where deadlock is concerned, a predictable and zerolatency deadlock detection scheme could be very useful. © Springer-Verlag 2004.