Distributed 2-level deadlock detection algorithms and scheduling algorithm for OR-model

Abstract

Algorithms presented in this thesis are based on general request models, OR model and AND/OR model. In the models, each task may wish for any one of a number of resources. For such a general model, this proposes distributed deadlock detection algorithms and a real-time scheduling algorithm. This thesis proposes two deadlock detection algorithms: an algorithm for OR-model and an algorithm for AND/OR model. The algorithm for OR-MODel is useful for a system of processes with CSP-like$^1$ communication. In the model, the existence of a cycle in a wait-for graph is the necessary condition of a deadlocked state. The proposed algorithm does not send control messages for deadlock detection to all dependents at the same time, but it sends a message to only one selected node of several dependents until a cycle is detected. If any cyclic wait-for state is not found, there is not any deadlock. When several separate cycles of wait-for processes are found by several initators, all deadlocks are detected by communication of leader processes of those cycles. For the AND/OR model, there does not appear to be a familiar construct of graph theory to describe a deadlock situation in terms of the wait-for graph. But, the existence of a cycle in a wait-for graph is the necessary condition of grid using energy minimization. To obtain more accurate boundary information, we devise a hexagonal boundary label model modified from the square boundary label model defined by Geman et al.(1990) and the penalty patterns reasonable for the hexagonal grid. Because of smaller quantization errors and reasonable penalty patterns, the results of the hexagonal boundary label model are better than those for the square boundary label model.