SYSTEM FOR PREDICTING PROCESS DEADLOCK STATES IN COMPUTER SYSTEMS

Abstract

The article presents the architecture and experimental evaluation of a system for predicting process deadlock states in computer systems. In contrast to reactive mechanisms that detect cyclic resource dependencies only after a deadlock has been formed, the proposed system is intended to identify a hazardous tendency in advance by analyzing current process signatures. The architecture integrates operating-system data access, process-change monitoring, process-signature construction, current and historical data storage, automated fuzzy-rule-base construction, deadlock-state prediction, and preventive control. The training procedure based on Fuzzy C-Means clustering is separated from real-time prediction, in which priority is given to processes currently waiting for a resource. Four output classes are used: normal, potentially risky, pre-deadlock, and critical. For the pre-deadlock and critical classes, the preventive action is restricted to restarting a non-system process with the shortest residence time in the system, which reduces the amount of discarded computation while breaking a potentially dangerous resource-waiting sequence. The system was evaluated by discrete-event simulation and by experiments on a physical server. For workloads containing 5% and 10% potentially deadlocking processes, the repeated real-system experiments yielded Accuracy values of 99.82% and 99.65%, Precision of 98.39% and 98.49%, and F1-score of 98.19% and 98.25%, respectively. The prediction was generated on average 3.1-3.6 ms before the actual deadlock, while the average processing time was 8.6-8.8% lower than that of reactive resource-allocation-graph analysis.