Efficient Techniques for Monitoring Distributed Computations

点击翻译按钮获取中文摘要

Given the inherent concurrent, non-centralized nature of distributed systems, monitoring and analyzing their behavior is a complex and costly task. However, it is a powerful capability, which can be utilized for such diverse purposes as process control, and the development and debugging of distributed computations. This research is investigating the development of techniques to reduce the computational, communication and storage costs related to the task of monitoring distributed computations to recognize user-defined event concurrences. This is accomplished by utilizing such general strategies as decentralizing analysis responsibilities and sharing monitoring data. Techniques are being developed to collect and organize information pertaining to the monitored behavior. These strategies incorporate timing knowledge available within the system into schemes to capture consistent views of the states of monitored components. Techniques to integrate separate views of the system state are being utilized to reduce the storage and processing overhead of constructing and maintaining the state views. The placement of monitors throughout the system and the assignment of recognition responsibilities to these monitors is being addressed. Event hierarchies for detecting behavior related to specific process synchronization conditions are being defined. Techniques are being developed to allocate to monitors the responsibility of collecting, organizing and analyzing the data pertaining to the defined activity. The techniques are designed and analyzed based on the interrelated (and sometimes conflicting) goals of limiting the impact of the monitoring activity on the processors where the monitored components reside and on the communication medium. Finally, runtime optimizations are being developed to adapt the monitoring strategies to the actual behavior exhibited by the computation while it is being monitored. The possible use of such techniques as preprocessing and filtering state data is being studied. Techniques to transfer or suspend monitoring responsibilities are being developed as a means of adapting initial monitoring decisions to the behavior of the computation in order to increase the efficiency of the monitoring strategy.