Data Mining Based Diagnosis in Resource Management

There are different solutions to resource allocation problems in Resource Management Systems (RMS). One of the most sophisticated ways to solve these problems is an adjustment to Quality-of-Service (QoS) settings during runtime. These settings affect the trade-off between the resource usage and the quality of the services the executed tasks create. But, to be able to determine the optimal reactive changes to current QoS settings in an acceptable time, knowledge of the resource allocation problem’s cause is necessary. This is especially significant in an environment with real-time constraints. Without this knowledge other solutions could be initiated, still an improvement to the current resource allocation, but the optimal compromise between resource requirements and QoS is likely to be missed. A resource management system (RMS) with the ability to adjust QoS settings can solve more resource allocation problems than one providing reallocation measures only. But problem-depending only optimal changes to QoS settings can solve the problem within timing constraints and thus prevent expensive system failures. Depending on the environment a RMS is used in, the failures could be a huge financial loss or even a threat to human lives. But the knowledge of a problem’s cause does not only help to solve the problem within existing timing constraints and to guarantee feasibility of the executed tasks, but helps to maximize the quality of the generated services as well. To detect upcoming problems in time, forecasting mechanisms can be integrated into the RMS. They can predict a problem in the near future, early enough for the system to react. For diagnosis of resource management problems, data mining can be applied to determine the cause of an allocation problem. The techniques implemented in this work are the k-nearest neighbor analysis and decision trees. Both techniques will make their predictions based on prior created resource allocation snapshots referring to problem cases with known cause.