Server overload detection and prediction using pattern classification

One of the key factors in customer satisfaction is the application performance. In traditional settings, it is usually not very difficult to manually detect a performance problem, however, with the advent of ultra-large-scale (ULS) systems [4], manual performance monitoring and prediction becomes tedious and would thus ideally require automation. A typical situation in such a system is depicted by Figure 1, in which a server overload occurs when approximately 500 requests are handled per second by the system. In order to prevent the overloaded state, we should be able to predict this state when the system is handling approximately 400 requests per second, so that it can be scaled up. Automating this prediction is typically hard, because many factors influence performance, and it is typically the human mind that excels at making the right (subjective) decisions based on multiple factors. It is our aim to automate performance prediction, for which we have two distinct goals in mind: (1) warn the system administrator for the need of an impending hardware upscaling and (2) provide an automatic overload prevention mechanism. An application in which such an automated prediction mechanism is very useful is in self-adaptive systems, which are capable of adapting their own behavior according to changes in the environment and the system itself [5]. Having such a mechanism will improve the quality of service as it helps these systems decide when to scale up. In this paper, we propose an approach for server overload prediction. An important aspect of our overload prediction mechanism is the performance monitoring method. Our performance monitoring is based on measuring a wide variety of so-called performance counters [1], such as the Memory\ Available Mbytes and Processor\%Processor Time counters. Rather than defining exact threshold values for the monitored performance counters, we propose to use pattern classification, which can assist with recognizing complex performance counter patterns. Paper accepted as a poster for the 8th International Conference on Autonomic Computing (ICAC 2011)