Cognitive performance application

摘要

This work shows that combining the techniques of neural networking and predictive analytics with the fundamental concepts of computing performance optimization is genuine in many ways. It has the potentials to: (1) reduce infrastructure upgrade costs (2) reduce human interactions, by enabling the system to learn, analyze, and make decisions on its own, and (3) generalize the solutions to other performance problems. This paper attempts to tackle a JVM performance optimization from a different dimension and in a way that can be scaled to other common utilized resources, such as file systems, static contents, search engines, web services...etc. It shows how to build a framework that monitors the performance metrics to determine patterns leading to bottleneck incidents and then benchmark these performance metrics. The framework uses artificial neural network in its core to accomplish this first steps with immediate benefit of eliminating the need to a domain expert analyzing which of these metrics is more important or has more weight on constituting the bottleneck condition, and hence enable the system to deal with more ambiguous situations. The framework uses an analytics engine, to establish predictive patterns between the system bottleneck and library of factors in order to establish an early alert system and thus enhancing the weight of the bottleneck signal. Finally the framework acts in defense when the deadlock signal is triggered from the learning and/or the analytics engine through streaming down concurrent transactions into a temporarily queuing data structure. We put our model into a test and built a simulation to quantify the added benefit of each component of our framework. The results are proven to demonstrate the immediate benefit of our framework and open doors for other future work.