Using Deep Neural Networks to Predict Performance of Software Applications

It is of paramount importance that software applications stay performant as the number of users increases due to certain predictable events such as sales, promotional campaigns and seasons. Modeling, typically using Queuing Networks, is commonly used to predict the application’s performance characteristics to help in finding the extra required computing resources.Building Queueing Network Models to represent the various hardware and software components is a complex and ad-hoc task due to the lack of generalized models. Additionally, obtaining model parameters, known as Service Demands, is a time consuming inaccurate process. Furthermore, solving such complex models requires the use of slow and resource consuming simulation techniques. Simplifying models and solutions to rectify these issues results of inaccurate predictions leading to wasted computing resources.In this work, we propose the use of Deep Neural Networks to build software application models to predict performance characteristics under varying workloads. Unlike Queuing Networks, Deep Neural Networks are built using well-established generalized approaches. Additionally, their input data can be obtained from the applications performance monitoring data or using typical performance measuring techniques. Finally, solving Deep Neural Networks is way more efficient than the simulation techniques of Queuing Networks.