A modern cloud system, such as an application server or database, includes hundreds of knobs that engineers must tune to get the best performance out of the system. Besides, these systems are dynamic in different dimensions. They are continually updating its architecture as well as its incoming workload changes over time. Usually, the ops team tries to tailor all dynamics with a single and static apps configuration. This configuration is supposed to capture the average of all possible environmental and application changes. However, using a single and static configuration for the application-life eventually leads it to waste or suffer from the lack of resources. Such performance engineering requires expert knowledge and substantial time. In this project, we ask the question: can we use statistical methods to identify the best parameter configuration for a given system?

In collaboration with IBM, we are developing SmartTuning, a tool that tunes cloud applications autonomously. SmartTuning observes and classifies the workload flowing into the application and continually does proactive and continuous experimentation of the application's different configurations, looking for those that optimize it for the given workload type. SmartTuning does statistical analysis to find out the configuration that tailors the application best. The analysis considers three key aspects of the application: the previous configurations applied to the application, the application's performance measured for those different configurations, and the unique characteristics of the previous workloads observed flowing into the application. Based on these aspects, SmartTuning samples a new configuration that is likely to improve the application performance given a current workload being observed. A key advantage of the SmartTuning approach is that it identifies different workloads if there are changes over time and adapts the configuration accordingly, keeping the application tuned automatically. Since SmartTuning observes and tunes automatically, it minimizes the need for the engineering team to instrument their code or try esoteric configuration on their environments continually to maximize the performance of running their application in a cost-effective manner.

People

1. Adalberto R. Sampaio Junior
2. Mayank Tiwary