Machine Learning-Based Prefetch Optimization for Data Center Applications
The dynamic nature of data centers, which sees frequent release of applications and upgrading of servers, makes it difficult to fine tune their performance. Yet, the significance of performance tuning cannot be overlooked as even a single-digit performance improvement of a data center can greatly reduce both cost as well as power consumption for organizations. This paper is an attempt to study the different processor prefetch configurations and their effectiveness in improving the performance of memory system and the overall data center. The study is based on the worst and best configurations data, ranging from 1.4% to 75.1%, taken from11 important data center applications. The comparison shows a wide performance gap between the two. The paper concluded that the parameter value optimization problem can be tackled by judiciously exploiting the machine learning-based framework, designed and implemented during the study. The viability of this framework is shown through multiple experiments that helped to get a clear understanding of the sensitivity of the framework with regard to several influential components such as machine learning algorithms, training datasets selection and modeling or problem formulation methods. The framework achieved performance within 1% of the best performance of any single configuration for the same set of applications.