The complexity of enterprise data centers has grown steadily in recent years, primarily due to continuous evolution resulting from small incremental changes to system hardware and software. This project focuses on a modeling and analysis approach for improving the manageability of complex data centers. The project focuses on (i) scalable, adaptive monitoring techniques for very large data centers, (ii) multi-scale modeling of complex data-center applications, and (iii) model-driven analysis to predict system performance, bottlenecks and capacity requirements. The adaptive monitoring approach involves the use of statistical data mining and learning to dynamically vary what systems are monitored and when. The multi-scale models involve capturing system behavior at large, medium and small scales by modeling application interactions at different levels of abstraction: at the data center level, at the application level, and at the individual server level. Model-driven analysis brings together adaptive monitoring and the multi-scale models for capacity provisioning, bottleneck detection and performance prediction. The modeling and analysis techniques will be validated using real-world data center applications running on a Linux-based data center testbed. Broader impacts include source code release and sharing of anonymized trace data to enable further experimentation by other researchers and the participation of undergraduate students in this research via summer REU projects.
