SHF-Small: Robust Methodologies for Effective Data Center Management

College Of William And Mary, Williamsburg VA

Investigators

Abstract

Despite the ubiquity of data centers, little is known about their effective management. Consolidation of multiple applications with diverse and changing resource requirements is common in data centers as hardware resources are abundant and opportunities for better system usage are plenty, as are opportunities to degrade individual application performance due to unregulated performance interference between applications and system resources. Is it possible to maximize resource usage while respecting individual application performance targets or is it an oxymoron to simultaneously meet such conflicting measures? In this project, a solution methodology to the above difficult problem is proposed using a three-pronged approach. First, a detailed large scale performance study on several thousands of data center servers within a time period that spans two years is going to be conducted. This study provides a micro and macro view of current workload requirements, of workload resource demands on basic resource components including CPU, memory, disk, and their temporal evolution. This analysis provides a baseline for the development of scalable and efficient resource management in data centers. Second, extensive experimentation on basic components of data centers is going to quantify performance interference among different classes of applications due to consolidation. This experimentation drives the development of a light-weight profiler that is system- and application-agnostic. The methodology captures application resource demands via non-intrusive low-level measurements that are provided via standard tools.The experimental observations have the potential to drive the development of resource allocation policies in data centers both at the micro level (i.e., at specific hardware components that are used as data center building blocks) and at the macro level (i.e., at the data center as a whole). Third, a queueing-theory based tool is developed that uses as input the resource demands measured by the profiler to accurately predict application scalability under homogeneous and heterogeneous consolidations. The model can be used to predict the application and system performance under virtualized environments at the micro and macro levels, and provide consolidation suggestions such that pre-defined user- or system-specified performance targets are met. The proposed methodologies have the potential to improve the effectiveness of resource allocation in data centers that operate under complex workloads and show excellent potential for allocation solutions that meet pre-defined user and system performance targets. This research will affect the state-of-the-practice via industrial collaborations, especially IBM Research and NEC Research Labs. More broadly, this research has the potential to make a strong impact in management of in-production data centers. Through this project, several students will be prepared to better meet industry demands in the areas of performance modeling and resource allocation in complex environments.