Power-hungry Graphics processing unit (GPU) accelerators are ubiquitous in high performance computing data centers today. The advent of GPU virtualization frameworks have introduced new opportunities for better management of GPU resources by decoupling them from application execution. However, there are significant challenges in improving the energy efficiency and managing the peak power usage of GPU-enabled server clusters. The underlying system infrastructure shows complex power consumption characteristics depending on the placement of GPU workloads across various compute nodes, power-phases and cabinets in a datacenter. Furthermore, GPU resources need to be scheduled dynamically in the face of time-varying resource demand while keeping the power usage below the peak power constraints. In this talk, Palden will describe the implementation and evaluation of an automated power manager that addresses the above challenges. For GPU virtualization, it uses VOCL, a framework for virtualized execution of OpenCL applications.

Bio:
Palden Lama is a PhD candidate at the Dept. of Computer Science, University of Colorado at Colorado Springs. His research interest is in developing self-managing autonomic resource provisioning methods to control the performance of heterogeneous Internet applications and the power consumption of the underlying servers in a virtualized data center. He is supervised by Dr. Xiaobo Zhou. As a research aide, he is working with Dr. Pavan Balaji, Argonne National Laboratory.