Title: Transparent Virtualization of Graphics Processing Units

Speaker: Dr. Pavan Balaji

Time:  15：00-17：30pm,  Monday,  Aug. 22, 2011

Venue: Room 337, 3rd Floor, Building No. 5, State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences

Abstract:
Current programming models for Graphics Processing Units (GPUs), such as OpenCL and CUDA, require each computational node to be equipped with one or more local GPUs for applications to be able to use them. Thus, for an application to use GPUs, it has to be allocated on a physical node equipped with one or more GPUs. If the GPU has to be shut down for maintenance, the entire application has to be migrated to another physical node with similar hardware. Recent advances in virtualization techniques, on the other hand, have advocated decoupling the application view of "local hardware resources" from the physical hardware itself, thus allowing applications to transparently utilize remote hardware. Such techniques both improve the utilization of hardware resources (because multiple users could use them simultaneously), and improve the maintainability of these resources by allowing system administrators to dynamically change the mapping of "virtual hardware" to actual "physical hardware". In this seminar, I will talk about VOCL (standing for Virtual OpenCL), a new implementation of the OpenCL programming model that allows applications to view all GPUs in the system (including remote GPUs) as local virtual GPUs. I will describe the overall idea of VOCL, its expected role in GPU virtualization techniques, the research ideas that went into designing and optimizing it, and some performance numbers. I will also briefly talk about some of the extensions for VOCL that we are currently working on to allow efficient coordinated resource management with virtualized GPU resources.

报告人简介：
Dr. Pavan Balaji holds a joint appointment as an Assistant Computer Scientist at the Argonne National Laboratory and a research fellow of the Computation Institute at the University of Chicago. He had received his Ph.D. from the Computer Science and Engineering department at the Ohio State University. His research interests include parallel programming models and middleware for communication and I/O, high-speed interconnects, efficient protocol stacks, cloud computing systems, and job scheduling and resource management. He has more than 80 publications in these areas and has delivered nearly 120 talks and tutorials at various conferences and research institutes. He has received several awards for his research activities including an Outstanding Researcher award at the Ohio State University, the Director's Technical Achievement award at Los Alamos National Laboratory, and several best paper and other awards. Dr. Balaji has also chaired or edited more than 30 journals, conferences and workshops including CCGrid, SC, JHPCA, ICPP, IEEE Micro, Hot Interconnects, P2S2 workshop, and ICCCN; and served as a technical program committee member in numerous conferences and workshops. He is a member of the IEEE and ACM. More details about Dr. Balaji are available at http://www.mcs.anl.gov/~balaji.