Scale-right Provisioning

Mazin Yousif, Systems Technology Lab, Corporate Technology Group, Intel Corporation

Tuesday, April 4
12:15pm - SENSQ 5317

Refreshments at noon

Hosted by Rami Mellhem

Abstract

This talk first summarizes the diversity of research in the Systems Technology Lab (STL) in the Corporate Technology Group (CTG) at Intel Corporation and then zooms into one research topic referred to as Scale-right Provisioning (SrP). STL researchers look into building real, valuable, and a long-term differentiation for Intel platforms focusing on research related to power management (e.g., efficient power budgeting and low power computing), platforms physicals (e.g., thermal and acoustics), infrastructures (e.g., many cores and partitioning) and capabilities (e.g., reliability and virtualization), as well as trust (e.g. content protection and managed identities), manageability and autonomics. In addition to conducting internal research, STL has established a number of collaboration efforts with selected professors in academic institutions through grants or joint research. SrP refers to the ability to independently scale up/down compute and I/O resources to efficiently react to runtime changes (e.g., workload and fault) and to enforce high-level policies related to Quality of Service (QoS), power management and service delivery. The goal is to significantly reduce Total Cost of Ownership (TCO) through improved resource utilization, streamlining deployment, enabling autonomics and utility computing. To realize these goals, three approaches will be discussed here; all of which could be deployed on a collection of standalone servers or blade servers in an enclosure. Approach I is based on running a Virtual Machine Monitor (VMM) on each server with the VMM launching a number of Virtual Machines (VM). Approach II is based on running a single VMM across all the blades (or servers) in an enclosure, referred to as an Enclosure-Wide VMM (E-VMM). Approach III is similar to approach II, but instead of relying on software to build the E-VMM, a firmware/hardware approach is adopted. Here a set of Resources Aggregation Capabilities (RAC) are defined to extend firmware and chipsets. The latter two approaches degenerate the servers into a large pool of compute and I/O resources (network and storage). Under well-defined constrains and high-level policies, VMs can span any collection of such compute and I/O resources. Of course, there are advantages and disadvantages to each approach, but it is expected that approach III would outperform the first two approaches in terms of resources utilization and performance. All the three approaches require extensive management infrastructure to manage the enclosure and the pool of resources. The management infrastructure may include a Management-dedicated VM (M-VM) in each blade server and an enclosure-wide Global Management Platform (GMP). GMP makes decisions on resources allocation, scheduling, power management and QoS through continuous communication channels with the M-VMs.

Biography of Speaker

Mazin Yousif is a Principal Engineer in the Corporate Technology Group of Intel Corporation in Hillsboro, OR. He currently leads a team that focuses on platform provisioning & virtualization to enable platform autonomics & Capacity on Demand (CoD). Prior to that, Mazin was in the Enterprise Product Group focusing on InfiniBand and datacenter I/O interconnects. During his involvement with the InfiniBand Architecture, Mazin was chairing the InfiniBand Trade Association (IBTA) Management Working Group. Mazin finished his Masters and Ph.D. degrees from the Pennsylvania State University in 1987 and 1992, respectively. From 1993 to 1995, he was an assistant professor in the Computer Science Department at Louisiana Tech University. Mazin worked for IBM's xSeries Server Division in Research Triangle Park (RTP), NC, from 1995 - 2000. During his tenure in industry, Mazin served as adjunct professor at Duke, NCSU and currently OGI. His research interests include Computer Architecture, Clustered Architectures, workload characterization, Networking and Performance Evaluation. He has published 50+ articles in his areas of research. Mazin has chaired several conferences and workshops, was in the program committees of many others, and led several panels. Mazin is in the advisory board of the Journal of Pervasive Computing and Communications (JPCC), editor in Cluster Computing, the Journal of Networks, Software Tools and Applications and an editor in the Journal of Autonomic and Trusted Computing (JoATC). He is an IEEE senior member.