Seminar: Some Operating System Issues in Large-scale High Performance Computing

Speaker: Hong Ong, Ph.D.

Computer Science Research Group, Oak Ridge National Laboratory, USA

Date/Time: 3rd May 2007, 13:00hrs

Location: G74, Lyle Building

Abstract:

High-end scientific applications have a diverse range of usage and programming models, and system services requirements that place a wide range of demands on their operating and runtime systems. Unix, Linux and other Unix derivatives are the most popular operating systems (OS)in use today for high-end scientific computing but with few fundamental mechanisms to support parallel systems. Although these traditional, full-featured OS can support application’s needs, their generality comes at a cost in term of scalability and performance. As we move toward petascale systems with hundreds of thousands of processors and multiple levels of parallelism, the relative cost of this overhead will increase. This talk will describe a number of investigations that are being undertaken at ORNL for addressing some of these OS issues. These efforts include exploring Linux-based single system image (SSI) for resource management; minimizing OS interferences/noise due to asynchronous system events; fast, parallel I/O storage via Lustre; and reliability and adaptability via system-level virtualization.