Combining Parallelism, Virtualization, Heterogeneity and Reliability: Some cu...

Google Tech Talks October 28, 2008 ABSTRACT This talk will begin with an overview of the Computer Systems group within the College of Engineering and IT at The Australian National University. These fall under the Themes of Bio-Engineering, Robotics, Advanced Runtime Systems, Performance Analysis, Parallel Processing, Operating Systems. Depending on audience interest, projects under the latter three themes will be discussed in detail. These include: OpenMP for Contemporary Clusters: state-of-the-art for distributed shared memory based systems, the handling of heterogeneity and utilization of advanced networking technologies (Infinband). High Performance Numerical Computing on Service-Oriented Architectures: this work involves the extension of the Symphony programming paradigm (Platform Computing), originally developed for financial applications running on enterprise grids. The desirable properties of this model include inherent load balancing in a heterogeneous environment, fault tolerance and relative simplicity of programming. The challenge arises in enabling compute tasks to effectively communicate with low overheads, while retaining most of these advantages. Virtualized HPC Clusters: Virtualization has many advantages in the context of a data center with a heterogeneous cluster of sub-clusters. Work on evaluating the performance of virtualized communication configurations is described, together with a framework for scheduling for taking advantage of virtual machine migration, based on the Xen hypervisor. Simulation and Performance Evaluation Frameworks for NUMA Clusters: We discuss a standard and detailed design for multiprocessor computer simulation. Both can be parallelized and a multiprocessor host, with the performance being limited by the inherent imbalance in the work required to simulate each CPU for a fixed interval of simulated time. We will also describe why issues in validation are particularly problematic. Recent work includes incorporating these techniques into dynamic binary translation frameworks (Valgrind) and extending them to cluster computers. Multicore Computing: With the recent donation of an UltraSPARC T2 by Sun Microsystems, we will discuss current work and future plans to expand multicore computing into our teaching and research programs. Speaker: Dr Peter Strazdins, Australian National University Peter is a Senior Lecturer in the Computer Systems Group of the Department of Computer Science at the Australian National University. He graduated with a PhD from the ANU in 1990 in the area of programmable systolic arrays, and from that time until 2002 was involved in the ANU-Fujitsu CAP Program in Parallel Computing. He has worked in the areas of parallel linear algebra, multiprocessor computer simulation, performance evaluation and developing middleware and virtualization techniques for high performance computing.