Keynote Speaker: Thomas Sterling

Dr. Thomas Sterling is a Principal Scientist at the NASA Jet Propulsion Laboratory and a Faculty Associate at the California Institute of Technology. The Principal Investigator on the HTMT architecture project, he leads a team of over a dozen institutions in an in-depth exploration of an innovative approach to achieving Petaflops-scale computing by integrating advanced technologies in a dynamic adaptive latency management structure. Sterling is also engaged in the development of advanced Processor-in-Memory execution models, and the invention of Continuum Computer Architecture, a general purpose cellular architecture approach to ultra-scale computing. A graduate of MIT (Ph.D, 1984) and Hertz Fellow, Dr. Sterling has engaged in research studies in parallel computing systems architecture and software over the last two decades. He is best known for his development (with colleague Don Becker) of Beowulf-class PC cluster computing when at the NASA Goddard Space Flight Center and as a leader of the National Petaflops Initiative while at Caltech and JPL. He is the co-author of the books: "Enabling Technologies for Petaflops Computing" and "How to Build a Beowulf" and holds six patents.

PC Clusters - The Convergence of High Performance Systems Architecture
Dr. Thomas Sterling
Center for Advanced Computing Research
California Institute of Technology
and
High Performance Computing Group
NASA Jet Propulsion Laboratory

The last year and a half has firmly established cluster systems as the leading approach to scalable high performance computing and PC clusters such as Beowulf-class systems as a major form of clusters. Rapid advances in PC-based SMP nodes, system area networking, and resource management software during this period have transformed a once limited domain of parallel processing into a robust, versatile, and broadly applicable methodology demonstrating unprecedented price-performance advantage with respect to classical vector and MPP type systems. With order of magnitude superiority in cost, Beowulf-class systems and other types of PC clusters have opened up new application domains to low cost parallel computing in science, engineering, industry, commerce, business, and entertainment as well as playing an important role in the explosion of internet and web applications and services. Yet, as important as cost is, so too is the implied convergence to a single general system architecture and programming methodology, supported by many vendors of both hardware and software and towards which education of future programmers can be targeted. The question surrounding The last year and a half has firmly established cluster systems as the leading approach to scalable high performance computing and PC clusters such as Beowulf-class systems as a major form of clusters. Rapid advances in PC-based SMP nodes, system area networking, and resource management software during this period have transformed a once limited domain of parallel processing into a robust, versatile, and broadly applicable methodology demonstrating unprecedented price-performance advantage with respect to classical vector and MPP type systems. With order of magnitude superiority in cost, Beowulf-class systems and other types of PC clusters have opened up new application domains to low cost parallel computing in science, engineering, industry, commerce, business, and entertainment as well as playing an important role in the explosion of internet and web applications and services. Yet, as important as cost is, so too is the implied convergence to a single general system architecture and programming methodology, supported by many vendors of both hardware and software and towards which education of future programmers can be targeted. The question surrounding Beowulf-class systems is: can they ever approach the capability of the MPP systems they have begun to replace. This talk will examine the gap between emergent Beowulf-class systems and MPPs to provide a realistic measure of the discrepancy of the two systems. Data from both simple test codes and realistic benchmark kernels run on both classes of system will expose the similarities and discrepancies between the two. It will be shown that there are three general categories of applications: those for which both MPPs and Beowulfs perform as well, those for which MPPs have been shown to be superior in performance, and those for which both classes suffer due to poor memory handling and latency hiding characteristics. Anticipated advances in technology for Beowulfs will be considered in terms of their expected impact in closing the gap of the second category.