Until the mid 1990's, supercomputing, high performance computing, or high end computing was reserved to an elite few who had access to systems costing millions (even tens of millions) of dollars. Large vector supercomputers and massively parallel processing systems (MPP) could deliver in the range of 1 to 50 Gflops.

But because these systems were usually shared among a community of users, the actual capability provided was often a small part of the peak capability of the system. For example, the majority of users of a C-90 capable of 16 Gflops peak performance would run on a single "head" of only 1 Gflops peak performance while other users space shared the other processors of the system at the same time.

Many large science and engineering applications were either not run or were greatly simplified in fidelity and size to run for shorter time on smaller parts of such big machines, or worse yet, on scientific workstations. But many other potential applications were never executed at all for want of adequate computing resources due to over subscription of existing shared supercomputing centers.

With the advent of local area networks of desktop scientific workstations, some environments tentatively explored and employed "cycle harvesting" applying unused workstations at off-hours to do embarrassingly parallel tasks; usually running the same program on many different machines at the same time with different input data sets. Workstation clusters were explored by the University of Wisconsin, UC Berkeley, and other sites demonstrating the possibilities of clustered computing. And research was conducted in the development of new network technology that might be used to integrate workstations.

But problems of cost, customized hardware, and proprietary software limited their installed base and use. At this time, a new class of clustered computing system was devised by a small group at the NASA Goddard Space Flight Center that overcame these difficulties and explored what has become the most rapidly growing type of parallel processing.

At the beginning of 1994, the Beowulf project undertook to assemble a cluster of PCs and to evaluate their utility as a scalable system for scientific computation using only mass market commodity off-the-shelf hardware and widely available open source software. It was only at that time that the capability of such hardware was good enough and the cost low enough to potentially enable this new class of computing: PC clusters.

The Beowulf project adopted the inchoate Linux operating system because it was free and open software, avoiding the legal complications of using Unix and providing the software for making necessary changes. The Beowulf project developed the majority of Ethernet drivers included today as well as channel bonding to support multiple simultaneous and many other low level tools for managing clusters of PCs.

Equally important was the pathfinding work in applying these systems to real world scientific applications. It was found that in some cases Beowulf could equal in performance that of much more costly machines and in many cases they provided a price-performance advantage of an order-of-magnitude or more.

Through a series of successive generations of systems, these capabilities grew with more than a Gflops sustained operation being achieved on a system costing less than $50K in 1996 and 10 Gflops on 120 processors in 1997, also winning the Gordon Bell prize for price-performance two years in a row. Today, a number of commercial concerns provide low cost Beowulf-class systems and other PC clusters with an installed base in industry, government labs, and academia. A number of these systems are now included on the Top 500 list of the world's most powerful computer systems.

This presentation will discuss the motivation and importance of Beowulf-class computing, its hardware and software elements, and its history from inception of 16 processor systems to present day systems up to a thousand processors.

Biographical Sketch:

Dr. Thomas Sterling received his Ph.D. as a Hertz Fellow from MIT in 1984 and has held research scientist positions with the Harris Corporation's Advanced Technology Department, the IDA Supercomputing Research Center, and the USRA Center of Excellence in Space Data and Information Sciences. In 1996, Dr.Sterling began a joint appointment with the NASA Jet Propulsion Laboratory and the California Institute of Technology. He is a Principle Scientist in the Jet Propulsion Laboratory's High Performance Computing group, and he is a Faculty Associate at the California Institute of Technology's Center for Advanced Computing Research.

For the last 20 years, Dr. Sterling has engaged in applied research in parallel processing hardware and software systems for high performance computing. He was a developer of the Concert shared memory multiprocessor, the YARC static dataflow computer, and the Associative Template Dataflow computer concept, and he has conducted extensive studies of distributed shared memory cache coherence systems.

In 1994, Dr. Sterling led the team at the NASA Goddard Space Flight Center that developed the first Beowulf-class PC clusters including the Ethernet networking software for the Linux operating system. In 1999, he co-authored the MIT Press book "How to Build a Beowulf".

Since 1994, Dr. Sterling has been a leader in the national Petaflops initiative, chairing two workshops on Petaflops systems development and chairing the subgroup on the Petaflops computing implementation plan for the President's Information Technology Advisory Committee. He chaired both the first and second Conferences on Enabling Petaflops Computing in 1994 and 1999. He is also an author of the book, "Enabling Technologies for Petaflops Computing" published by MIT Press in 1995.

Dr. Sterling is the Principal Investigator for the interdisciplinary Hybrid Technology Multithreaded (HTMT) architecture research project sponsored by NASA, NSA, NSF, and DARPA involving a collaboration of more than a dozen cooperating research institutions. The HTMT project is developing an adaptive, latency tolerant, Petaflops-scale computer employing superconductor, optical, and processor-in-memory technologies.

Dr. Sterling holds six patents, and was the winner of 1997 Gordon Bell Prize for Price Performance.

For further information:

How To Build A Beowulf , Sterling, Salmon, et. al.
Beolinks (Caltech)
The Beowulf Project (CACR-Caltech)
Beowulf.org
The Beowulf Underground
The Legend of Beowulf (Anglo-Saxon epic poem)

A streaming video recording of this presentation will be available after the lecture through Dr. Dobb's TechNetCast (http://www.technetcast.com).
Check site for schedule and details.