7.2.4 Bandwidth Greedy and Frugal

		10 Conclusions and Views

		This book has explored a new computational opportunity that delivers the capabilities of earlier supercomputers into the hands of small laboratories, groups, or even individuals. Beowulf-class computing systems have enabled a revolution by breaking the price/performance barrier and by putting control of system configuration and operation directly in the hands of the end-user. The possibilities implied by Beowulf, some of which have largely gone unexplored, are breathtaking, although several factors conspire to impede broader usage of Beowulf-class systems in the work place. In the concluding comments to this book, these issues are considered as they relate to the practical capabilities and limitations of Beowulf-class systems and their future evolution.

		10.1 New Generation Beowulfs

		Changes are occurring almost daily in the core technologies used to implement Beowulf-class systems. The earliest Beowulfs could sustain less than 5 Mflops per node on favorable (but real) applications with 10 Mbps communication links. As this book goes to print, the 500 MHz Intel Pentium III processor and the 533 MHz DEC Alpha processor are being incorporated into Beowulf systems. The AMD K-7 and Intel's Merced chip are on the horizon, promising performance between two and four times the last generation of systems. Although costs have yet to drop sufficiently, both Myrinet and Gigabit Ethernet communications technologies (aided by the new PCI-2 bus which is wider and runs at a higher clock rate) will provide a near-term path to 1 Gbps communication channels. Several companies, including Cisco, 3Com, Extreme and Lucent offer Ethernet switches supporting a mix of Fast and Gigabit links and backplanes with many Gbps of internal bandwidth, all at commodity prices. All these trends broaden the domain of applications and algorithms that may be effectively performed on Beowulf-class systems and will permit larger configurations to be implemented with off-the-shelf components.

		By the end of 2001, Gflops performance, GByte memory capacity, and Gbps network bandwidth Beowulf nodes will be commonplace. Processor nodes of 2 Gflops peak performance will be achieved shortly thereafter by employing the newest processors in symmetric multiprocessor node configurations. Peak performance to cost will approach $1/Mflops, enabling Teraflops peak performance Beowulfs at a cost of about a million dollars by the year 2002 or earlier. That is well within the budget range of many moderate sized computing centers. In 1999, machines of that capacity qualify as the fastest in the world.