High Performance Technologies Inc, a professional services firm from Reston, Virginia, has won a contract from the US Department of Commerce to supply a large cluster of Alpha-based Linux servers to the National Oceanic and Atmospheric Administration’s Forecast System Laboratory. The system, to be installed at the Boulder, Colorado, based lab, will be used to […]
High Performance Technologies Inc, a professional services firm from Reston, Virginia, has won a contract from the US Department of Commerce to supply a large cluster of Alpha-based Linux servers to the National Oceanic and Atmospheric Administration’s Forecast System Laboratory. The system, to be installed at the Boulder, Colorado, based lab, will be used to improve existing weather forecast models and develop new ones. HPTi has partnered with Compaq Computer Corp, the University of Virginia, processor interconnect company Myricom Inc and storage systems company Patuxent Technology Partners LLC on the project.
The contract was tendered at a fixed price and won on performance merits, according to David Rhodes, director of high performance systems integration at HPCi. The list of other companies bidding for the deal wasn’t revealed, but are thought to have included Silicon Graphics Inc, IBM Corp, Compaq Computer Corp (separately, with its Wildfire technology), Sun Microsystems Inc and Hewlett- Packard Co. The traditional companies are now starting to see competition in the supercomputer sector from lower-cost alternatives using the Linux operating system and simple server nodes.
Although the size of the contract is valued at only $15m, HPTi claims the system will grow into one of the fastest supercomputers in the world. At initial installation, the system will run at one third of a TeraFlop, growing to four TeraFlops by the final upgrade in 2002 – 20 times faster than the current FSL system. There are three phases: 256 processors initially, a further 256 by month 13, and a total of 1,024 Alpha 21264 processors by month 34. Nodes will be single processor in the earlier stages, with dual processor nodes introduced towards the end.
Arcadia, California-based Myricom, formed in 1994 and privately held, says its Myrinet high performance processor interconnect can scale up to thousands of processors, and says it will continue to improve performance and lower cost over the next year and over the life of the FSL contract. Earlier this year, the company was demonstrating a cluster of five Compaq XP1000 Monet workstations at LinuxWorld. Its interconnect technology is also in use at the University of Illinois National Center for Supercomputing Applications to connect up a 128-node 256- processor cluster of PCs from Hewlett-Packard Co and Compaq. The company has approximately 35 employees.
Patuxent, formed by two veterans from Cray Research Inc and one from Data General Corp in mid 1997, has been working as a consultant on RAID storage systems, and claims to be a pioneer of storage area networks and robotic mass storage. It has offices in Maryland, Texas, Oregon and Virginia. The University of Virginia was the site of HPTi’s benchmarking activities, and hosted the live test demonstration to NOAA and FSL. The University has been working with HPTi on the application of advanced clustering technology.
While based on the Beowulf clustering technologies included within Linux, Rhodes says the HPTi machine is not related to hobbyist efforts to link multiple low-cost servers together, but is a real system, with all the fault-tolerance, batch queuing and other features you’d expect. The HPTi approach, however, differs from the drift towards distributed symmetrical multiprocessing architectures that most supercomputer companies, and even IBM with the SP2, are now moving to. Many existing users with MPI message passing interface applications find themselves faced with a major upgrade task when moving from their existing simple architecture systems. Also, says Rhodes, the bandwidth of such systems can’t keep pace when adding high numbers of processors. We’re keeping the architecture simple, says Rhodes. Last May, Compaq began installing a comparable Linux-based cluster of Alpha servers for the Cplant project at Sandia National Laboratories, now up to 812 processors.
This is HPCi’s first foray into systems integration, says Rhodes, and there will be more to come. The company’s typical projects have combined parallel servers with high performance networking, multimedia databases, online analytical processing, visualization and systems engineering. Formed in 1992, HPCi has grown from four to over 180 employees in that time.