Sun, Stanford Partner for Hybrid Supercomputer Research

Published: June 22, 2006

by Timothy Prickett Morgan

Stanford University is not only where Sun Microsystems gets its name--technically, it is Stanford University Network Microsystems--but it is also where most of the executives who founded Sun came from. Stanford established a computer science degree in 1965, and is noteworthy for its new Gates Computer Science Building, which opened in 1996 and is the result of a $6 million gift from Microsoft founder, Bill Gates. While other schools and academic institutions are always looking to build a bigger supercomputer, Stanford has tended to be more conservative. But that may be changing.

This week, Sun and Stanford announced that they were collaborating on a new hybrid supercomputer architecture, one that would test the feasibility of mixing different kinds of computers into a supercomputing grid. While the Sun-Stanford project is modest, the two have big plans for the new Computational Earth and Environmental Science research lab that Sun Chairman Scott McNealy (who got his MBA from Stanford) opened this week.

Perhaps Stanford's most famous recent supercomputer is code-named Merrimac, which was a proposed design from 2003 that would use custom processors, ASICs, and interconnection fabrics to create a machine that would scale from a 2 teraflops workstation that cost $20,000 to a 1 petaflops massively parallel supercomputer that cost $20 million. This is an order of magnitude better than current designs can deliver in terms of price/performance and about three times the performance of current top-end supercomputers. This is a very exotic machine, and it is unclear when it will be operational. In the meantime, Stanford researchers need some iron to play with, and that is where Sun, one of its progeny comes in.

Sun has donated a relatively modest setup to Stanford to start: 64 nodes of Sun's two-socket V20z Opteron machines, a four-node cluster of the four-socket V40z Opteron machines, and a cluster of Sun Fire 6900 servers with 192 GB of shared memory. The Opteron machines are linked to each other through MPI protocols, while the Sun Fire 6900s run OpenMP software. The whole shebang runs Solaris 10 and is linked with TopSpin InfiniBand networking from Cisco Systems; the servers are equipped with Sun's own storage arrays and its SAMQFS supercomputer file system. This is not a lot of iron, admittedly.

The interesting bit of the new Stanford supercomputer is that this cluster of machines has been equipped with Sun's Grid Engine middleware, for harvesting clock cycles, and includes Java-based pre-processing software that Sun's consulting group created for a telecom company that allows it to examine workloads before they run and then dispatch them to the appropriate part of the cluster so they will run best. As the tests proceed, Stanford is expected to build out this machine and use it to support a wide variety of climate change, oil and gas exploration, and other related energy applications. Chevron and BP have also kicked in dough for the center, by the way.