Mac cluster rises in supercomputing ranks

According to updated performance figures, an inexpensive cluster built from Macintosh G5 PCs could rank as the third most powerful supercomputer in the world

A supercomputer formed of a cluster of Macintosh G5 PCs has inched higher in the list of the world's fastest machines, ahead of all but two rivals -- and its performance could still improve, according to the system's architect.

According to the latest performance figures from Virginia Tech's Terascale Cluster, nicknamed the Big Mac, the system is computing at 9.55 trillion operations per second, or teraflops. That puts it behind only Japan's Earth Simulator, at 35.8 teraflops, and the ASCI Q supercomputer, at 13.8 teraflops. The figures were posted in a report by Jack Dongarra, a University of Tennessee computer scientist who maintains the top 500 list.

The Big Mac is turning heads in the high-performance computing world because it has been built for just over $5m (£3.2m), from off-the-shelf Macintosh PCs, in about six months. Top-ranked supercomputers traditionally cost hundreds of millions of pounds and can take years to construct.

The Big Mac's performance figures have fluctuated as its performance is tuned. Early tests using only a few of the cluster's 2,200 processors showed a higher estimated performance, while more recently the system reached 8.1 teraflops, which would have placed it at No. 4 in the supercomputing list. Virginia Tech's goal was to build one of the world's five fastest supercomputers.

The system's architect, Srinidhi Varadarajan, said at the O'Reilly Mac OS X Conference in California on Tuesday that further tuning could boost performance by another 10 percent, according to reports. The final supercomputing list will be revealed in mid-November at the International Supercomputer Conference.

The cluster was assembed from 1,100 dual-CPU Macs using IBM's 64-bit G5 processor. Varadarajan said the university paid the full $3,000 price tag for the PCs, and said he had at first considered using processors from Intel or AMD. Intel's Itanium 2 was too slow, while AMD's Opteron was too expensive, Varadarajan said. He said he is planning to upgrade the cluster to Apple's new Panther version of the Mac OS X operating system shortly.

The servers in the cluster are connected through 24 high-speed Infiniband switches from Mellanox Technologies. Infiniband, which was developed by a consortium of server and storage companies, provides greater bandwidth than other interconnect technologies on the market, such as Miranet, and can often cost less. The cluster also uses a cooling system from Liebert, a division of Emerson Network Power, as well as Gigabit Ethernet switches from Cisco Systems.

Virginia Tech will use the cluster to perform research on nanoscale electronics, chemistry, aerodynamics, molecular statics, computational acoustics and molecular modelling, among other tasks.

Clustering, which involves linking hundreds or thousands of computers to take on massive projects, has opened the supercomputing market up to companies other than those like IBM and Cray that have established backgrounds in big iron. Dell has emerged as one of the leaders in selling clusters to research institutions such as Cornell University. Utah's Linux Networx, meanwhile, has won contracts to install systems at Los Alamos and other national research laboratories.

Japan's Earth Simulator, with 5,120 custom processors, was measured at 35.8 teraflops last year, and is estimated to have cost up to $250m. ASCI Q, a Hewlett-Packard machine running on 8,192 Alpha chips, is ranked No. 2 at 13.8 teraflops. The third-ranked system on last year's list is, like the Big Mac, a cluster: it was built by Linux Networx for Lawrence Livermore National Laboratory from 2,304 2.4GHz Xeon chips, and runs at 7.6 teraflops.

This cluster would drop behind the Big Mac and another HP-built machine powered by Intel's Itanium 2 processors, which is set to enter this year's list at 8.6 teraflops, according to Dongarra's figures.

CNET's Ina Fried and Michael Kanellos contributed to this report.