Fujitsu Bioserver Uses Economy Chips
By Salvatore Salamone
January 12, 2004 | Fujitsu Ltd. has developed and is now testing the BioServer, a supercomputer with a novel architecture that uses cheaper chips to deliver high-performance computing (HPC) to drug discovery researchers.
While much of the excitement in HPC today centers on the use of large systems and clusters built upon 64-bit commodity or pricey proprietary processors, BioServer uses low-performance processors normally found in printers and cell phones.
Specifically, the BioServer uses Fujitsu’s FR-V series embedded processors, which are much less powerful than the Intel Pentium or Itanium processors.
FR-V chips consume about one-fiftieth of the power of a typical PC processor. Such low power consumption translates into low heat generation, enabling many FR-V processors to be packed closely together without worry about heat-caused problems. Unlike traditional supercomputing systems, the BioServer “can sit in an office with no special cooling,” says Michael McManus, a scientist with Fujitsu America.
As many as 1,920 FR-Vs can be stuffed onto a single rack roughly 2 feet wide by 3 feet deep by 6.5 feet tall. Each CPU operates independently, running its own genomics simulation program. For this reason, Fujitsu calls the BioServer a massively parallel simulation server.
The system uses axLinux, a version of Linux for embedded processors from Axe Inc. Essentially the collection of processors operates as a Linux cluster.
Compared to a normal cluster, the BioServer takes up one-tenth the space and consumes one-tenth the power, according to Fujitsu, which results in much lower cost of ownership.
Two test systems are currently being validated. BioServer #1 is being used by Fujitsu and ZoeGene, a subsidiary of Mitsubishi Chemical that develops and licenses tools for genomics-based drug development. This system has 1,920 FR-Vs, each with 256 MB of memory, and connects to a 1.5 terabyte storage system.
“The BioServer gives us access to 2,000 times the calculating power of the desktop computers we have been using, removing a bottleneck on drug design and making the development of new compounds exceedingly more efficient,” says Masayuki Mitsuka, director of business development at ZoeGene.
ZoeGene is using the BioServer to run GROMACS, a molecular dynamics program developed at University of Groningen, Netherlands; and several informatics programs developed by Fujitsu including a molecular modeling program called CAChe, a protein modeling program called BioMedCAChe ActiveSite, a gene and protein information analysis tool called GeneSphere Enterprise Edition, and a quantum mechanics program called MOPAC.
The work with ZoeGene that combines the HPC capabilities of the BioServer with Fujitsu informatics software is illustrative of the systems approach Fujitsu is taking with its Genomics-based Drug Design Solutions. In some ways, this solutions program is comparable to IBM’s Blue Gene effort in that both are research projects that are pushing HPC hardware and informatics software development.
A second slightly smaller test system, BioServer #2 with 1,280 FR-Vs, will be used for aptamer research being done as part of Japan’s New Energy and Industrial Technology Development Organization’s (NEDO) bioinformatics development project.
Low Power Rules
Fujitsu’s efforts to shrink power and space requirements reflects an industry trend. “There has been a shifting emphasis,” says Bill Dally, professor in the computer systems laboratory at Stanford University. “There’s now more emphasis on the footprint and on the power consumption.”
Even IBM’s Blue Gene, which uses processors that are computationally powerful, was designed with low electrical power consumption, a small physical size, and heat dissipation in mind (see “Blue Gene is Cool for 2006,” July 2003 Bio-IT World, page 1).
Speaking at the SC2003 conference in Phoenix last November, Dally noted that supercomputers are composed of three main elements: arithmetic processors, memory, and bandwidth.
Arithmetic components and memory are quite inexpensive, Dally said. The same cannot be said for the interconnection technology that links processors, memory, and nodes within a supercomputer. “It’s the bandwidth that’s the cost killer today,” Dally says.
The farther apart components are within a system, the higher the connection costs. He estimates it costs about $2.50 per gigabyte to provide bandwidth across a system backplane, $5 per gigabyte across a 10-meter cable, and $25 per gigabyte across a fiber link.
For this reason, new system architectures are being explored -- architectures that minimize the distance between elements to reduce the bandwidth costs. “If you keep power down to keep links short, this makes bandwidth less expensive,” Dally says.