Sometimes it turns out that bigger isn’t better, especially when it comes to compute clusters. “Anything that gives more power in a smaller footprint is good,” said Jeffrey Skolnick of his most recent high-performance computing purchase: an IBM BladeCenter cluster. Skolnick’s previous HPC toy, a 4,000-processor Dell cluster that was christened in 2002 with much fanfare — and a few sound bites from CEO Michael Dell — won’t be thrown in the dumpster just yet, but will be “augmented” by the IBM system, Skolnick said.
Skolnick, director of the University at Buffalo’s Center of Excellence in Bioinformatics, could be described as a borderline cluster junkie. With a research focus on protein structure prediction, he’s been a regular customer of high-powered systems for most of his career. Before moving to UB, his lab at the Donald Danforth Plant Science Center’s Laboratory of Computational Genomics was equipped with a 1,040-processor, 335-gigaflop Beowulf cluster, and installing the 5.8-teraflop 4,000-processor system was one of the first items on his to-do list upon arriving at UB. [BioInform 06-10-02].
But after a few years, and a few “technical issues” with the Dell system, “one’s appetite increases,” Skolnick said, “and the next logical step was to expand the cluster.”
UB put out a request for proposals to a number of different vendors, and IBM presented “the most compelling package” that met the “key issues” of raw processor speed and cache size, Skolnick said. The 1.32-teraflop system contains 266 IBM eServer BladeCenter HS20 systems running Red Hat Advance Server 2.1 Linux. Each system has two 2.8 GHz Intel Xeon processors and 1.0 GB of memory. In addition, IBM is also providing a storage system comprising seven IBM xSeries 345 Intel servers connected to 5 TB of TotalStorage FAStT700 storage servers. The original Dell cluster used a 16-TB EMC storage area network.
Skolnick was unable to provide the purchase price for the system, but said that UB got “good value for the money.”
The Dell cluster will remain in place in the machine room that was built to house it, “because it’s still got about another year of effective life,” Skolnick said, but it appears that his Dell days are over. UB invited the company to bid on the upgrade, he said, but the request was met with “what I would best describe as disinterest.”
IBM, on the other hand, expressed “an extraordinary degree of interest” in working with UB. As a result, the hardware deal is concurrent with a collaborative relationship between the center and IBM Research. Isidore Rigoutsos, a pattern-recognition expert who works in IBM’s computational biology research group, will work with Skolnick’s team on functional identification of proteins. In addition, Skolnick said he’s hoping that the collaboration may provide early access to the Blue Gene protein-folding supercomputer currently in development at IBM Research.
Skolnick said that in addition to performance considerations, the blade system made practical sense because of its higher density, which packs more power into a smaller space. In addition, he pointed out, blades use less power than standard CPUs and require less air conditioning. The Dell cluster, for example, weighs in at 80,000 pounds, uses two miles of wiring, and requires 160 tons of air conditioning. “You don’t want to be in that room when the air conditioning turns off,” Skolnick warned. “It can get up to 150 degrees in a few minutes.”
It looks like UB’s machine room may not go unused for long, even if the Dell cluster goes belly-up. Skolnick is still thinking big. “My vision is to have tens of thousands of these things cranking away,” he said. After all, he noted, “if I increase the density by 50 percent, that means I can have 50 percent more computers!”