Virginia Polytechnic Institute and State University

Cost-Conscious Supercomputing

Optimization

Pulling Out Incredible Efficiencies

Kazushige Goto, a visiting scientist at the University of Texas, Austin, added still more customization to the supercomputer. Goto optimized the BLAS (Basic Linear Algebra Subprograms) libraries, a function that Lockhart says was instrumental in achieving the supercomputer’s 10.28 teraflop processing speed benchmark.

“By going with the Power Mac G5 computers, we were able to hit our budget and meet all of our performance goals.”

— Jason Lockhart

Notes Lockhart, “These libraries have to be optimized by hand for each processor architecture at the assembly-code level. It’s usually a very long, tedious process, trying to eke out every gigaflop possible. But Mr. Goto got some early access to the Power Mac G5 systems, and he was able to pull some incredible efficiencies out of them — in the mid-90 percent range — in only five or six weeks. The theoretical limit of a dual-processor Power Mac G5 is 16 gigaflops, or 8 gigaflops per processor. He was getting 95 percent efficiency on a single processor, which was amazing for the short amount of time he had.”

Dr. Dhabaleswar Panda of Ohio State University supplied the interface library (called MVAPICH) that enables the InfiniBand support. Since the library was developed under a Linux variant, Varadarajan had the straightforward task of bringing it over to Mac OS X, which is a variant of BSD. “Srinidhi did the porting work himself, with support from Dr. Panda’s grad students,” says Lockhart. “They were very instrumental in helping us pull the library into Mac OS X, so we could create our cluster of machines. Without it, we’d have been hard-pressed to get anywhere with the project.”

Since Apple provides all of its processor optimization tools for free, Virginia Tech was able to quickly access the powerful optimization capabilities built into the PowerPC G5. Collectively named CHUD (Computer Hardware Understanding and Development), they include profilers, processor simulators, and visualization tools. All are included as part of Apple’s free Xcode development environment.

Introduction 1 2 3 4 5 6 7 8 9