Apple has carved out a niche on the desktop, and has a loyal following in graphics design, video editing, and other similar technical fields. More recently, it has stormed the world with the iPod. But few think of the company as a player in the supercomputing field. Yet that is exactly where its Xserve G5 line of servers is making waves.
When one thinks of Apple, one thinks first of the desktop market. Its Xserve G5 server line is changing this image. The Hypersonic Missile Technology team at COLSA is clustering G5 systems to design scramjet engines and missiles more economically.
Case in point: the Hypersonic Missile Technology (HMT) team at Huntsville, Alabama based COLSA Corp. has adopted a supercomputing platform consisting of 1,566 of the 64-bit, dual-processor Apple Xserve G5 servers for its MACH5 (Multiple Advanced Computers for Hypersonics) cluster. This gives the company 60 times more computational power than its current production machine. Result: It can now do calculations in one day that previously took two months.
He says the primary issue that led to the Apple acquisition was the need for speed. The grid size of these CFD problems is on the order of one to 10 million cells that require 12 to 16 equations to be solved in each cell.
“Flight tests are extremely costly, so this technology enables us to pick up the development and design pace via more iterations,” said Dr. Anthony DiRienzo, executive vice president of Colsa Corporation. “We will no longer have design engineers waiting two months for an answer.”
Colsa models the complex aero-thermodynamics of hypersonic flight for the Research, Development, and Engineering Command (RDECOM) of the U.S. Army at nearby Redstone Arsenal. It has been using an IBM RS/6000 with 284 Power3 Winterhawk 64-bit processors as its primary production machine for Computational Fluid Dynamics (CFD) processing. Additionally, it has developed eight other large clusters using a wide range of processors, including designs based on the Intel Xeon, AMD Opteron, and various IBM processors. The largest of these is a Xeon/AMD 512-processor machine. According to DiRienzo, most of these machines run on Linux.
He says the primary issue that led to the Apple acquisition was the need for speed. The grid size of these CFD problems is on the order of one to 10 million cells that require 12 to 16 equations to be solved in each cell. Runs could take one to two months on the IBM RS/6000.
“The design of hypersonic vehicles requires a faster turn around time,” said DiRienzo. “The Mach5 should be able to run comparably sized problems overnight.”
Earlier, Colsa had experimented with one of the first Apple clusters using 17 Power Mac G4 systems when clusters weren’t yet fashionable. When it decided to add supercomputing power to its operations, it tested everything out there including AMD Athlon, Intel Xeon, AMD Opteron, Intel Itanium2, and PowerPC G5. Colsa benchmarked a total of eight systems using its primary production application. This test involved simplified geometry with 2 million grid points and aero-thermodynamics of 12 chemical species in the atmosphere and engine combustion products. One cluster had a reliable processor but high power and heat requirements. Another had a capable but expensive processor that was four or five times more expensive than the G5.
Once Apple servers were selected, the physical installation took about eight days. The organization is now benchmarking the complete Apple supercluster in its production environment.
“We are performing both Linpack benchmarks as well as benchmarks specific to the particular CFD code that is used by our government customer,” said DiRienzo. “We expect to be in full production in February 2005.”
Based on the results achieved so far, he is happy with its performance. At its peak, the supercluster can exceed 25 teraflops. With this kind of power, the company expects to design scramjet engines and missiles more economically, with significant reduction in wind tunnel testing and in the number of very expensive flight tests.
In addition, DiRienzo notes several positive things about the Apple Xserve G5. The 1U server’s engineering, he says, makes a big difference in heat and energy management. Colsa has 3132 processors in 600 square feet of space and uses only 110 tons of air conditioning to cool them.
“Our experience with other 1U servers has been less than satisfactory in that we have had to modify some to dissipate the heat,” said DiRienzo. “This decent engineering allows the Apple server to run at maximum power when needed whereas other competitors do not.”
He further states that these servers are engineered to require less power than those of its competitors. According to Colsa’s figures, this can amount to as much as 50 percent less power than some of its rivals — a big difference in the monthly electric bill to its facility.
Not everything has been smooth sailing during the testing phase, though. According to Colsa, the main deployment snag concerns the interaction of Linpack tests with the MAC OS X 10.3 Panther operating system. When engineers run regular CFD code no problems are experienced. With Panther, however, some incompatibilities can crop up. As a result, Apple developed a firmware fix to the problem.
“We will be analyzing the goodness of the fix in the next month,” said DiRienzo. “Once we complete our CFD code benchmarking, we will shift our attention back to the Linpack benchmarks.”
His advice to others interested in procuring a supercomputing system is to let the market give a best value product. Colsa solicited six companies for bids on behalf of its government customer. The primary concern was performance, and the company took care not to weigh the solicitation toward any particular vendor.
“The result is a product that is significantly more cost effective than any machine in the top 20 of the Top500 supercomputing list,” said DiRienzo. “The Xserve G5 has a smaller footprint, uses less power, and it gives us the performance that we were looking for.”