Future Clusters
Penn State’s insatiable demand for research compute power continues unabated. Forty 4-way Sun V40z servers are en route, and it is talking about adding another 40 more in the near future. These boxes will also have 2.6 GHz AMD processors. For the past few months, Agarwala has been testing one such unit in-house. According to his benchmarks, he feels it scales well with regard to problems.
“With many of the problems we run, the V40z will speed up the end-to-end time to execution,” he says. “However, some computations don’t lend themselves well to multiple processors.”
These 4-processor machines will have 32 GB of memory. Their introduction is expected to multiply the options available to Penn State. Some computations take a very long time to process because they require an excessive amount of I/O from disk. By moving I/O to memory, researchers can solve problems that were formerly beyond them. Alternatively, specific situations might demand three processors in the 4-way Sun server sit idle. One CPU using all 32 GB of RAM might be the most efficient configuration for certain problems.
Agarwala is also leaning in the direction of dual-core technology. After the arrival of the 40 single-core Sun V40z’s, he believes it extremely likely the university will order dual-core V40z’s next. These are 4-way dual-core (eight processor) 2.2 GHz machines.
“We expect to have at least 100 V40z servers eventually so it is natural that many of these will end up being dual-core Sun machines,” says Agarwala. “I believe that dual-core technology is one of the most significant advances of the past five years.”
Agarwala’s imagination is already full of the possibilities. When attempting to solve research computations where 32 GB of memory is enough, he can throw eight chips at it instead of four, giving it a performance boost of 50 percent to 70 percent at the same cost.
“Sun’s new dual-core Opteron systems allow the university to almost double its processing core capacity while maintaining the same heat and real-estate footprint,” Agarwala said. “If I can get even 50 percent more processing performance, while using the same amount of power, then this will be huge.”