Home > Hardware

Computerworld - Discussions of computer performance are typically dominated by references to measures such as MIPS, MHz and MFLOPS.
But Wu-Chun Feng, a computer architect at Los Alamos National Laboratory in New Mexico, expects that to change during this decade. He says it's time for the computer community to adopt alternate metrics for evaluating performance. "It's about more than speed; it's about reliability, availability and efficiency," he says.
It's more than an esoteric semantic point. ASC Q, a giant supercomputer at Los Alamos, has 8,192 processors, and although each one is extremely reliable (as well as fast), there are so many of them that the machine overall fails about 114 times a month, or once per eight-hour shift.
The problem is heat, Feng says, and it's not just an issue in supercomputers. The power consumed and the heat given off per unit area in processor chips increases with Moore's Law, doubling every 18 to 24 months. Indeed, the power density of commodity processor chips used in PCs today is on a par with that inside a nuclear reactor. And the failure rate of a processor doubles with every increase in temperature of 10 degrees Celsius (18 degrees Fahrenheit), Feng says.
Running Cooler
But researchers are inventing clever ways to keep reliability up by keeping heat and power consumption down.
In 2002, Los Alamos built a 240-node computer called Green Destiny. For two years, it ran without a single failure in a dusty, unventilated warehouse where the average temperature was 85 F. The magic: It used processors from Transmeta Corp. that consumed just 6 watts each. In comparison, mainstream microprocessors at the time consumed about 100 watts.

Because of heat, the 8,192-CPU supercomputer at Los Alamos fails about 114 times a month. Image Credit: Los Alamos National Laboratory

Green Destiny's low-power, low-heat, high-reliability characteristics came from its architecture, independent of the applications running on it. But now Feng and his colleagues are developing software that can alter the processor's power consumption depending on the moment-to-moment needs of the application.
The power drawn by a processor is proportional to its frequency and voltage. Notebook computers today can scale back frequency and voltage after some period of user inactivity in order to conserve battery power. But that's not feasible in a number-crunching scientific computer or a big transaction-processing server. "The CPU almost always looks busy compared to doing a Microsoft Office document," Feng explains.
So the lab has developed and is now enhancing "dynamic scaling" software that learns the characteristics of the application as it runs. It's