Google: Server efficiency needs new recipe

BURLINGAME, Calif.--Chipmakers have been applying lessons learned in mobile computing to servers in an effort to increase efficiency by lowering power consumption. But a noted Google engineer threw some cold water on the approach on Monday, arguing the two styles of computing are too different.

"The data center is a different device than the key targets for mobile electronics, laptops, and mobile devices," said Luis Barroso, a Google engineer who closely studies the company's power consumption, speaking at the O'Reilly Velocity conference here.

And naturally, with at least hundreds of thousands of servers in operation and its data centers placed near power plants to cut electricity costs, Google is trying to get computing equipment makers more excited about efficiency.

"Maybe if you call this a land-held computer, perhaps they'll help us," he quipped, showing an aerial view of a sprawling Google data center.

The basic problem is that mobile devices and servers have different modes of activity.

Mobile devices have been improving through better exploitation of the fact that they spend a lot of time dormant with occasional bursts of activity. That lets processors and other electronics save power by spending most time in low-power sleep modes, then snapping awake for peak-power high-performance modes when necessary.

Google's servers, though, have the opposite type of activity: they spend most of their time doing modest amounts of work, with frenzied moments of peak activity and complete lulls a rarity, Barroso said. The measurements are based on measurements of about 5,000 servers performing four different Google applications, he added.

The company's servers simply can't go to sleep, he said. Each machine is "rarely fully idle," he said. "The fraction of time the servers are actually doing exactly nothing is very small."

Thus, Google is urging electronics designers to create products that more gracefully reduce power demands as activity diminishes. Servers naturally consume peak power at peak activity, but what's bad is that they still consume about half peak power when at zero activity.

Processors have gotten a bad rap for squandering ever more energy--indeed, Barroso himself, once a chip designer for Digital Equipment, has expressed such concerns. But chips actually are better than hard drives, memory, and network adapters at reducing power consumption during periods of moderate activity.

Some sophisticated hard drives, for example, can slow down their rotational speed to save power during periods of lower activity. However, "They need to bump to higher RPM to do something useful," to read and write, he said, unlike processors, which can actually still process data when in low-activity modes.

[Marcus Westrup]

Which of course, misses the point.
Two very big power users are not directly part of the computer core. Power supplies are rarely efficient (60 to 80% depending on load). The need to air condition an entire room - when you only need to cool a few square inches of electronics - wastes much more.

Start with redesigning the computer center. An integrated power system for all the machines would be more efficient and reduce the heat load. Next, vent the CPU / Hard Drive heat directly outside. These devices run warm anyway, and would be fine with ordinary outside air for cooling (as long as they get enough air flow).

[greeny73]

Another step from Google to be more green - if only other big companies would start following their lead.
check out www.cutyourfootprint.com to find a good profile of how green Google (and other big companies) actually are...

[carlhage]

Mobile computing and servers aren't that different. Both run >95% idle on the average (when not in sleep mode) and waste most of the power. Laptops have a much better transactions/watt rating (and lower total cost) because of better care in power management. Servers have multi-cores, but don't power down the cores when not needed. A better solution is a multi-core processor, with a special low power chip that can run low duty background processing and take near 0 power when idle. It should be possible to make a system where energy consumption is near 0 at idle and scales linearly. The problem is we have no standardized benchmarks and manufacturers don't disclose power usage, so have no incentive to make them cheaper to use.