SPEC, Google pressure component makers on power

[Daniel ~]

Written by Daniel
Thursday, 20 December 2007

SPEC, Google pressure component makers on power

By Jon Stokes | Published: December 20, 2007 - 06:44AM CT
ArsTechnica

In May of last year, SPEC began soliciting feedback from the industry on a forthcoming slate of power efficiency benchmarks. The first of those benchmarks, SPECpower_ssj2008, is now out, but a recent report from Google indicates that the big villains in datacenter power consumption are now memory and networking components that don't provide much in the way of power optimization..... Front Page

[cloasters]

I guess hard drives by the dozen really add to the power bill. More efficient memory may not be too far away.

[Gizmo]

I'm no expert, but I'm unsure how much gain can be achieved with memory technology.

The DRAM memory cell is just a transistor and a capacitor, so there's nothing that can be done there. The output buffers and sense amplifiers consume a substantial amount of power, but the vast majority of that is while driving the signaling lines (i.e. when active, as in most of the time the machine is on), so there's not much that can be gained there either. There might be the potential for a 10 or 15 percent power savings there somewhere, but that's about it, I'd think.

Hard drives are in a somewhat similar position; the vast majority of the power consumption comes from the spindle motor and head actuators (i.e. mechanical stuff). Powering down the spindle saves bundles of power but also introduces huge latencies. They MIGHT be able to mitigate that by incorporating rather substantial caches (like on the order of 256 MB or so), but cache is expensive compared to disk storage. In addition, every bit of cache RAM you add eats into the power savings from spinning down the drive. At some point, you are no longer gaining anything. However, another thought that crosses my mind is that they might be able to reduce the spindle motor speed to some lower level, rather than stopping altogether; e.g. a drive with a 15K RPM spindle might have a set of lower power modes that run at 10K and 5K RPM. This would increase latency as compared to full speed, but could represent a substantial savings in power as compared with full speed operation, while mitigating the latency issues of spinning the drive up from a dead stop.

[cloasters]

Sounds good to me, perhaps you could "stagger" balance banks of drives so that the lag time wouldn't be objectionable? Merry Christmas, and thank you for teaching me a LOT about the world of IT!