Power-sucking vampires
More is not always better
My main work machine, a souped-up blue and white Macintosh, is of a
product generation that doesn't even support PCI deep sleep as newer
Mac desktop machines do, much less turn off the fan in the power
supply when it's purportedly sleeping. So there's not a whole lot I
can do there until it's time to replace this box. I also have an
OpenBSD box, a Micron PII-450, which ran 24/7 and did an excellent
jet-engine simulation under my desk. Once the electrons became
scarce, that box was turned off, and is now on only when necessary.
But for my work, having a Unix box up and running full-time is a very
useful thing, and having to fire it up and bring it down gets to be
quite a pain very quickly. So, I started looking into Unix server
appliances.
Cobalt's current-generation Qube3 server is powered by an AMD K6-2 processor and, according to Cobalt, the whole system consumes up to 45W at peak load. The Qube2, its predecessor, uses a little-endian MIPS processor, and only consumes about 25W max. Cobalt appears to be going in the wrong direction with its power consumption.
This is rather odd considering that the Qube and the Raq product families share hardware generations, and the Raq is designed as a 1U, 19" rack-mountable device. If I were a co-location (colo) facility, I'd be interested in devices that consumed less power because they also generate less heat, which therefore requires less cooling. I wouldn't be at all surprised if colo companies started charging premiums for client hardware that was excessively power-hungry.
In the house
But far more than just computers soak up power in most contemporary
households. How many devices are left on permanently without regard
for their power draw? I used to leave my cable TV settop box running
all the time until I realized that it generated a substantial amount
of heat, and it made that heat from the stuff that came in through
the power cord.
As homes get even more wired, more and more devices will be connected to the mains as well as various networks: Ethernet for Internet connectivity, FireWire/IEEE1394 for media devices, wireless for untethered devices (except when they're resting/syncing in their cradles), etc. All these things will soak up electricity; now is the time to start communicating to the manufacturers of these devices that power consumption matters to their customers.
(Those that think homes of the future won't have at least one or two servers in them are in for quite a surprise. And soon.)
Transmeta inside
If Cobalt is moving in the wrong power consumption direction, one of
its direct competitors, Rebel.com, isn't. Rebel's first generation of
NetWinder servers, which -- like Cobalt's -- run Linux, used a
StrongARM processor as their CPU. However, as nifty as the StrongARM
is, and as portable as Linux is, the Rebel.com folks decided to
switch to a 533MHz Transmeta 5400 processor for the second generation
of NetWinders. The NetWinder 3100, decidedly non-cheap at $1700 for
the lowest-end model, has very slick specs, and I'll report on it in
more detail once I've had a chance to use one.
However, what's notable in this context is that despite gaining CPU horsepower, the NetWinder 3100 actually uses less power than its predecessor. According to Rebel, the StrongARM-based NetWinder consumed 14.5W of power at a ready idle, and 17.4W when running full tilt. In contrast, the new 3100, whose integer performance is described as being over 4x that of its immediate ancestor, ready idles at 7.8W and maxes out at 14.1W.
I certainly find the idea of one of these little servers interesting since if I want to have one running 24/7 in my home, I really care about how much power it consumes. Sure, counting fractional watts may seem excessively miserly, but those fractions add up over time.
Smarter consumption
Another reason that the 3100 uses as little power as it does is that
it contains a 2.5" hard drive, ostensibly designed for laptop
computers rather than desktop machines. And as one can divine from
reading the specs of laptop and desktop drives on any of the
mechanism manufacturers' sites, the smaller drives tend to allow more
granular power saving modes, not to mention consume less power in
general.
When looking at a system's power consumption, it's more than just the CPU that's chewing up electricity. The rest of a motherboard's components use power, too, as does mass storage. In a recent thread on Slashdot, even Linux kernel guru Alan Cox has observed that enabling power saving features on his machines was worthwhile, even if much of the benefit derived from monitors powering down.
It appears, though, that letting sub-systems with enough context to make a reasonable decision about power consumption decide how much power to use is a good idea. Transmeta uses this principle to good effect in its processors, since the code morphing software has a great deal of context about the code it's executing and can tell the chip when to use more or less power. Many other peripherals can do the same.
Clever drives
I recently had an interesting conversation with Steve Hetzler, a
Fellow at IBM's Almaden Research Center. His research is focused on
improving IBM's hard drives and one of his areas of expertise is
power management. The conversation went on for the better part of an
hour and I learnt a ton of stuff, but what really struck home with me
was the idea that the more intelligence you can put into a
peripheral, the more interesting things you can do there.
One of Hetzler's precepts is to keep the drive focused on what it really is doing as opposed to what the OS thinks it's doing. This means that many simplistic power saving schemes, like inactivity timeouts, are actually counterproductive since they often spin drives down at inopportune moments and force a spin-up, wasting power.
IBM's drives can be told via a standard ATA API how to balance performance vs conservation; it's left up to the drive to meet the goal. While the laptop drives currently are far more sophisticated in this regard than desktop drives, drive manufacturers now have considerable incentive to make their drives power smart. It turns out there's is quite a lot you can do to save power on a drive beside spinning down the platters.
Some might argue that servers aren't the right place to be saving power, since servers are all about responsiveness. The short answer to this contingent is that there's plenty that can be done to save power without substantially reducing responsiveness. Once the need arises, you can still turn on everything and run all-out to respond to requests. But most servers spend most of their time idling and wasting power. Rather than buying into the simplistic servers-must-suck-power argument, why not think of ways of lowering that consumption whenever possible?
Not as idle as it should be
A final example: I use an Ecrix tape drive for nightly network backups. The
drive is on all the time because it's in a closet, but it's actually
in use for less than an hour per day. Yet the drive fires up its fan
frequently to cool its insides, since the unit draws enough power
even when supposedly idle to heat the tape to the point where it
needs cooling. Repeated requests for a low-power mode, even a simple
time-based one, and even at the expense of taking more time to power
back up when it's time to actually perform a backup, has thus far met
with no success. The response has been that until there is enough
customer demand, it's unlikely to happen.
Perhaps its time customers started demanding these features from their gadgets.
ZDNet columnist Stephan Somogyi recently replaced all the incandescent lightbulbs in his home that weren't on dimmers with low-power compact flourescents. He is awaiting his next power bill.