“Pageviews per kilowatt-hour (kWh)”
Noted by Jim Buckmaster (Craigslist), they keep getting kicked out of colocation sites. I’ve heard this a lot lately, the biggest headache at highly scaled sites is the power consumption of all the server hardware. Perhaps it’s simply a side-effect of the horizontal scaling (scale-out): lots of cheaper machines, rather than some big iron (but the latter is not affordable, and might not be suitable for this kind of job anyway).
So the challenge lies with chip manufactureres (CPU, RAM) and hardware companies: provide high performance with lower power consumption. One can probably simply look at 64-bit CPUs there, with the memory addressing capabilities and relative low cost (AMD Opterons being ahead right now, but Intel is coming out with new stuff), 32-bit for this type of server won’t make much sense any more. Of course this is when talking about new hardware setups.
Re the pageviews, I think Jim said they’re currently at around 120,000 pageviews per kWh.
More cores per box is one way to do this – there’s mostly fixed per box overhead. More CPU power per box also helps with load balancing and that’s a scalability issue if page build times vary significantly – it was one of the longer-running Wikipedia aggravations.
“The UltraSPARC T1 processor with CoolThreads technology is the highest-throughput and most eco-responsible processor ever created. It’s a breakthrough discovery for reducing data center energy consumption, while dramatically increasing throughput. Its 32 simultaneous processing threads, drawing about as much power as a light bulb, give you the best performance per watt of any processor available.”
Source: http://www.sun.com/processors/UltraSPARC-T1/
Well, that’s the PR. How does paying $180 per request per second instead of $92 per request per second sound? Some T2000 MediaWiki tests. The machine was returned to Sun so they could work out why it was so slow at PHP, possibly the cache miss rate in tight loops.
Hi,
What makes you say that they “keep getting kicked out of colocation [facilities]”? Cite?
Regardless, isn’t getting more pageviews-per-kWh a more efficient use of energy, thus using less to achieve more?
I wonder if maybe you’ve accidentally misinterpreted what Jim actually said. For example, Russel Shaw at blogs.zdnet.com (http://blogs.zdnet.com/ip-telephony/?p=1192) quotes Jim thus:
[…begin quote…]
“We do worry about how to maximize page views for kilowatt hours,” Jim said. “We’re up to 150,000 pages per kilowatt hour, and got out of a co-lo (co-located facility) because of that.”
[…end quote…]
I don’t know if that’s an accurate quote, but it says to me that they were able to reduce their physical deployment (“getting out of a colo” = getting rid of an expense, not getting booted and having to find a new one) thanks to improved efficiency a la increased pageviews-per-kWh.
Besides, every colo provider I’ve ever used charges for power. They’ll either include a certain size circuit (e.g. 80% of 10Amps) as part of a contract, or they’ll actually bill you based on your actual usage, but one way or another they charge you for it. Assuming they’re not incompetent (perhaps not always a sound assumption), they won’t charge you less than it costs them, so it makes no sense for a colo facility to get rid of a customer based on the customer’s increased use of a commodity that can be marked up. There are stories of Google, in its early days, frustrating colo vendors by sucking up dramatically more power per square foot of cage space than the vendor expected, but that was years ago – today, you’d be hard-pressed to find a colo vendor that doesn’t have provisions for such things spelled out in the contract.
I’m not trying to be combative or anything – just trying to get to the bottom of this. :)
Graham
I did not take that the comment was literal, I regarded it in the context of an anecdote. I don’t think it’s relevant to debate that further, as it distracts from the otherwise interesting aspects of the story (which don not rely on the details of this).