Hacker News Reader

Comment by storborg | Hacker News Reader

While these numbers are sort of interesting, they are misleading, because the 1.18 cents/search number that the author provides is not actually the cost of a single search.

A more useful analysis would be to estimate the marginal cost of each additional search (e.g. if I go make a search right now, how much extra does that cost Google?).

Likewise, there is a similar number that represents the marginal cost of each additional webpage that Google indexes.

View on HN

Wouldn't that be zero at most off-peak times? Machines are gonna use electricity whether you're using them or not.

Power consumption will vary based on usage on modern machines and Google could do automatic boot/shutdown of machines to match demand. But the marginal cost in electricity alone should still be pretty small.

but its not about running power - its about cooling costs, which IMHO makes up the bulk of a data center's expenditure.

I'm not sure if there has been any innovation in this area (shut off power AND cooling for certain areas in the data center)

Sure, but you don't need to cool heat you haven't generated.

I dont think it works that way - in centralised cooling spaces, they have to keep their cooling machines running. So, it costs the same whether you are generating that amount of heat or not.

To cut down on cooling, you need to architect your farm to have smaller, independent cooling systems. However, your system must also be architected such that weird things like "only machine is active in my cooling cluster, so I cant shut down the cooling for that area" cannot happen.

I'm not sure if things like that have been solved.

The AC will be on 24/7 but surely it will use less power if it has to get rid of less heat.

Correct, HVAC systems are designed to handle variable capacity and to stand down as demand lessens in the same way as a data center is. It's such long standing practice that most recent homes even have variable speed compressors on the condensing unit and multi-speed fans on the air handler.

I think pedrocr means that the baseline would be moved down, so that less cooling in total would be required, not that they reduce the amount of cooling used during usage dips.

There has been a form of innovation - in some server farms that consist of thousands of interchangeable computers, they simply don't try to cool them, and instead just exchange air with the outdoors (possibly with a bit of filtering). If your software doesn't mind having nodes frequently drop out due to overheating, then it apparently is cheaper to replace melted CPUs than to keep them all cool.

http://blogs.technet.com/b/msdatacenters/archive/2011/01/04/...

Which is also why more data centers, including Google data centers, are being built in cool climates. If you're exchanging for cold outside air, you get to keep more of your servers from melting down without paying more for cooling.

According to this link someone else found [1], they claim to use 0.0003 kWh electricity per amortized search, which at e.g. 0.06 USD/kWh (typical [2]) is 1.8*10^-5 USD, or 0.26% of search costs (according to the estimates in the article).

[1] http://googleblog.blogspot.com/2009/01/powering-google-searc...

[2] http://www.eia.doe.gov/cneaf/electricity/epm/table5_6_a.html

(In a related estimate, 0.0003 kWh/search means 54 CPU-core-seconds of indexing per search, assuming (without any justification and probably wrongly) that (i) most electricity consumption is by CPUs during indexing and (ii) they are using x86 machines with about 20W/core power consumption. I'm sure someone else here can fix/improve this estimate.)

I would guess that in such a case the largest portion of cost would be bandwidth.

The calculations are quite tricky given the granularity of data the author is working with. I agree that a marginal cost calulation is a bit more useful as the cost of development staff is not really relevant here as this is an R&D cost. The exception being those employees that work on technologies to sustain the growth of pages and searches and the IT staff to maintain the infrastructure. Without management accounting data the figures will be rough estimates at best.

Well, you would need to calculate the marginal cost if you were really interested in what a particular search costs, but I think this number is actually says much more.

It is the amount that Google is spending on average on a search and thus it is the amount that Google has to earn back on average per search in order to be profitable.

Also the 0.68 cents is cost of search-related revenues divided by number of searches. However, presumably there are only revenues when someone clicks on an ad on the search results page. So a more accurate number would also factor in the ratio of ad clicks to searches. If there's one ad click per 10 pages on average, these numbers are off by an order of magnitude.

I think you're mistaken, his calculating is not dependent on click-through rates.

You're right... I misunderstood it.

"A more useful analysis would be to estimate the marginal cost of each additional search (e.g. if I go make a search right now, how much extra does that cost Google?)."

Interesting, but tricky. Is the price of the new servers they need to buy for the peak processing marginal or fixed cost?

Marginal, but you would have to multiply it by the fraction of times that happens, so it would still be quite low.

I agree in principle, but this fraction would be easy to get if they just had to buy a server after a fixed number of clicks. But they only have to if the peak usage grows. Peak usage grows, they buy some servers, then go on for some time processing huge amounts of clicks. Things might get even more complicated if they decide to spend the (now) excess processor power on something else.

Fixed, because it happens in the long run. That is, they can't buy another server in response to your additional search.

Reply on news.ycombinator.com