EC2 Spot Instance Pricing Algorithm Reverse Engineered

ladypine.livejournal.com

10 points·timf··4 comments

It seems plausible to me that demand for spot capacity is very low because most customers either don’t understand how to use it or the savings doesn’t cover the additional risk. I wonder if the reserve price correlates with electricity cost or weather; following James Hamilton’s advice, the natural reserve price for a server would be the cost of powering and cooling it (at any lower price, you’d be better off turning it off).

Spot instances are used to sell spare capacity. This means the provider is basically selling computing power on machines that have to stay turned on, to allow for flexibility in accommodating needs of deals with better SLAs (e.g., on-demand and reserved instances). Hence, the natural reserve price would not be the cost of powering a server, but rather the difference between the cost of an idle and active server.

the natural reserve price would not be the cost of powering a server, but rather the difference between the cost of an idle and active server.

For an idle slot in a partially-allocated server, I would agree, and I think that cost is near zero. But EC2 may also have more completely idle servers than are necessary for SLAs and thus can be powered down.

My guess is that those servers, which can be powered down, are not offered as spot instances at all. A spot instance usually pays less than half the price of an on-demand, for the same power consumption. Offering servers that can be shut down at half the price would cannibalize the main deals the provider offers, and thus make no sense.

Reply on news.ycombinator.com