The Glacier posts generated much discussion that revealed some non-intuitive ideas about Amazon Web Services.
- AWS doesn’t care about profits on Glacier. They have other reasons for offering a cheap archive service.
- Glacier is based on S3. They do some things, like spinning down disks, using old disks, advanced erasure codes, special low-power controllers, high-density enclosures etc. to drive costs down.
- Powered down disks. That’s it.
- It’s a wildly over-provisioned group of servers/disks. Only a few can be powered on at any time.
- It’s tape. No evidence for that from vendors and Amazon denied it, but . . . .
Other than the first and last arguments these are saying, in different ways, Glacier is justifiable based on power and power provisioning savings. Remember Glacier is still, after the recent price cuts, 1/3rd the price of S3.
What data do we have on power and power provisioning?
Mighty power rangers
From a Google paper, Power Provisioning for a Warehouse-sized Computer:
Typical datacenter building costs fall between $10 and $20 per deployed Watt of peak critical power (power for computing equipment only, excluding cooling and other ancillary loads) [25], while electricity costs in the U.S. are approximately $0.80/Watt-year (less than that in areas where large datacenters tend to be deployed).
That translates to ≈10-20 Watt-years is equal the cost of provisioning one Watt.
The paper, The Cost of a Cloud: Research Problems in Data Center Networks from Microsoft, with now-AWS architect James Hamilton as a co-author, presented a table breaking costs down further:
A later paper from July 2012, Battery Provisioning and Associated Costs for Data Center Power Capping from UCSD researchers, uses more recent data:
These sources are in rough agreement on one point: the cost of power and power provisioning – including UPS, cooling, wiring etc. – is less than half of the total cost of a data center. Even if power and power provisioning were free, the savings would be only 40-50% of the total data center expense.
The StorageMojo take
We can rule out power savings as the sole driver of Glacier pricing. AWS must save money on the infrastructure side as well.
Unless, of course, AWS is irrational. Then all bets are off.
StorageMojo’s assumption is that AWS is rational: they want every service to make money, and given Amazon’s low margins and Mr. Bezos’ finance background, they have to. But it can be rational to have a money-losing service, as long as the service covers all the variable costs and makes a contribution to the fixed costs.
Could Glacier be a money-losing service? Cost accounting is a dark art, but typically a business will only offer such a service – dry cleaners laundering shirts for instance – if there’s customer demand and they’ll lose business if they don’t offer it. The key: cover your variable costs and make a contribution to your fixed costs.
Is someone else offering cheap cloud archiving that AWS is worried about? Readers?
Courteous comments welcome, of course.
I know it’s boring, but I’m betting that Glacier is tape-based…
From Amazon Glacier FAQ:
“Amazon Glacier is an extremely low-cost storage service that provides secure, durable, and flexible storage for data backup and archival.”
You have all the characteristics of a storage system based on … TAPE!
Now, IMHO, AWS still makes a decent profit even with the recent price cut…
only google with it’s unlimited war chest of advertising income is a competitor and they are way behind on the software side I expect. Depreciation is an frequently gamed accounting gimmick. If it doesn’t accurately represent the total purchase cost of the hardware divided by the service life then it’s inclusion in and relative size in the pie is bogus.
Does Glacier make money hand over fist? Probably not. EC2, S3 and RDS and the other services are the money makers and on their own justify the building they sit in. If you don’t have to run chillers or turbo fans in the Glacier room since the power draw is a fraction of the S3 counterpart, then the cost of the service declines even more.
I don’t know what it is, but I think a lot of people decided to hate on Blu-Ray and that is that.
Blu-Ray came in around the time Apple started getting mindshare and the “cool” thing was to watch video on a tiny screen on your phone and watching HD video on a TV was something your grandfather does.
I have a portable Blu-Ray writer that is pretty amazing in terms of write speed. It may not be fashionable, but sending Blu-Rays in the mail is a practical way to send 100GB+ data sets, and it’s much cheaper than the “fashionable” LTE networks run by Verizon and AT&T which cost $10/GB.
I refer you to the Amazon Shareholders Letter of 1997 for Bezos’ view on how to make money and how that might influence your thoughts regarding rational/irrational.
He’s playing the long game…
http://benhorowitz.files.wordpress.com/2010/05/amzn_shareholder-letter-20072.pdf
Maybe Glacier isn’t *what* it is, but *where* it is. Knock cooling/etc out of the picture and more of your costs are gone…
😉
Robin,
Fascinating post. The view from the dc/cloud provider’s perspective is very interesting. There is, I think, a divergent view from the user’s perspective, who (in the private cloud space anyway) adds loads of software costs into the mix. Flash deployments that eliminate servers with all that they entail in software licensing and maintenance costs is gigantic. Software providers can respond by leaving server-based licensing behind but that tends to be a really difficult process.
The other issue not addressed which is really big is WCC/ROIC and risk, both time and scale related. It’s really easy to say in hind sight that power/space levels aren’t that bad. It’s another thing entirely to contemplate the initial budgeting decision to build out the traditional hdd-based dc with all the servers and all the power/space costs that have to be provisioned to make it work. Smaller, cheaper, less complicated means more agility and speed and less risk. That’s the approach that wins every time.