Amazon Starts Charging For Cloud Computing Resources By the Second

AmiMoJo writes: "Back in the old days, you needed to buy or lease a server if you needed access to compute power," remembers Amazon's AWS blog. "When Amazon launched EC2 back in 2006, the ability to use an instance for an hour, and to pay only for that hour, was big news. The pay-as-you-go model inspired our customers to think about new ways to develop, test, and run applications of all types."

But now from the 2nd of October, Amazon will start billing Linux virtual machines by the second, with a one minute minimum.

Michael!

[Impy the Impiuos Imp]

"After observing your usage statistics, an hour is way too long. I am now charging by the second, one minute minimum."

That's what she said!

Re:

[uncqual]

And don't forget the charge for 'kilocore ticks'.

Impact on cost?

[Ronin Developer]

I wonder what the net effect will be on the cost to users? Will it reduce costs or, ultimately, raise them? Perhaps, this is explained in the article...guess I should read it.

Re:Impact on cost?

[Zocalo]

They're claiming it'll reduce them, but I suspect it's more nuanced than that. If you've been chopping your work up into smaller enough slices to qualified for free usage (15 minutes or less?) then it's absolutely going to cost you more, because that option seems to be going away. For everyone else though, I guess the de-facto subsidies of those in the former category are going to be coming off your bottom line, so yeah, there's probably going to a reduction for you.

You can bet that Amazon has crunched the numbers and the net overall revenue from EC2 will be going up though.

Re:

[Hognoxious]

Depends how it rounds, doesn't it?

Re: Impact on cost?

[Zocalo]

Not true, it depends on how they adjust the rates they charge per timeslot vs. how many people were getting in under that 15 minute (or whatever it is) free window. If you are now going to be paying a total of $1/s for your EC2 server farm, then the breakeven point would be at $3,600/hour for that same server farm - if you're previous bill was more than $3,600/hour - including any dead time - then you are going to see a reduction in total costs. If you are in the situation where time is money and you are deliberately chopping up a task into a lot of parallel tasks so it finsishes faster and swallowing the resultant expense of having multiple servers sitting idle when the job completes, then the savings could add up pretty fast. The potential win for Amazon comes from all those who were previously getting free minutes through massive parallelisation to get under the free usage bar; that previously written off revenue is now going to be billed, and if Amazon has done their homework almost certainly will exceed the savings made by those in the former group.

Re:

[rtb61]

How about straight up sales pitch. Our computing cycles only cost $100 per minute, buying the hardware outright would cost $10,000. Which would you rather spend $100 (trying to calculate how many minutes you would spend becomes a major computing exercise) or $10,000 (which is the easy answer). End of the year and that is easy total, instead of spending say $250,000 doing it yourself, you find you have spend $1,500,000, youch and now you don't have the expertise to do it any more and must rebuild it at major

Re:Impact on cost?

[dacut]

Disclaimer: I work for AWS, but I'm speaking personally here.

This will always be cheaper for on-demand users. Previously, you were charged a full hour for any fractional usage. As soon as you start an instance, you're being billed -- even a start and stop a second later counted as 1 hour. (There is a free tier: you get 750 hours/month of t2.micro usage on Linux, RHEL, SLES, or Windows, during the first 12 months.) [amazon.com]

Let's say you had a batch job that ran for 12 minutes, 4x/day, on Amazon Linux on a c4.large instance in the Oregon region ($0.100/instance-hour). Before this change, you would have paid $12.20/month (4 instance-hours/day x 30.5 days/month x $0.100/instance-hour). Starting October 2, you will pay $2.44/month (0.8 instance-hours/day x 30.5 days/month x $0.100/instance-hour).

AWS believes that cloud computing is going to be a high volume, relatively low margin business, and Amazon is very comfortable with these types of businesses. AWS has had (as of this writing) 62 price reductions in the last 9 years, largely in the absence of any competitive pressure. (And, since I pay for my personal usage -- no, we don't get a free lunch here! -- that's kept me happy as a customer.) Internally, it's a relentlessly customer-obsessed culture -- you can (and I have!) stopped a VP mid-speech by saying, "Wait, I don't think that's the right thing for the customer!" (We're also a very data-driven culture, so you're expected to have data to support this, of course. :-) )

Hope this helps clear up some of the confusion. Note that there are some cases where billing will continue to be per-hour (or fraction thereof), such as marketplace usage -- Jeff Barr's blog post [amazon.com] has all the details.

Great!

[Opportunist]

I no longer need to buy expensive machines when I need to break just one key or find the clear text for a singe hash!

I always wanted to steal identities, but the large up-front costs always made me shy away from it. But no more!

Thank you, Amazon!

Closing a loophole

[cunina]

There's a small but (until now) growing cottage industry that has helped heavy users of AWS get compute time very, very cheaply. Basically, if you can divide a huge compute job up into a large number of short, nearly stateless jobs, then you can launch a bunch of nodes, run them for less than fifteen minutes each, terminate them after the jobs run, and not have to pay for CPU time. This new move by Amazon puts that to an end.

Re:Closing a loophole

[tomhath]

I suspected the real rationale was something like this. It's even hinted at in TFA:

Many of our customers are dreaming up applications for EC2 that can make good use of a large number of instances for shorter amounts of time, sometimes just a few minutes.

I can't blame Amazon for the change though. In fact, it seems that charging by the hour was more underhanded.

Re:

[hord]

This is one of the first things I thought of when I was exploring the free tier... How many minutes per month? Broken into how small of chunks? How much actual useful CPU time do I get for free in a "month" and how I can (ab)use that? I can use another instance to coordinate everything for $12/month.

Re:

[fluffernutter]

There is a large industry of providers that don't nickle and time people to host severs like Amazon does.

Re: Closing a loophole

[cunina]

Absolutely. Now that they've closed the loophole, people can use that (and pay for it).

Re:

[dacut]

As mentioned previously, I work for AWS, but I'm speaking personally here.

I'm not sure where this idea of "run them for less than fifteen minutes each, terminate them after the jobs run, and not have to pay for CPU time" has come from; this is the first I've heard of it. To my knowledge, this has never been the case. Currently usage from 0-59.999... minutes is billed as 1 instance-hour; 60.0-119.999... as 2 instance-hours; etc. Starting October 2, you will be billed by the second (with a 1 minute minimum).

Re:

[Anonymous Coward]

I'm curious where you are all getting this "Less Than 15 Minute" isn't charged exception from? We run thousands of instances in AWS and this has never been the case, as far as I'm aware of.

Their docs seem to confirm as much:
"When you run an instance for 10 minutes, stop the instance, and then start the instance again, you are billed for two instance-hours."
https://aws.amazon.com/premiumsupport/knowledge-center/ec2-instance-hour-billing/

The change to per-second billing will be a massive cost savings for us

Re:

[hord]

I remember cycle count auditing... I always wondered who paid for the billing system.

Everything old is new again.

[alexhs]

Back in the old days, you needed to buy or lease a server if you needed access to compute power," remembers Amazon's AWS blog.

Someone didn't learn History [wikipedia.org], again.

In the 1960s, [...] users were charged rent for the terminal, a charge for hours of connect time, a charge for seconds of CPU time, and a charge for kilobyte-months of disk storage.

Re:

[Waffle Iron]

The difference is in the price.

I was stuck in one of the last CS classes at my university that used punched cards and batch processing. We got charged several dollars of "funny money" every time we ran a process that constructed a binary tree with a couple of hundred elements. (Between the threat of draining our accounts and the ~1 hour turn-around time per run, I guess we really had an incentive to pour over our code for bugs offline. The whole experience sucked.)

Now, the price for same computation at Amaz

Re:

[swillden]

In the older case, one person could probably rack up enough charges to pay for their own minicomputer

Heh. In the early 90s I was playing with fractals (Julia sets) on the university mainframe. It was an awesome machine for the purpose because it had a vector processor, which allowed me to compute 16 pixels at a time. I had been playing with this for months, always running my jobs at "idle" priority so as to avoid interfering with anyone's real work, when I was called up to the computing center director's office. He wanted me to explain what I was doing and showed me the "bill" for my usage, which was over

Re:

[Shoten]

Back in the old days, you needed to buy or lease a server if you needed access to compute power," remembers Amazon's AWS blog.

Someone didn't learn History [wikipedia.org], again.

In the 1960s, [...] users were charged rent for the terminal, a charge for hours of connect time, a charge for seconds of CPU time, and a charge for kilobyte-months of disk storage.

I think you don't understand the difference between the pricing models, or the actual history you're referring to.

In the 1960s, you were charged rent for the terminal, yes. That was above and beyond the contract just to have access to the mainframe in the first place. You couldn't just go and buy a few seconds...or even a few hours...of compute time. You had to buy access on an annual basis.

Even so, the company type (service bureaus) that you're talking about were more like car rental companies. They ha

Re: Everything old is new again.

[Zero__Kelvin]

No, he understands and you don't. One was literally charged accordingly to the number of CPU clock ticks your program required to run.

Re:

[jmccue]

In the 1960s, you were charged rent for the terminal .... AWS and Azure represent the first time that such access has been available to anyone

I cannot speak about the 60s, but in the early 80s I worked for a mini-computer company that also rented out time on an IBM mainframe they owned. Usage and charges were based upon CPU Time and Disk I/O, this sound very similar to Amazon. I remember printing out the usage/billing reports on green bar for finance. The only difference I see is due to technology and AFAIK Amazon does not charge for disk I/O.

Re: Everything old is new again.

[Entrope]

That business model never really went away, it just became less prominent. For example, you can still buy a "Unisys ClearPath Forward(tm) Dorado" system (their marketing department keeps up with the latest fads, at least...) with a limit on how many MIPS it lets you use.

Is "The creimer Monologues" next?

[tepples]

Is it now trendy to accuse numerous posters of being creimer the way it used to be trendy to accuse numerous posters of being twitter [slashdot.org]?

I love this so much

[Just Some Guy]

This is great for my workload, which looks a lot like "spin a large cluster up, slam it for about 20 minutes, then shut it down". Since the previous minimum was 1 hour, I foresee massive drops in our on-demand costs. It also means that the complicated scheduler we were considering, which was going to be optimized for keeping machines loaded until they were x hours and 58 minutes old, can be tossed out. Seriously, thanks Amazon!

Billing by the second, not charging by the second

[mveloso]

AWS always did accounting by the second, but the minimum charge was by the hour.

This makes AWS cheaper for those who are doing smaller workloads.

Makes life easier for everyone (except Amazon)

[Anonymous Coward]

This is great news IMHO.

Up until now if you wanted to get your money's worth you needed your EC2 workloads to work efficiently with an EC2 Instance Hour. As soon as the hour begins you were paying for the whole thing. So you have to organise your scaling policies, batch processing and anything else around making sure that your EC2 instance is doing something useful for the whole hour.

This is going to reduce effort of cost minimisation, with the immediate impact being cheaper bills for heavy and varied EC2 u