Amazon EC2 freezes Moore's law | Hacker News Reader

Amazon EC2 freezes Moore's law

46 pointsandrewstuart45 comments

The power of an Amazon EC2 small instance has remained the same for years http://aws.amazon.com/ec2/instance-types

With EC2 there is no steady increase each year in computing power. It's like Moore's law has been frozen in time. EC2 server instances are looking decidedly old and underpowered.

Amazon gets all the benefit of hardware costs decreasing and computing power increasing over time, but no increase in power gets passed through to EC2 users.

Very sad. Slow old cloud computing servers are starting to make running your own powerful modern, cheap, fast servers look attractive.

Will cloud computing vendors start to seriously compete on computing power?

View on HN

I don't necessary agree with this post. Many m1 hosts have been upgraded from the original Opteron 2218 hardware to Xeon E5430. In the US East region you may still end up getting an m1.small 2218 though. EC2 fixes CPU allocation, so you won't get very good performance on an m1.small... 1 ECU just isn't designed to provide much compute power (1-1.2 GHz) With spot pricing, you can get small instances for typically about $0.04/hr which is reasonable if you just need a simple web server (EC2 gives you 100 Mb/s+ uplink and generally very good bandwidth & low latency).

The newer m2 instances use x5550 Nehelem hardware and perform quite well. You'll generally get the most compute power for the money with an m2.xlarge instance (6.5 ECUs) which generally performs much better than the supposedly 20 ECU c1.xlarge Xeon E5410 instance. m2.xlarge can generally be had for around $0.17-0.30/hr (spot pricing) which is very reasonable for 17GB ram (compare to Rackspace Cloud $0.96/hr for 16GB cloud server running on less powerful Opteron 2374 hardware).

If you are interested in other providers, I've done a complete specs/performance/price analysis of 20 different cloud providers including EC2 here: http://blog.cloudharmony.com/2010/05/what-is-ecu-cpu-benchma...

Amazon lowered the price of Linux instances by 15% in November 2009 (from 10 cents/hr to 8.5 cents/hr). Source: http://developer.amazonwebservices.com/connect/entry.jspa?ex...

If that doesn't seem like much, you also have to consider that other costs of providing the service may increase to offset hardware savings due to Moore's law. Think of the networking hardware, energy required to cool the datacenter, bandwidth, HVAC repair and maintenance, real estate, staffing, insurance. These costs are not likely to decrease at the same rate as computer hardware costs, and some, like rent and payroll actually increase.

I'd prefer more computing power to lower costs.

With cloud computing, lower costs = more computing power. If you were running 20 instances before, with the 15% reduction in price, you can now run 23 for the same price.

The whole point of the cloud is that you pay for a set amount of computing power. If you want more computing power get the high CPU instance. Everything is virtualized down to the incremental unit of computing power that is desirable. It's pretty optimal in my opinion and I really appreciate the price decreases overtime.

I was just thinking the exact opposite of this in the shower this morning.

Amazon is one of the few businesses I'm aware of that goes out of its way to keep its profit margin consistent, to the point of regularly lowering the price it charges for its AWS offerings.

It's in their best interest to do so in the long run. Sure, they could fix prices and watch as a 1U EC2 instance gets more profitable over time and 100GB S3 bucket does the same, but eventually somebody else would build a business by squeezing in at a lower price.

As it stands, Amazon has the Cloud game beaten. They're just miles ahead of the next guy. Since the only reason to consider switching is price, it's a genius move on their part to keep walking the price down. If you want to beat them, you have to figure out how to build a better product and squeak it in cheaper.

I just don't see that happening.

They've offered more and more powerful instances as hardware has improved, and they offer spot pricing allowing the market to set the price for spare cycles.

Offering smaller instances is a good thing for a variety of reasons from simple cost savings when you don't need much to efficiently slicing up arbitrary hardware for embarassingly parallel compute jobs.

At root it seems you're essentially asking Amazon to replace all their datacenter servers every 18 months to keep up with processor technology.

I don't think the economics work out for that. There have been price cuts, and I bet there will be more as Amazon pays off the hardware. Once their investment is amortized and the infrastructure is paid for, at that point the only costs on the balance sheet are the operating costs and prices would bottom out.

Worth pointing out too is that Moore's law governs transistors/cm2/$, not CPU speed as such. Thus Amazon could take advantage of Moore's law simply to fit more instances on a given chip. They may not necessarily pass these lower costs onto the customer _as rate decreases_ if the current rates are competitive in the market, and at the moment they appear to be - why would Amazon bite into their own profit margins for no reason?

I have a pet theory (which I think leads to wrong conclusions at the moment, but hey) that once CPU becomes a utility, Jevons paradox will apply and prices will start to decrease more slowly as opposed to collapsing like it has to date. FWIW I wrote the thing up here: http://ianso.blogspot.com/2010/05/jevons-paradox-moores-law-...

EC2 has been growing rapidly and Amazon has been installing newer (and thus cheaper) generations of hardware, so on average their hardware should be fairly new. Even once they reach steady state, the average perf/$ of their hardware should increase as old stuff is replaced. Clearly Amazon's costs per ECU are declining even though they don't replace equipment every 18 months. I think you're right that the lack of competition is why they aren't lowering prices much.

Don't confuse an EC2 instance with real hardware. It's a convenient measure of computing resources used for billing purposes.

I'm not sure why you say this. When Amazon says that you get 1.7GB of RAM, that's what you get. It's not an abstraction. The price of DRAM seems to be falling a lot faster than Amazon's prices.

The price of DRAM isn't falling though. A year and a half ago 6GB of DDR3 was cheaper than it is now. Sure over the long term its getting cheaper, but it takes awhile.

At least Linode recently gave everyone a 42% RAM increase. Good point though, most of the hardware progress has been pure margin for the cloud vendors.

I'd say Linode are about the only people passing on the savings.

(Disclaimer: I use my linode for just about freakin' everything)

Linode is VPS, EC2 is cloud computing. Although I agree that Linode is probably the best VPS provider around, EC2 provides a lot of cloud features that Linode does not including CloudWatch monitoring, autoscaling, external EBS storage, greater memory and compute power options, AMIs, virtual private cloud. So Linode vs EC2 is not exactly an apples to apples comparison. If all you need is VPS, then Linode is really the best option available.

>Linode is VPS, EC2 is cloud computing.

No diff in utilization. You can still use the API elastically spin up new instances, if at a different scale.

>...$features..

I guess, if that sort of thing matters to you. I don't really like relying on third parties for utilities like that.

> If all you need is VPS, then Linode is really the best option available.

K, but has anybody been making their machines faster?

> No diff in utilization. You can still use the API elastically spin up new instances, if at a different scale. They both have APIs, but EC2 has hourly billing versus monthly for Linode (they issue service credits pro-rated daily if you don't use a VPS for the entire month, refunds are manual and require a processing fee). EC2 also offers more options in terms of CPU and memory (3+ different CPUs and 1.7-32GB versus 0.5-4GB and single L5520 CPU architecture). They are both elastic, but from a billing perspective, Linode's elasticity isn't too practical.

> I guess, if that sort of thing matters to you. I don't really like relying on third parties for utilities like that. It may matter for your project, if you need automatic-elasticity, more durable off-instance storage, better security, more compute power, more memory. Linode can't provide these features.

> K, but has anybody been making their machines faster? Sure, EC2 has upgraded their m1 hardware the past couple of years, GoGrid is upgrading to Westmere later this year, Storm on Demand has upgraded some of their hardware, Flexiscale has upgraded, and others.

Linode provides 512 all the way to 20G instances, fwiw.

You're right, I meant 0.5-14GB, I hadn't noticed their new 16 and 20GB instances.

The point is that you are not actually using their machines. You're using virtual machines that have only ever given you a fraction of the machine's total power.

>only ever given you a fraction of the machine's total power

And cloud instances aren't precisely that, albeit generalized and abstracted further?

The thing is, computers are not getting faster anymore.

The only thing that is changing is parallel processing. So amazon is still selling a single CPU, and that single CPUs speed has hardly changed in years.

You should get more memory and disk space though.

Performance/watt is still increasing, though. This helps Amazon's margins.

But watts/dollar is quickly decreasing. I wonder if performance per dollar spent on energy is increasing or decreasing. I wouldn't be surprised if it's decreasing.

Actually, if prices are not increasing, then they are falling. Prices fall by the inflation rate. Let inflation rate = 3% then the price is falling 3% per year. Not major but still a decrease.

Add to that that EC2 prices actually are falling, from .10 to .085/hr for linux boxes for example, and it seems like a pretty good deal.

If you want to get the lowest pricing on EC2, use a spot instance request where your bid price >= standard pricing (i.e. 0.085/hr for m1.small). This will essentially get you a spot priced, but indefinitely persistent instance and you'll typically pay 1/2 or less than the standard pricing. I've had a spot instance running for 3 months paying about 45% off list price on average.

Can the spot price rise above the standard pricing? If so, are you able to switch to standard pricing without losing your instance(s)?

It could, but it is extremely rare. Check out http://cloudexchange.org to see historical pricing. A spot instance cannot be changed to standard pricing.

Do you actually, when running your own server cluster, upgrade every 18 months to keep up with Moore's law? I know I certainly don't. It was only just recently we phased out x86 boxes.

Will cloud computing compete over computing power? That depends, does anyone turn to cloud computing for serious computing performance?

When EC2 came out, your only option was 1 compute unit (CU) with 2GB ram. Now they have high CPU and high memory options up to 26CU and 68GB ram.

If you want power, you can certainly get it.

Couldn't write a couple paragraphs on a blog and then link to that?

Amazon's cut prices, and the high end of their offerings have gone up. Yes, they're going to take a disproportionate share of the benefits from speed if they can get away with it. They also took all the risk on the outlay and used to charge comparatively less per dollar of provisioned hardware when they bought those chips. So I don't think they're ripping anyone off, yet.

If it becomes a bad deal, people are free to go elsewhere - likely quite easily, what with EC2's architecture.

Are you really complaining about a lack of blogspam? Dear lord.

I'd prefer people say as little as is needed to spur thought and conversation without directing me to some ghetto blogspot page that lurched into existence a week prior.

Excuse me, why is this being downvoted?

Maybe people don't see a particular need for this to be a blog post?

Don't ask meta questions like that please, distracts from the conversation at hand.

Well geez, sorry for ruining your morning.

My post was approx 20% meta question, 80% substantive, by word count.

See? You keep getting downvoted. Exercise some self-control and dignity and focus on the topic at hand. I'm violating my own habits here to try to show you what you're doing.

And getting back to the subject, are you aware of any companies that have been upgrading their servers beyond just RAM (such as linode)?

Yeah, Amazon. Cutting prices across the board at the low-medium end and adding new options at the high end. As I said in my original comment.

> Sorry if you couldn't get that in between all of my awful, antisocial behavior.

The griping and passive-aggressive comments are what's causing the down votes. Cut it out

As for the price cuts, at the small-medium size ranges, the cuts haven't been extensive, and certainly not to the same scale as what Moore's law would indicate. The new options at the high-end aren't an upgrade if they cost more (which is in fact the case).

At present the best deal I've been able to find in terms of $/performance has been Linode, but they aren't the most reliable in the world so far.

Moore's law is a law of economics, not nature. Competition drives the performance improvements coming out of Intel and AMD, and competition will drive the rate at which savings are passed along to the consumer in cloud service providers.

It's not even really a law of economics or nature. Or a law. It's a state of a tendency if the rate of expansion of transistor density.

At present, it's more economical (if you don't need programmatic elastic expansion of your cluster) to rent/lease dedicated hardware or set up your own. This is because hardware has been getting faster than they have been upgrading.

>Competition drives the performance improvements coming out of Intel and AMD, and competition will drive the rate at which savings are passed along to the consumer in cloud service providers.

The point of this discussion is that this hasn't happened yet with any known majors beyond RAM upgrades and small amounts of price-twiddling.

I am going to move past the OP's face-value argument.

The point is that amazon pricing is fairly static. You get X for $Y. However as the years go by, the value of X decreases yet you are still charged $Y. So basically going to a hosted service I might get a significantly faster machine for my buck. Sure it won't scale as quickly but that might be a minor issue, pick your battles so to speak.

Amazon needs to take into account this, and instead of upgrading everything, just start deflating prices of the old and buying newer hardware every 2 years or so and charging full price for the new hardware. This keeps them well balanced with technology. In the end this just means that instead of buying 1 server, you will buy 2 for the same price outdated servers and vuala.

However thats not quite how servers work and its not like their cost-for-operations is going down with the years as well.

amazon offers a product. people buy it. if people were concerned with getting more for their money or following moore's law, they'd buy something else. instead, people buy from amazon because they're an industry leader and have good community resources available.

moore's law is about the level of modern hardware capabilities. what servers amazon uses isn't going to affect that much.

Reply on news.ycombinator.com