"Amazon need to lift their game in terms of CPU performance. They offer a great service that obviously extends far beyond a simple CPU benchmark. But when you can get the same work done in Joyent significantly faster for the comparable price, you’ll get far more mileage per instance, which is ultimately going to save the dollars."
Or, "Wow, Amazon sure makes bank on this EC2 stuff, they are selling kilo-core-seconds for a bunch more than Joyent is and they are still the most widely used provider."
There have been a couple of good benchmarks around which instances get you the most 'bang' for the buck, but its pretty clear that Amazon is making a ton of money here.
A good comparison is a rack of 40 SuperMicro 1U dual socket servers with 192G of ram in them. You can create 480 dedicated instances with 16 GB of RAM each, and run the whole thing in a colo facility for $5K/month (power, networking, and floor space). That is 0.2 cents per second, or a bit more than $10/month running them 100% of the time.
That, $0.2cents/second of course, is on the margin. The real challenge to doing this is, of course, not the hardware/hvac/colo costs, but the several millions (tens of millions?) of dollars associated with building the software infrastructure required to scale/deploy the CPU/Storage/Services dynamically. And then there is the well staffed devops/NOC environment required. And the single-minded dedication to building these services so reliable, that you can run your primary business on. Oh, and course having a massively scaled business to both offset all those devops/noc/engineering efforts, and beta-test your scaling environment/server/storage infrastructure.
Other than Amazon, I can only think of Google being in a position in which such an enterprise can take off.
We tried this at Loudcloud, and it turns out you need more than a bunch of smart people, lots of money, and single-minded focus towards building on-demand infrastructure.
Of course today you can get over 90% of the required software off the shelf, see cloudstack and openstack, both used in clusters of 1000s of production servers running 10,000s of VMs.
I recently tried out Joyent SmartOS for a NodeJS consulting gig I'm doing. After about 30 seconds, I killed the vm and switched to Ubuntu 12 server.
I don't care if SmartOS is 'faster' or has dtrace. The fact that it is missing basic things like top (no prstat isn't anything close to top) and 'standard' things like upstart made me want to run for the hills. The Joyent wiki documentation on SmartOS is abysmal at best. Even the pkgsrc repository system feels ancient and pales in comparison to what apt is. I have no desire to spend time re-inventing basic tools that linux has had for ages.
The image version of SmartOS w/ NodeJS isn't up-to-date with the latest Node release and it starts up MongoDB by default. While this may be good for just getting things running quickly, what happens if I don't want Mongo running at all? Generate my own images? No thanks.
I know, lots of hate... but it isn't all bad. Joyent as a cloud platform is great. Basic ab testing against NodeJS had me serving 9900 concurrent connections without any issues on their least expensive vm (which runs on a blazing fast network at 10gb/s). We were on Rackspace before and with a network limited to 20Mbps/s (at the lowest tier) we just couldn't get the throughput we needed.
tl;dr:
SmartOS is a train wreck. Don't waste your time.
Joyent as a platform is great.
Why bother with EC2 when Joyent is faster and less expensive?
For whatever it's worth, we are very actively working to make SmartOS a more hospitable environment to those coming from other systems. I'm not sure if you're interested in the details here (after all, it sounds like you might have spent longer on the HN comment than on the "30 seconds" with SmartOS), but if you provision the default base smartos64 image (currently version 1.6.3), "top" is available via pkgsrc ("pkgin in top"). Not only that, but due to popular demand, we've also ported "htop" to SmartOS ("pkgin in htop"). In terms of node.js version, I think what you're referring to is the node found in the base platform (/usr/bin/node), not the node available via pkgsrc (which will appear as /opt/local/bin/node). This is definitely a point of confusion, and we've rectified it in the next version of our software by not having the node version that we use in the base platform be visible to SmartMachines (that is, there isn't a /usr/bin/node).
More generally, we're definitely committed to making SmartOS the best environment for the modern cloud; I'm sorry that your experience was frustrating, and if you have any additional concrete feedback on what you'd like to see improved, we're all ears. And finally, I know you said that you "don't care" about DTrace, but I just can't help my (biased) self; if you're developing a node.js app, you really should check out some of the things that SmartOS can uniquely do with respect to node:
As a new customer, I feel like Joyent needs to focus on simplifying things quite a bit. I want to run NodeJS on a server. I install the aptly titled 'SmartMachine Node.js' product and expect it to be the right version and the basic tools installed to get the job done. Instead, I get this morass of oh, you didn't install the right machine or oh, node is in some other directory now. Can you see how that might be distracting?
Sure, I didn't spend a lot of time with it, why should I? I saw that it was missing a great deal of basic things out of the box. I decided quickly to not waste my time learning more esoteric stuff that is only useful if I stick with Joyent and SmartOS.
You are an absolute genius for creating dtrace and it is nice to have but certainly isn't a necessity for most applications. Your example is voxer which is a total edge case since they've been pushing the node envelope since day one.
In the end, the reality is that SmartOS doesn't feel modern at all, it feels like Windows98. Actually, it feels like SunOS back when I had pizza boxes under my desk 15 years ago. I could spend time giving you tons of (free) feedback about what's wrong with SmartOS or I could just use Ubuntu and move on.
9900+ concurrent connections on a $22/mo Ubuntu box is 'good enough' for me.
The things you are complaining about -- top vs. prstat, dtrace, upstart vs. smf, clear and comprehensive documentation vs. man pages written in an afternoon or blog posts on the internet, apt vs. ips -- these are things that are objectively better on Solaris in a quantifiable way. If they make you feel like you're working with Windows 98, I don't think the problem is with the tools, but somewhere else.
A few days ago there was an article about Linux monoculture. It's even worse than that, it's an Ubuntu monoculture.
If you really believe there is more information in SmartOS' wiki than in man pages & other proper documentation available about Linux, I don't think the problem is with the Linux monoculture, but somewhere else.
You're complaining that SmartOS doesn't ship with inferior technology? That's not a valid criticism. Perhaps a valid criticism might be that Joyent doesn't make it clear enough why prstat, DTrace and SMF are superior to Linux alternatives, and where would you find documentation for all these things. It's understandable that you searched the SmartOS wiki, but the Oracle Solaris documentation would have been better if you were not familiar with Solaris.
Joyent has to make the experience better by streamlining documentation, not by compromising its core values that make SmartOS great. The quality of the Solaris documentation is great, it's just not organized how Linux people expect it to be and they don't know where to find it.
We used SmartOS (or its predecessor) for 3+ years. The experience was overall very positive and performance was great. Joyent's reliability is outstanding. Sure, some of our software was dated but everything worked.
The bigger issue is Joyent's product lifespans. We've had to migrate our servers twice in that three or so years. Because of that, I moved to a different cloud provider. I don't know if I can expect better.
The only outage I can recall was recently when they began the process of EOLing the current product we use from them.
You can very easily install whatever version of Node.js you want using this bash script: https://gist.github.com/a0d40f1ff1679ed582a5. Also, MongoDB isn't even installed by default on the Node.js instance.
As for your bash script... I stopped compiling binaries ages ago, why would I want to start again when I can just use Ubuntu and the official NodeJS PPA?
Sorting 5 million random numbers takes ~20secs - and this is the fastest benchmark?
Wait there buddy.... First of all, why would OS choice be of so much significance here, after all this is all executed mainly on the CPU (not much IO). The only difference would be that the random() function might return different things on different systems (libc difference?)... or maybe not (I'm not familiar with ruby's internals).
So I'm not sure what this measures at all...
At least you can run the benchmark shootout (CPU), or something that can test the I/O.
- No one pick a host based on a single microbenchmark, please. Game out what your app would need to achieve your target performance level with what each host offers. Amazon still might lose (sounds like that happened with TripAdvisor) but at least you've made your decision a sane way.
- For the best price, you'd want to use a reserved instance if this is a webserver, and perhaps a spot instance if it's a batch/HPC task. You'd want to use high-CPU instances if you really are all about quickly sorting 5M numbers. If your app is or might be come database-bound, DB perf's super key.
- I'd love to hear more broadly about the practical differences between the clouds--whether anyone seems to be better at managing noisy neighbors than the next guy, network and I/O performance, etc. One CPU-bound task can't really be the end of it.
Did the math on how to get cheaper ECU-hours from Amazon. Prices below are per ECU-hour, not per instance-hour; '1yr' is the one-year reserved instance price with the upfront cost amortized; 'spot' is the spot price in the US West (Oregon) region when I looked just now.
Small: 6.5 cents/ecu-hr
Small, 1yr: 4.7 cents/ecu-hr
Small, spot: 1 cent/ecu-hr
High-CPU: 3.3 cents/ecu-hr
High-CPU, 1yr: 1.9 cents/ecu-hr
High-CPU, spot: 0.56 cents/ecu-hr
If RAM matters, high-CPU (7G, or 3.5/core) > Small (1.7G) > EC2 Micro and Joyent extra-small (~0.5G). Again, don't rely too much on figures like these; look at all your needs (not just performance, everything) and try to figure out what'll work for you.
cc2.8xlarge is the best in terms of price/ECU-hours. You get a whopping 60GB of RAM, 10Gbps network interface. Downsides are: EBS only, us-east-1 only.
Ah, it is indeed cheaper, but I'm getting different specific numbers: 240 cents/88 ecu-hr == 2.7 cents/ecu-hr on demand; less if reserved or spot. cc2.8xl is also in the Oregon and Ireland regions, but not others as far as I can tell.
1) Funny that scaling up to one of their largest instance types can improve bang-for-buck, on this metric at least. 2) I wish Amazon made it simpler to sort all this out.
I have been asking, and wondering why would anyone want to use AWS. It is insanely expensive. It is Slow on CPU, Slow on I/O, Low on Bandwidth, Low on memory, and Zero Support. Compared to many other alternatives that gives you 3 – 4 times the on all the above plus better support for the same price. Even their High CPU and Memory plan is still on the expensive side, only the High I/O SSD offers some value. AWS works great when you are in the mid range of expanding, then you can calculate your usage and book their server in advance using Reserved Instances. Instagram and Netflix as example. But that is hassle you have to put though, there are alternatives that are priced and perform just as well without using Reserved Instances. And once you got past the mid range, a managed dedicated server and some instant cloud severs should be much more cost effective.
Apart from being forced by VC ( because they are heavily invested in Amazon ) i could not understand why startup are choosing AWS at all.
You're generalizing too much if you believe that the whole AWS stack is bad.
S3 is very reliable, and I'm not aware of anything as reliable (we're talking eleven nines)/simple/cheap/distributed.
EC2 gives you the ability to manage instances in several continents with a dead simple API, which is unheard of anywhere else at this price.
ELB is very reliable, simple to use, and saves you hours of setup.
My point is that if you really wonder why anyone would want to use AWS, you might not be considering AWS, but only EC2, and in a very non-international way.
Arh. I should have written EC2. Sorry for the confusion. ( Will double check next time before i make a post )
I do realize there is other AWS services that are great and simple to use. But for hosting i dont think anyone should choose EC2.
We use AWS (and with great satisfaction) because it provides a full offering. We don't just need a bunch of servers, we also need databases, one or more queues, backup storage, caching and so forth. AWS provides all of these services, offers one (reasonably clear) dashboard for them and makes it insanely easy to get them all running and tied together. As for the Zero Support thing: our experience is that it is quite easy to get support from people at Amazon. I've been emailing extensively with one of their platform architects which was actually a very nice experience.
I have been asking, and wondering why would anyone want to use AWS
It's dead-simple to get started.
It's widely known. I can convince my CTO to use Amazon for cloud-hosting. He's never heard of Rackspace or Joylent - people who are serious about running their infrastructure on the cloud will research and know about them but for quick experimenting and companies just getting their feet wet, Amazon has a big edge here.
I don't understand why Joyent clings to a dead operating system.
Are they planning to single-handedly keep solaris alive?
Cloud computing is a commodity. Many (most?) customers use multiple clouds in parallel. Learning and maintaining the intricacies of a niche OS in this environment is a nonsensical proposition.
And for what? For a few percentage points on a CPU benchmark? Seriously?
If you think this benchmark is impressive... just do some I/O benchmarking. It blows EC2 away much harder than it does for CPU. ZFS on local disks with SSDs for L2ARC, and no virtualization overhead (this obviously applies to SmartOS). Oh, and the performance is predictable and doesn't vary in time. And you have DTrace to understand your application.
To me, the weakest part of AWS is always on the technical support - you have no right to create a ticket to them unless you pay for the support fee (as contrast to Joyent /Rackspace/Linode), even you have discovered a bug in their system!
On the other hand, I also agree with @latchkey that SmartOS /Ubuntu is an important factor when we are choosing the platform - for a lot of people, ease of use and larger community is more important than cutting edge features like dtrace and zfs.
If you serve the US market Joyent is maybe good, if you serve the EU market you have way better deals with EU providers. Joyent have some servers in EU but they are just collocating. EC2 is available worldwide for good latency.
Joyent love micro benchmarks because they allow one customer to grab a lot of CPU in burst, but if you are one of the other customers on the same hardware node, it is better to have a bit slower all the time than high variability like Joyent. Predictability when you run your infrastructure is better.
Also, Joyent will kick you out of their infrastructure if they consider at their sole discretion that they are not making enough money out of you, but this is my personal experience and may not be yours. So basically, they broke the trust I placed in them, I am migrating out at the moment, so take my comment as coming from someone with a bad experience with them.
Bare metal, the best is OVH. You get a server in 1h in France or in Canada. You can also get instances for pretty cheap but I have only used them as throw away test instances. Oh, they have free bandwidth between all their datacenters, even between Canada and France as they managed their own transatlantic links.
They have a very decent "private cloud" offer. The feedback is good but I am a bare metal + Ganeti[1] to manage my VMs kind of person.
If you like pure instances, I have also instance running on Gandi[2] infrastructure. My blog made several times the front page of HN and without special caching it handled extremely well the load (in fact, you could barely see the increase in CPU usage in the monitoring graph while seeing a huge increase of the outbound bandwidth).
Another good provider is Hetzner[3]. You need to pay attention to the server you select (ECC memory or not) but then they are reliable (I had a server with them for 4 years or so without a single issue).
In Europe, AWS is in Irland, I suppose it was for tax purpose that they selected this location, but from a latency point of view, this is stupid if you want to deliver packets in the South or East of Europe (and even part of Scandinavia because most of the time your route does IE, NL or FR, and then goes North).
Sorry for the long post... but at the end of the day, do your calculations. AWS is great if like Netflix you scale from 20% to 100% during the day[4] (the hit on the cache gives you how they scale) but if you are a normal shop without day/night activity because you are at global level or so well optimized that you just not need to scale (a single dedicated server with SSD is incredibly powerful nowadays) you simply do not need "the cloud". Just test with 1k€, do a bit of maths, then select the best choice.
We also like & use most of those providers with our own food (http://comodit.com) to deploy/manage them, even if they are bare metal. We also mix them in hybrid cloud scenarios.
If you are okay with working on bare metal (i.e. no virtualization out of the box) than Hetzner (www.hetzner.de) is a very interesting partner. My experience with them is very good. Ex-colleagues are also happy with them.
At my last job, I did some evaluation between EC2 and Joyent with a node.js based application. Hands down, Joyent beat EC2 up -- there wasn't an explicit 1:1 correlation anywhere close, as I was seeing the performance with a m1.large go in line with what I was seeing for a small instance at joyent.
When you got right down to it, what killed that plan was that you were still using Amazon S3 for storage; the money that we would have saved in CPU/processor time we would be eaten alive with transferring that data back to S3.
We've been using Joyent in production for close to a year (coming from dedicated servers and EC2), and we've had nothing but an outstanding experience.
After seeing this, I thought I would spin up a server with Joyent, just to give it a try. After 10 minutes, the server is still not marked as "provisioning," which seems like an excessive amount of time.
I've had quiet a bit of experience provisioning machines on Joyent. 95% things work flawlessly, with 20-40 second boot times. However occasionally I'll get 10m+ boot times or machines that just fail to boot. Sounds like you got unlucky. Delete that machine and try again.
Running a benchmark on services like this can be really tricky.
It really depends on which cluster you were allocated and what other applications were running at that time. If someone else was running a heavy process at the same time, then that could explain some of the sluggishness..
Even still, the numbers are so drastically different that this question does deserve a second look. Someone from Joyent or EC2 should really put together a solid benchmark to prove once and for that their service is indeed better in a wide range of compute and networking categories.
I haven't tried out Joyent, but I've tested out Linode and Storm-on-Demand and they easily beat Amazon EC2. My benchmark was running our production image processing algorithm that utilizes OpenCV via python on Ubuntu Server (definitely CPU-limited).
FWIW, here are my benchmarks, along with pricing for Northern California regions:
Linode 512: 15.2 seconds, 65 cents per day
Storm-on-demand 1 GB: 13.05 seconds, 5 cents per hour
EC2-micro: 168 seconds, 2.5 cents per hour
EC2-small: 63 seconds, 9 cents per hour
EC2-medium (high CPU): 12.13 seconds, 18.6 cents per hour
That's very interesting! I converted the prices to be in the same units and did some calculations for better comparison.
Linode
Cost per day: $0.65
Time per image: 15.2
Images per day: 5684.2105263158
Cost per image: $0.0001143519
Images per dollar : 8744.9392712551
Storm-on-demand
Cost per day: $1.2
Time per image: 13.05
Images per day: 6620.6896551724
Cost per image: $0.00018125
Images per dollar : 5517.2413793104
EC2-micro
Cost per day: $0.84
Time per image: 168
Images per day: 514.2857142857
Cost per image: 0.0016333333
Images per dollar : 612.2448979592
EC2-small
Cost per day: 2.16
Time per image: 63
Images per day: 1371.4285714286
Cost per image: 0.001575
Images per dollar : 634.9206349206
EC2-medium
Cost per day: $4.464
Time per image: 12.13
Images per day: 7122.8359439407
Cost per image: 0.0006267167
Images per dollar : 1595.6173709545
An AWS micro is designed for burstable CPU loads, i.e. Websites. Performing a CPU intensive task will never do well. Amazon even says this in their own documentation.
The second fault I have is that articles like this, is that they only every look at the OnDemand price, which most people using AWS seriously don't pay. AWS has reserved pricing and spot both of which will dramatically drop the pricing. Maybe not as inexpensively as Joyent, but much much closer.
For Windows servers, Joyent is prohibitively more expensive than EC2... though their Linux prices trounce AWS' significantly. I wonder why that is. They also don't offer anything smaller than "medium".
Does anyone know a good cloud-based Windows host as an alternative to AWS?
Azure IaaS is even more expensive on a small scale (though it does get cheaper as you scale up). Azure PaaS would fit the bill very nicely, but someone at MS completely dropped the ball: you have no option of using SSL + your own domain!
You can have either your own domain name or SSL w/ their subdomain.
Who the bleep would settle for that? What kind of "enterprise solution" is this?
You probably mean the Azure Websites service. It doesn't support SSL on custom domains yet, but it's a preview version. That feature is expected before this service is released.
The PaaS offer per se is called Hosted Services and it does support SSL on custom domains.[1]
The price for the IaaS offer for a small Windows VM is $0.08/h ($57.60/month).[2] EC2 pricing for a small Windows VM is $0.115/h ($82.80/month).[3]
I tried out Joyent back when they did a year's free VM promotion with Facebook. I found Solaris to be very nice, but it required a little work to get software built for linux to build on it. Different C compiler for one thing. Usually for common pieces of software, like ruby there is a guide on the Joyent website for compiling.
I never used any cloud computing but hell that's slow. Just tried the "sort five million numbers" on my desktop and it took under 5 seconds (something above 5 in VM). I tried it in python though.. but I imagine ruby's performance would be about the same in this regard.
The `ruby` benchmark on my iMac i3 3GHZ took `14.5` seconds. Inside of a CentOS 5.3 x64 virtual machine on Softlayer with 2 cpus and 1GB of memory got `15.6` seconds. Not sure how Joyent and especially AWS performed so poorly on that benchmark.
"Amazon need to lift their game in terms of CPU performance. They offer a great service that obviously extends far beyond a simple CPU benchmark. But when you can get the same work done in Joyent significantly faster for the comparable price, you’ll get far more mileage per instance, which is ultimately going to save the dollars."
Or, "Wow, Amazon sure makes bank on this EC2 stuff, they are selling kilo-core-seconds for a bunch more than Joyent is and they are still the most widely used provider."
There have been a couple of good benchmarks around which instances get you the most 'bang' for the buck, but its pretty clear that Amazon is making a ton of money here.
A good comparison is a rack of 40 SuperMicro 1U dual socket servers with 192G of ram in them. You can create 480 dedicated instances with 16 GB of RAM each, and run the whole thing in a colo facility for $5K/month (power, networking, and floor space). That is 0.2 cents per second, or a bit more than $10/month running them 100% of the time.
Not a bad business to be in.