[Amazon EC2] for only 45c/hr you get this!

#1

I was researching Amazons EC2 program, and had stumbled upon something interesting. With their ‘Spot Request’ system you can get an instance of cc2.8xlarge (88 compute units) for only 45cents/hr. It stays that price for the most part, spiking to 50cents once a week if that.
I tried using hamachi to create a vpn with my own workstation, and it works with one instance running, but trying to add another from the same AMI doesnt work (still trying to figure out a fix). Important part is creating it as a Windows Server 2008 r2 for clusters. After you configure it you can actually use it on any of their instance types. I am lacking in the whole backburner department so I haven’t played with that, but running .bat files is super fast and slick! There are of course data transfer fees but they are negligible for an individual.

As far as I know, this is WAY cheaper and just as fast as the point and click renderfarms. Need it done in an hour? Make a dozen of them, it’s literally two clicks once youve created the AMI.
Ask any questions if you want, I’m still learning as I go.

#2

Nice, Ive been pondering setting up a machine or two for rendering, let us know how you get on.

#3

this might benefit you some… this is to part 3 of his series on setting up EC2 and backburner.

http://vimeo.com/27994567

#4

interested in your results…will be following this.

#5

I’m assuming software that requires a license per render machine is incapable of utilizing cloud services like this?

Makes me almost want to jump from maya to 3dsmax so I get unlimited MR licenses…stupid maya

#6

Okay, for the last 4 hours I have been fighting through buggy non-responsive web-interfaces, failed connection, webpages that never loads and password decryption that doesn’t work. And so far, I haven’t had the change to see a single instance desktop. That’s laughable(in a sad and disappointing way) in my opinion.

#7

Sorry to hear that, read this and go through it step by step and if it still isn’t working I dont know what to say. They provide an image you can use, without having to worry about password decryption and such. It’s for C4D but if you get it working you can just install Maya or whatever, change the password to your own and save it as a new image.
http://cosmocyte.com/animation-techniques/

#8

Hey man,

Wow! This looks super cool. I’ve read a lot about this for rendering, but I’ve never seen an example of someone who got it to work.

How does it compare to RR or Rebus price wise? Do the licences install ok? Can you think of this like your renting a beastly dual xeon by the hour?

Thanks,
AJ

#9

I got around it. had to use internet explorer, apparently they don’t support Chrome or Firefox.

Chrome works until you press a link “instances” and then the script dies. and you cannot browse to other functions. and it cannot decrypt passwords.

Firefox just returns HTTP Status 400 - BadRequest. upon login.

Internet Explorer works, however you constantly get a warning box suggesting to close a faulty script.Very very shaky interface!

Thanx for the links cdenny! :slight_smile:

I think I know a way to automate the IP setup and login to render manager.
so you only have to prepare the manager, launch the render instances and press render.

#10

Good to hear! Interesting tidbit about the browser, I’ve been using Chrome and haven’t had issues with it, I’ve had no ‘script’ stuff popup. I’m using the latest version on windows 7 pro x64. As for the ip dealio, I’d love to hear your solution, the only thing I know of, to keep a static ip for the manager, is to create and associate an elastic ip to the manager so it stays the same. this costs $.005 for every hour it isnt associated with an instance, and data out costs 1c/gb of data to the other instances. It works out to be only $3.50 just sitting there.

I was messing with license stuff last night, and I’m not sure how its working so far. I’ve been using the trial for licensing, but it counts down just like any other (I tried turning off internet time update thinking it would just revert, but it doesn’t). I’m still looking into how that all works; the only issue is with creating the job and using the manager, the other instances dont need to mess with it because Maya allows up to 6-8 nodes with mental ray. I’m going to try to create a vpn and use my actual workstation as the manager because I have a static ip, but then the flow of data will be limited to around 300kb/s which is just too low in my opinion. If you have good internet this may not be an issue. One other annoying issue is the naming, I checked the box that says ‘use ip <xxx>’ for dynamic naming, but the computer name is still just amazon-xxxx or whatever.

Compared to rebusfarm, it’s much more complicated, may take more time just to set it up and stuff, but its much cheaper. Based on 3.9c/hr/ghz, a single cc2.8xlarge instance with maya’s 45c/hr rate would be equivalent to paying 3.45/hr (about ten times the cost). There are other costs with data usage, etc, but that is nearly negligible at 12cents/gb. As a student this is awesome, but if your a busy professional or such, going with a dedicated service like rebusfarm may still be a better option, because the time you spend setting it up and making it work and dealing with issues might not be worth it. Zync render utilizes Amazon EC2 farm, and while I don’t have access to it, I hope their rates are better than rebus farm. I was thinking about upgrading or building a better computer, but I think I’ll stick with this EC2 method for now because there is no upfront cost ( which I cant afford anyway).
EDIT: One thing to note is that if you used a comparable machine locally you’d have to pay electricity for it, which would be about 1000watts on 100% load. Around here it’s 11c/kwh, so just consider that a saving as well.
I plan on just setting up the renders and then creating .bat files to run on two instances, with the files being synced by dropbox. This may not work with the licensing, I’ll find out soon enough.
Good luck!

#11

Oh dear… I found the issue! It was Avast webRep a browser plugin.
Now chrome works just fine… Sry, I never thought a simple thing like webRep would screw it up that bad. :smiley:

In regards to the idea i had it has become quite a complicated mess with wbscripts, batch files and Unix commandline tools… And I’m not quite there yet :argh:
I may look into writing a program/tool instead, but this will be some other time.

In regards to licenses: I have not setup licenses so far(no maya install). just playing with Backburner and the integration between the Instances. But I am working on a way for the Instances to derive there “Backburner” configurations from the manager server when they are launched. Might be able to port that function to the license part. At least for floating licenses. That i cannot test from home, since I only have single license here.

#12

Just a little update:
I built another one from scratch because, no matter what I did, changing the computer to ‘dynamic naming’ in the EC2Config Settings the name would stay the same. This caused problems. I ran into trouble with the elastic ip. Backburner connected, shared drives were visible etc, but I kept getting an “Error Exit on 211”. This appeared to be the slave not ‘finding’ the files, which made no sense, I guess I just dont know how to use it or something.

One other problem, using \servername\blah to connect wouldn’t work on anything other than the host (where you could just do \localhost anyway). The private ip worked, but that isn’t static. The elastic ip worked on clients, but the host couldn’t “map to network drive” it’s own shared folder using the elastic ip, it would just pop up a login window where you couldn’t log in. I ditched that and downloaded Deadline because you can have two free slaves, and like I said way earlier, these cc2.8xlarge instances are beastly. First ami I built, the name doesn’t change (it’s borked) so that is the ‘manager’, it can be its own slave so I just set the repository to \servername\DeadlineRepository . I “released” the elastic ip because I can take the five minutes to manually network the shared drive and connect the second instance to the ‘manager’, it’s literally just two numbers, it’s nice that \servername\blah works on the host and \ip address\blah on the slave. If using the instance as both a manager and a slave is horrible I’ll just get the elastic ip back and make the ‘manager’ instance a cheap low powered 5c/hr computer. I’m loving Deadline so far, even though I have no clue how to use it for the most part.

I can’t stress how awesome and cheap this is, people were raving over how awesome it was when they were using the normal instances with 26cpu units, and not these compute clusters with 88.5 cpu units for 20c more. If I could only figure out the whole networking, folder access part of this, you could easily create a small personal, on demand, renderfarm that is 10x faster than an overclocked 3930k rig for $3.60/hr, but with no overhead. As a college student on a tight budget this makes me happy. Am I unclear on anything? I feel like I’ve explained how it works pretty well, a lot of it is just trying over and over to get it right. I’ve only spent $1.41 setting this up so far, ha.

#13

did you go through Judd’s entire vimeo series on setting up backburner in EC2?.. i know its’ a bit slow, but he does a pretty thorough explanation of everything.

#14

do me a favour
send a job to EC2 and send the same files to Rebusfarm
(register and you get 10 points free at Rebus automaticly)

lets compare the prices and speed Rebus vs EC2
I doubt your math “about 10 times more” is correct.
I am curriouse to see the result !

Andre
Rebusfarm.net

#15

Well considering I used your calculator it’s probably pretty accurate. (this is only true when you use the cc2.8xlarge instance, which cause problems because it’s ‘cluster edition’)
BUT
Screw this, it just doesn’t work, I’ve spent hours messing with this shit and it just isn’t worth it. It was a fun-ish experiment that just failed, just bill your clients for using rebus or something. I got it all networking correctly using a VPC and EIP, BUT THERE’S ALWAYS A PROBLEM, like backburner telling me it rendered 20 frames in 6 seconds, but the frames don’t exist, and believe me, I checked the settings a thousand times. Just buy another freaking computer or use rebus. I have wasted too much time for too little return. Also, with the .10cents/gb transfer rate I’ve realized those costs can start to add up when you have thousand of files and stuff.
Good luck if you try to make it work, but I’m done. The only thing I’m going to try is using Deadlines 2 node edition, or just use bat rendering, I’m done trying to make backburner work with amazon.

#16

The thing with amazon is you have to supply your own software license, install and setup the OS, and get the VM’s set up to network render. Pro renderfarms like Rebus and renderrocket supply the software for you and already have all the renderfarm network management working for you so that should rightfully increase the cost of using them – along with the fact that you don’t have to spend 4 hours fiddling around learning how to get amazon to render your particular software.

I think it’d be a pain to set up the VM’s as their own self contained renderfarm on a short notice, and even more of a pain if not impossible to work out the firewall, security, and network address issues to somehow tie it into your current local renderfarm or access your license server.

If you’re using a 3d app that allows infinite rendernodes for free or if you want to attempt to set up and install throw-away trial editions of software that’ll expire after 30 days, go for it. You have nothing to lose but time if it doesn’t work.

On the flipside, if you’re a working professional who needs a reliable bulletproof solution, then I’d still opt to use an established pro dedicated renderfarm and charge the client for the cost difference or pay it yourself if the rush time management problem was your fault.

Even if it’s a rare occasion where you need a huge job rendered by tomorrow - I’d be even more inclined to pay for something that will absolutely get the job done in time without spending hours setting up cloud VM’s that might end up just rendering out blank frames and leaving you with only 2/3 of the day left to render the files.

and I really do wish some of these render engine companies would update their licensing to not charge per machine, but rather charge for specific rendering features, speed optimizations, access to weekly beta builds with bug fixes, and different levels of tech support, etc.

Small studios often can’t realistically afford to use some of the computationally expensive render settings. It doesn’t do much good to have access to these high end features if they’re too slow to actually use in production because you don’t have enough machines. I think it’s rather ironic and the wrong way to go about charging for render licensing.

#17

Hi all.

No,no, it works.

I made money with AWS EC2.
I ran 20 cc2.8xlarge in Spot Mode in a VPC.
My clients earn hours.

I’m working on Blender + Cycles in GPU.
There’s a good strart.

Best Regards. (sorry for my poor english).

PhiKa.

#18

Reawakening this thread as I’ve found very little on the topic of actual EC2 render performance (3dsmax) until we did some testing ourselves to better understand if it’d be something investing in over buying more kit in the office - here’s what we found (based on a very basic setup, and not being particularly scientific so please consider this more anecdotal!)

Amazon Setup
1x Backburner 2012 running on m1.large (single socket 4 core xeon 2.5ghz)
4x 3DS Max 2013 running on CC2.8xlarge (dual socket 8 core xeon 2.6ghz)

In reality, the cpu specifications don’t give the performance we originally expected. I’m not entirely sure why, but I’m sure hypervisors, some shared resources and other hidden stuff is getting in the way.

Compared with local machines we have, these are rough rendertimes for a test BB job:

Amazon EC2 CC2.8xlarge (XeonE5 2x8 cores @ 2.6GHz : 41.6GHz) 5 mins
Local Rendernode (SandyBridge i7 6 cores @ 3.2 : 19.2GHz) 6.5 mins
Local Workstation (Xeon E3 4 cores @ 3.5GHz : 14GHz) 10 mins

We were hoping to see the Amazon instances to be knocking frames out in 1.5x - 1.8x faster than a local renderbox can smash them so were a little disappointed when they were coming in at only 1.2x considering they supposedly had twice the raw available GHz. There were very little maps being used in the scene, and actual scene translation time was fairly minimal (precached geometry, no FG/GI, no proxys) and kept to around 15 secs. Once rendering the machines were using 100% of the available CPU resources (according to Task Manager).

In total it takes about 30 minutes to spin all 5 machines up from cold and get a file rendering via backburner. We’ve not yet got a successful VPN from an EC2 instance to our studio so having to transfer files off the EC2 instances via a webdav link so the bottleneck in our experience is with our download speed. In tests, we can shift approx 1GB a minute off the EC2 onto other cloud storage. note: Don’t use Dropbox - it’s hobbled to 400KBPs (it seems in the UK).

In summary - it could be very useful when needing extra juice during a panic (until we expand our local render kit). Probably works out to about £8/hour to employ 4 extra machines with minimal setup fuss. Downside is performance ain’t all what we expected/hoped. Our studio internet connection is the biggest immediate problem (a fixed always on EFM to the green box sadly limited to 4mbps up/down for nearly £6K/year - thanks BT for never digging up the road around here!). Plugins/node licensing is a big unknown - will try krakatoa or Fume next time we need it and see what headaches that creates.

The 3ds max file we were testing with:
150meg layout file referencing a 120 meg XrefScene file
100 meg of bitmaps
Mental Ray, ray tracing, cached geometry, no render elements, no other caches/mr proxys
3840x1080 resolution, framebuffer off during render
No plugins or fancy stuff

Hope some of this was useful for other EC2 dabblers…

#19

http://vimeo.com/27993626
http://vimeo.com/27994510
http://vimeo.com/27994567
http://vimeo.com/27996749
http://vimeo.com/27996798

#20

I realize this is an old thread but while we’re on the subject again I’d like to add something. A few applications come with support for rendering on Amazon EC2 out of the box. For example in Houdini you don’t have to do anything special and they have machine images already setup for you. They had this back in 2009.

http://www.sidefx.com/index.php?option=com_content&task=view&id=1630

Like sentry66 said, Amazon is cheaper for a reason. If you’re using software that requires licenses per machine those costs are already covered with other render farm services and they have a pipeline and queue manager already setup. There’s also other charges to consider with Amazon like bandwidth and storage which are billed separately.

This can be a slippery slope, cloud computing in general can be very expensive compared to purchasing the equipment and software outright. If you need 50 machines for one day a year then cloud computing is a good way to go. If you need 50 machines five days a week all year long then cloud computing is a horrible idea and will end up costing many times more than setting up your own farm. It takes only a few months of regular usage for the costs of cloud computing to exceed that of purchasing outright (YMMV).