EC2 TESLA GPU instances

Associate
Joined
27 Nov 2011
Posts
41
Hello!

So I just joined these forums (should have done earlier tbh, been buying rigs and kit here since 2001).

I noticed the DC projects forums so decided id join up and say hi (im doing Einstein @ home)
my comps - http://einstein.phys.uwm.edu/hosts_user.php?userid=572821

Also wanted to ask if anyone had tried using EC2 instances to do crunching.

I currently have a quad XL instance with 2 TESLA M2050 GPU's and 16 processors. My problem so far, I have been unable to setup windows server 2008 with an appropriate set of CUDA drivers (despite amazon saying the instance comes with the drivers, the GPU is not recognised by BOINC). The BOINC project page shows me as having the 2 TESLA GPU's.

Just wondering if anyone else had tried EC2 GPU for crunching and how they got it tor work.

At current costs, it is $2.60 an hour to run this instance ad-hoc on EC2 so quite pricey if you want to do it for a while (lucky for me im not footing the bill). Its cheaper to reserve an instance annually if you need one.

/edit - Halz can you do an E@H update please?
 
Welcome to the team. :)

Don't have any experience of your problem personally so I'm afraid I can't help. Have you tried asking on the Einstein or Boinc boards?
 
First question, does Einstein@Home even support them? It's definitely worth asking 'over there'.

Also, really silly question, but does EC2 support CUDA? (I know how Azure works but don't know much about EC2)

And finally, ask around if there's something in Server 2008 that might prevent BOINC from using a GPU. There's definitely the 'protected mode' stuff, which is a no-no, but being a server OS means Windows is configured not to use 3D by default.
 
Last edited:
yeah the Quad XL GPU instance has win server 2008 R2 with CUDA toolkit and drivers installed, i figured it would work off the bat but no such luck.

Despite frequent googling, I have not been able to find a 'CUDA for nubs' guide on how to do this on EC2, I did find a post by a German guy who said he'd done it but not how.
 
How do you access the EC2 server? Is it by RDP? If so then no cuda apps will work, you will have to use VNC or something that doesn't load the Terminal server gfx driver.
 
How do you access the EC2 server? Is it by RDP? If so then no cuda apps will work, you will have to use VNC or something that doesn't load the Terminal server gfx driver.

I was using RDP yeah, ok I will install vnc on it and try that. I very interested to see how these TESLA units perform vs other GPU's in the nvidia line.

The 8800 GT in one of my machines does a binary pulsar search in about 1h 30m, not sure how fast the newer nvida gpus go.
 
Biffa, good catch re: RDP.

Beware VNC's hopeless security (unless your compute instance is behind some other firewall or there's been a miracle and someone has actually secured it). BOINC can do remote administration which might be a reasonable choice (BOINC isn't secure either, but nonetheless is somewhat less likely than VNC to open the system up to being rooted - EC2 is a popular target for hackers).
 
Last edited:
The other option is if there is some sort of EC2 console interface, like there is with VMware. Or use something like PCAnywhere
 
yeah, security is pretty good, you setup the firewall when you build the instance, by default there is only ssh / RDP enabled (linux / windows) but you can dynamically change this as the instance is running.

All the instances in the US data centres seem to come out of the same ip range in Washington state so probably quite easy to target 100's of them with a bit of research and a port scan.
 
If you poke a hole in the firewall to allow VNC through then that means there's no firewall. :)

But unless it's doing something else critical then the advantage of using compute nodes is that it's easy to tear one down and start another.
 
Nothing critical and no sensitive personal/company data etc, just figured since the company I work for wants me to test various different instances I would put them to good use.
 
Put VNC listening on a wierd port, and put an ACL on the firewall to only allow you IP through and you will lower the chances. Certainly its no worse than it being open to RDP
 
Back
Top Bottom