Folding Forum

Transferred from the thread on the backup Google group here:
http://groups.google.com/group/foldinga ... 81397d1f83

I think I'm almost done with my Folding@home on EC2 experiments. Thanks to theMASS for a ton of help in getting different approaches to work and in analyzing the results at 4am a couple nights.

My initial impression is that EC2 is probably not cost effective for Folding@home compared to alternatives like buying cheap hardware or (obviously) running it on spare cycles on systems you already use.

It has been, however, an interesting experiment in running applications on EC2 and prompted me to build a 64-bit large/extra large AMI for Ubuntu 7.10 Gutsy which should be released soon.

Tests:

For the first test, I used Christer Edwards' folding installer for Ubuntu:
http://ubuntu-tutorials.com/category/folding/
This seems to create a single worker (not sure of terminology here) for each CPU (though feedback from the author indicates it should be using the SMP client on multi-CPU systems).

The small instance costs $0.10/hour to run (1/2 of a dual core CPU). It started 29 Nov 06:22:21 and has completed 1200000 out of 1500000 steps (80%) as of 2 Dec 22:12:59. The live FAH log is available directly from the small folding instance while I leave it running:
http://ec2folding.azeelo.com/folding/FAHlog.txt
Browse around that server for other info including Munin charts, cpuinfo, and config.cfg.

The large instance costs $0.40/hour to run (2 dual core CPUs) and took about 24+ hours to complete its first 2 WUs (concurrently). The live FAH logs are available directly from the medium folding instance while I leave it running:
http://ec2folding-large.azeelo.com/fold ... FAHlog.txt
http://ec2folding-large.azeelo.com/fold ... FAHlog.txt

I tried a large instance using the SMP client following instructions here:
http://folding.stanford.edu/English/FAQ-SMP#ntoc29
I don't have the results handy, but they did not appear to be very good from discussions with theMASS.

I tried an xlarge instance using the non-SMP client for a while at $0.80/hour and its performance per WU seemed equivalent to the small instance.

The SMP client claims to be optimized for 4 cores, so I didn't try it on the 8 core xlarge instance.

Results:

If my calculations are correct, the large instance was folding at about $5/WU and the small and xlarge instances were about $11/WU. The difference between large and xlarge doesn't make any sense to me which leads me to...

Caveats:

- I just learned about Folding@home so I may not have installed and tweaked the configuration to get maximum performance out of EC2.

- I may not understand how to configure EC2 to get the maximum performance out of it, especially for this application, especially on 64-bit.

- The sample WUs that were assigned may have been unusually hard. Apparently not all WUs are created equal, and I'm not sure the best way to judge a platform's performance without running it for a long time (which gets expensive on EC2).

Followup:

Followup on the idle CPU part of the thread: I believe that the "top" program is reporting idle percentage incorrectly on the small instance type. Both vmstat and Munin reported that the CPU was completely in use with zero percent idle, while top consistently showed 25% idle.

Followup on "stolen" CPU cycles: The correct term is apparently "steal" not "stolen". Approximately 50% of the perceived dual-core CPU gets assigned to steal as the VM increases CPU usage on the small instance type. The large and xlarge instance types have 2 and 4 complete dual-core CPUs available so no steal shows up.

Offer:

I'd hate for my beginner attempts to be taken as the last word in EC2 folding, but since there seems to be nobody else trying it out or publishing the results I figured I should toss out what I had.

If anybody else would like to take a look at folding on EC2, I can provide AMIs, recipes, and advice for getting it running. If you are an expert with the folding software but don't want to learn how to use EC2, I can even fire up some instances on EC2 and give you ssh access to study them.

Best,
--
Eric Hammond
http://www.anvilon.com/

You're paying to fold !?!

toTOW wrote:You're paying to fold !?!

One way or another, you pay.

Yes but it's different if your machine is used for something else ...

Well so ... what PPD do you get on EC2 Is the PPD/$ interesting compared to a personal machine

As there was not clarity around the performance of the SMP folding client on Amazon EC2 large instances, I've kicked off a few that I'll run for a day or so. While they are running you can view the live log files and other information here:

http://ec2folding-large-smp-1.azeelo.com/
http://ec2folding-large-smp-2.azeelo.com/
http://ec2folding-large-smp-3.azeelo.com/

Each large instance has 2 dual-core CPUs (see the web servers for stats) running Ubuntu 7.10 Gutsy in a xen-based VM at a cost of $.40/hour.

The SMP client was installed using these instructions:
http://folding.stanford.edu/English/FAQ-SMP#ntoc29

You can see the client.cfg files on the web servers above.

I am so new to folding, I don't even know what "PPD" or "PPD/$" are but I'm running this so that folks who do know can figure that out. I don't plan to continue paying for folding after this experiment is over, but was doing this as a research project to see how this particular type of application fit Amazon EC2.

It seems folks in the folding community had similar questions about Amazon EC2, so we're bridging the gap and making a connection here.

In the off chance that anybody is willing to pay for folding and they don't know how to use Amazon EC2 themselves, I will make this offer:

I will run folding under your username/team on Amazon EC2 for as many concurrent large instances as you'd like for as long as you'd like, provided you fund the cost in advance. Given that this is helping science, I don't need any profit.

If your job is big enough and long enough, you'll probably find somebody on this board wiling to do it with purchased hardware at a cheaper cost, but one of the beauties of EC2 is that I can get instances fired up on a moments notice (and throw them away when no longer needed).

I am currently running a system at home that uses a T7200 mobile processor on a desktop board that is doing in the range of 1200-1400 PPD. The system costs me about $.50 a week in power to run. The cost was arround $350-$400 to build at that time. I hope this gives you a close comparison to what your currently running.

PPD means Points per day.

You get this value by dividing the credits for the WU by the time it take to process it.

PPD/$ is an indicator to seen how much point per day you'll do for one $ invested

If you have question or need help, ask us, we'll be pleased to help you

I made a little operation based on the cost you gave us (0.4$/hour) ... it makes 290$ per month. It's very expensive (it's more interesting to rent a dedicated web server)

Judging by the CPU used in this machines (cpuinfo of you smp instances), it won't be interesting for folding (only opterons @ 2 GHz) ...

If they used intel chips, it would be better

esh wrote: http://ec2folding-large-smp-1.azeelo.com/
http://ec2folding-large-smp-2.azeelo.com/
http://ec2folding-large-smp-3.azeelo.com/

These large instances running the SMP client have been terminated, but here are the relevant logs for folks who want to study them:

http://s3a.notlong.com/20071203-ec2/ec2 ... FAHlog.txt
http://s3a.notlong.com/20071203-ec2/ec2 ... FAHlog.txt
http://s3a.notlong.com/20071203-ec2/ec2 ... FAHlog.txt

http://s3a.notlong.com/20071203-ec2/ec2 ... puinfo.txt
http://s3a.notlong.com/20071203-ec2/ec2 ... client.cfg

Definitively not worth the money they cost

Renting a dedicated web server will be cheaper than buying time on the EC2

So, what would one pay for CPU cycles? Is Amazon the only one offering this type of service? What PPD/$ be an acceptable one? I think that maintaining a PPD of say a stock E6600 on full time, at a $.12 per KWh rate and a minor premium for network access, cooling, and hardware upgrades would be acceptable. Now if you had a business rate account where electric is about $.08 per KWh. This would be part of the profit/upgrade margin.

I really dont see why people would pay to fold. But im glad the OP has given some stats which will make people think before they consider doing things like this in the future.

@OP

Why not just buy a Quad Core PC and stick it in a corner running Linux SMP, and ignore it. It will generate large amounts of points for your username/team, and in the winter months will help keep that room nice and warm. Its a win/win situation. If heat is a problem, maybe stick it in the garage during winter (if you have one), or cellar, or loft.

Alternatively you could use the computer for general day to day tasks, and run say 4 UniProc clients on say Windows, or one VM with SMP on windows and 1 Uniproc (if you do things that eat up CPU time). There are many different configs for different tasks and loading requirements.

Reasons to pay vs. ownership of boxen:

1. Cost of power (someone paying a greater amount of cost per KWh.)
2. Limited computer building skills (unless buying OEM)
3. Limited space for systems.
4. Lives in a hot climate year round and doesn't need that extra heat.
5. Maintainance
6. Initial costs ($400 upfront and running costs, vs a steady $20 a month)
7. Depriciaton of equipment
8. time (monitoring, upgrading, maintaining, etc.)
9. Ease of use. (like a "buy it now" option)
10. Ability to expand and contract donation to the cause based on income during the year.
ETC...

I think that it is a bit odd, but some will find the benifits to it.

In the previous forum, we had a similar discussion about people who auctioned FAH points on ebay.

I though that before any of you decide to start advertising a service here with some statement like "I've got a computer that I'd be glad to rent for $xxxx which can earn yyy PPD for your account" I'd better remind you that we prohibit commercial posts and will delete it just like we're trying to do with the other spam. No selling allowed.

Nevertheless, I think my point is that with a little negotiation, you ought to be able to find the same thing cheaper.