mfaktc for dummies

NBtarheel_33 · 2011-09-27, 09:16

So, I have become interested in dropping a few bucks and trying out the GPU clusters over on Amazon EC2 (you get 8 2.66-GHz Xeon cores IIRC and *two* Fermi GPUs - 448 cores each...insert Homer Simpson drooling noise...for $2.10/hour guaranteed or something like 80 cents an hour floating price).

Anyway, I think I just successfully donated four bucks or so, trying to get mfaktc up and running. I got the executable copied over, but when I went to run it, I kept getting an error that the file "libcudart.so.3" was nowhere to be found. Didn't get a single number tested. Do any of you mfaktc gurus know what this error message means, and how I might get around it? As far as I know (and can tell), the latest CUDA toolkit (4.0.17) is installed on the Amazon system.

Any thoughts are greatly appreciated. And if I were to get something like this up and running, how many GHz-days per hour could I reasonably expect to pull in?

Ralf Recker · 2011-09-27, 10:25

Quote:

Originally Posted by NBtarheel_33

I got the executable copied over, but when I went to run it, I kept getting an error that the file "libcudart.so.3" was nowhere to be found.

It seems that your binary was compiled with an older 3.x version of the CUDA toolkit.

Quote:

Originally Posted by NBtarheel_33

As far as I know (and can tell), the latest CUDA toolkit (4.0.17) is installed on the Amazon system.

Quick and dirty compilation instructions (CUDA Toolkit installed to /usr/local/cuda, /usr/local/cuda/bin/ not in the $PATH, your mileage may vary):

Code:

1. Download and unpack mfaktc-0.17.tar.gz
2. cd mfaktc-0.17/src
3. PATH=/usr/local/cuda/bin:$PATH make
4. cd ..

TheJudger · 2011-09-27, 12:15

Hi,

build from source or precompiled binary?
Did you try this one: http://mersenneforum.org/showpost.ph...postcount=1175

Oliver

NBtarheel_33 · 2011-09-29, 04:41

I tried it out last night, but on a Windows machine - it's easier (and quicker, which matters when you're renting the machine by the hour!) to set things up. I played with a 2x Quad-Core 2.93 GHz Nehalem with 2 NVIDIA Fermi GPUs.

On the GPU, I was getting through 65-bit assignments in the 292M range in about 40-45 seconds! They take 12-13 minutes on my 2007 Core2 Duo, and 18-20 minutes on my 2006 Pentium 4. I ran one instance of mfaktc on each GPU, and I also played around with P-1 on the eight Nehalem cores (with hyperthreading, so you're actually getting 16 threads). Stage 1 on a 50M exponent, using all eight cores/16 threads, looks to need about 3-4 hours! The system had 23GB of RAM, so Stage 2 of P-1 would be interesting. There are other non-GPU systems available with as much as 68GB of RAM - I wonder what would happen if I gave all of that to a single P-1... Hmm...

Didn't try CUDALucas yet. It's probably not economically feasible at $2+ per hour to try to run an entire LL.

The system cost $2.94 per hour, so for two TFs every 45 seconds, that's a cost of 1.84 cents per TF, with just the GPUs running. So, say it's 1.5 cents per TF with the CPU and GPU - that means it would cost around $345 to clear 292M-293M. We should probably add to the "You're Addicted to GIMPS When..." thread - "You know you're addicted to GIMPS when you rent high-performance computing clusters to process your assignments" ... LOL.

Dubslow · 2011-09-29, 05:06

As for hyperthreading, set each worker to run two threads, and P95 should automatically assign the helper threads to each hyperthread to maximize efficiency. That should also get assignments through quicker (which helps here) while getting the same overall throughput.

nucleon · 2011-10-02, 08:58

At $2/hr, you could of bought a GTX560Ti card for that after a week. (or even less depending on pricing in your local area)

A GTX560Ti needs about 4-5GHz* of CPU to saturate it.

A saturated GTX560Ti delivers about 120+ GHzdays/day* of TF output.

In that week, the rented machine may have done 3500GHz.days work.

So after a month, you'd be way ahead. Of course this assumes, you have 4-5GHz of local CPU grunt you can run 24*7.

-- Craig Meyers
*Figures are indicative only, your mileage may vary.

Christenson · 2011-10-06, 03:54

Don't forget to factor in electricity and a couple hundred bucks for the case with the motherboard and CPU on it.

NBtarheel_33 · 2011-10-11, 06:17

Quote:

Originally Posted by nucleon

At $2/hr, you could of bought a GTX560Ti card for that after a week. (or even less depending on pricing in your local area)

A GTX560Ti needs about 4-5GHz* of CPU to saturate it.

A saturated GTX560Ti delivers about 120+ GHzdays/day* of TF output.

In that week, the rented machine may have done 3500GHz.days work.

So after a month, you'd be way ahead. Of course this assumes, you have 4-5GHz of local CPU grunt you can run 24*7.

-- Craig Meyers
*Figures are indicative only, your mileage may vary.

Yeah, the only problem is that my 2006-vintage Pentium 4 Prescott would merely "dampen" that card, rather than "saturate" it. (Assuming that the magic smoke just didn't up and totally leave upon plugging the card into such an antiquated box.)

NBtarheel_33 · 2011-10-11, 06:21

Quote:

Originally Posted by Christenson

Don't forget to factor in electricity and a couple hundred bucks for the case with the motherboard and CPU on it.

I'm planning to do a build (my first) probably next year. At this point, I figured I'd just wait for Ivy Bridge to come out. Definitely want to include a couple of good GPUs.

Dubslow · 2011-10-12, 05:14

Ivy Bridge...
I can already feel my tastebuds moistening...

Christenson · 2011-10-13, 00:04

Quote:

Originally Posted by NBtarheel_33

I'm planning to do a build (my first) probably next year. At this point, I figured I'd just wait for Ivy Bridge to come out. Definitely want to include a couple of good GPUs.

The biggest problem with my first build was the bent-over pin 997 on the CPU socket...everything else was straightforward...Between Batalov and Lavalamp, I got a good system put together, the only regret was too little power supply to run a better GPU!

2011-09-27, 09:16	#1
NBtarheel_33 "Nathan" Jul 2008 Maryland, USA 5×223 Posts	mfaktc for dummies So, I have become interested in dropping a few bucks and trying out the GPU clusters over on Amazon EC2 (you get 8 2.66-GHz Xeon cores IIRC and two Fermi GPUs - 448 cores each...insert Homer Simpson drooling noise...for $2.10/hour guaranteed or something like 80 cents an hour floating price). Anyway, I think I just successfully donated four bucks or so, trying to get mfaktc up and running. I got the executable copied over, but when I went to run it, I kept getting an error that the file "libcudart.so.3" was nowhere to be found. Didn't get a single number tested. Do any of you mfaktc gurus know what this error message means, and how I might get around it? As far as I know (and can tell), the latest CUDA toolkit (4.0.17) is installed on the Amazon system. Any thoughts are greatly appreciated. And if I were to get something like this up and running, how many GHz-days per hour could I reasonably expect to pull in?

2011-09-27, 12:15	#3
TheJudger "Oliver" Mar 2005 Germany 5·223 Posts	Hi, build from source or precompiled binary? Did you try this one: http://mersenneforum.org/showpost.ph...postcount=1175 Oliver Last fiddled with by TheJudger on 2011-09-27 at 12:15

2011-09-29, 04:41	#4
NBtarheel_33 "Nathan" Jul 2008 Maryland, USA 45B₁₆ Posts	Thanks for the help I tried it out last night, but on a Windows machine - it's easier (and quicker, which matters when you're renting the machine by the hour!) to set things up. I played with a 2x Quad-Core 2.93 GHz Nehalem with 2 NVIDIA Fermi GPUs. On the GPU, I was getting through 65-bit assignments in the 292M range in about 40-45 seconds! They take 12-13 minutes on my 2007 Core2 Duo, and 18-20 minutes on my 2006 Pentium 4. I ran one instance of mfaktc on each GPU, and I also played around with P-1 on the eight Nehalem cores (with hyperthreading, so you're actually getting 16 threads). Stage 1 on a 50M exponent, using all eight cores/16 threads, looks to need about 3-4 hours! The system had 23GB of RAM, so Stage 2 of P-1 would be interesting. There are other non-GPU systems available with as much as 68GB of RAM - I wonder what would happen if I gave all of that to a single P-1... Hmm... Didn't try CUDALucas yet. It's probably not economically feasible at $2+ per hour to try to run an entire LL. The system cost $2.94 per hour, so for two TFs every 45 seconds, that's a cost of 1.84 cents per TF, with just the GPUs running. So, say it's 1.5 cents per TF with the CPU and GPU - that means it would cost around $345 to clear 292M-293M. We should probably add to the "You're Addicted to GIMPS When..." thread - "You know you're addicted to GIMPS when you rent high-performance computing clusters to process your assignments" ... LOL.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
mfaktc not starting in Mac OSX	bayanne	GPU Computing	0	2014-05-10 14:38
mfaktc on a Mac	bayanne	GPU Computing	0	2013-10-18 09:59
mfaktc (0.20)	fairsky	Software	9	2013-09-24 12:58
Politics on the Internet for Dummies	jasong	jasong	3	2013-09-19 05:05
mfaktc	tichy	GPU Computing	4	2010-12-03 21:51

2011-09-29, 05:06	#5
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3·29·83 Posts	As for hyperthreading, set each worker to run two threads, and P95 should automatically assign the helper threads to each hyperthread to maximize efficiency. That should also get assignments through quicker (which helps here) while getting the same overall throughput.

2011-10-02, 08:58	#6
nucleon Mar 2003 Melbourne 5×103 Posts	At $2/hr, you could of bought a GTX560Ti card for that after a week. (or even less depending on pricing in your local area) A GTX560Ti needs about 4-5GHz* of CPU to saturate it. A saturated GTX560Ti delivers about 120+ GHzdays/day* of TF output. In that week, the rented machine may have done 3500GHz.days work. So after a month, you'd be way ahead. Of course this assumes, you have 4-5GHz of local CPU grunt you can run 247. -- Craig Meyers Figures are indicative only, your mileage may vary.

2011-10-06, 03:54	#7
Christenson Dec 2010 Monticello 5×359 Posts	Don't forget to factor in electricity and a couple hundred bucks for the case with the motherboard and CPU on it.

2011-10-12, 05:14	#10
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3×29×83 Posts	Ivy Bridge... I can already feel my tastebuds moistening...