I'll post this howto here, because it could be useful for documentation:
1. It could help people configure other (similar) systems, and
2. It's not worth it. The Amazon servers cost more than the bitcoin earnings.
Anyway, here's what to do to get maximum juice out of one of these servers. They have 2 Tesla C2050 GPU cards, and 16x CPU cores. With these instructions, you'll run
poclbm on each GPU, and jgarzik's
cpuminer on all processors, all mining for deepbit.net.
First, start up an Amazon EC2 GPU (cg1.4xlarge) instance with Amazon's own cluster Linux distribution (I used ami-321eed5b). When it's up and running, log on and run the following commands. Cut'n'paste is fine but
remember to fix username/password/pooled mining server.
# set up path stuff
echo >> $HOME/.bash_profile
echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/tools/lib' >> $HOME/.bash_profile
echo 'export PATH=$PATH:$HOME/tools/bin' >> $HOME/.bash_profile
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$HOME/tools/lib
export PATH=$PATH:$HOME/tools/bin
cpus=$(cat /proc/cpuinfo | grep ^processor | wc -l)
# initial package config
sudo yum -y groupinstall "Development Tools"
sudo yum -y install git libcurl-devel python-devel screen rsync
# install yasm
git clone git://github.com/yasm/yasm.git
cd yasm
./autogen.sh
./configure --prefix=$HOME/tools
make -j $cpus
make install
cd -
# install and start cpuminer
git clone https://github.com/jgarzik/cpuminer.git
cd cpuminer
./autogen.sh
./configure
make -j $cpus
screen -d -m ./minerd --threads $cpus --algo sse2_64 --url http://deepbit.net:8332/ --userpass YOUR_EMAIL:YOUR_PASSWORD
cd -
# install numpy
git clone git://github.com/numpy/numpy.git numpy
cd numpy
git checkout remotes/origin/maintenance/1.6.x
sudo python setup.py install
cd -
# set up newer nvidia library
wget http://developer.download.nvidia.com/compute/cuda/3_2_prod/drivers/devdriver_3.2_linux_64_260.19.26.run
wget http://developer.download.nvidia.com/compute/cuda/3_2_prod/toolkit/cudatoolkit_3.2.16_linux_64_fedora13.run
sudo mv -v /lib/modules/$(uname -r)/kernel/drivers/video/nvidia.ko /root/
At this point, you need to reboot your server. The easiest way to do that is to run
sudo reboot. Log back in to the server after reboot, and continue. You'll need to interact with the NVIDIA installers, so you can't cut'n'paste everything here. You'll also need to edit the siteconf.py file in the command
vi siteconf.py and make sure it says
CL_ENABLE_DEVICE_FISSION = Falsecpus=$(cat /proc/cpuinfo | grep ^processor | wc -l)
# restart cpuminer
cd cpuminer
screen -d -m ./minerd --threads $cpus --algo sse2_64 --url http://deepbit.net:8332/ --userpass YOUR_EMAIL:YOUR_PASSWORD
cd -
sudo bash devdriver_3.2_linux_64_260.19.26.run
sudo bash cudatoolkit_3.2.16_linux_64_fedora13.run
# install pyopencl
git clone http://git.tiker.net/trees/pyopencl.git
cd pyopencl
sudo easy_install Mako
git submodule init
git submodule update
python configure.py --cl-inc-dir=/usr/local/cuda/include --cl-lib-dir=/usr/local/cuda/lib64
vi siteconf.py # set CL_ENABLE_DEVICE_FISSION = False
sudo make install
cd -
# get poclbm and start for each device
git clone https://github.com/m0mchil/poclbm.git
cd poclbm/
screen -d -m python poclbm.py -o deepbit.net -p 8332 -u YOUR_EMAIL --pass=YOUR_PASSWORD -v -w 256 --device 0
screen -d -m python poclbm.py -o deepbit.net -p 8332 -u YOUR_EMAIL --pass=YOUR_PASSWORD -v -w 256 --device 1
If you're not familiar with the screen program, it runs your programs on a virtual console. You can access them with
screen -r <id>, and you release them again with the key-combo
Ctrl-a Ctrl-d.
Now you should be running 2 poclbm.py instances and one instance of cpuminer with 16 threads. In my test, each GPU calculated ~75000khash/s, while each CPU core did approx 1400khash/s, for a grand total of approximately 170mhash/s.
...and that's not worth it given an instance price of $2.10 per hour.
