GeePS is a parameter server library that scales single-machine GPU machine learning applications (such as Caffe) to a cluster of machines.
Run the following command to download GeePS and (our slightly modified) Caffe:
git clone --recurse-submodules https://github.com/cuihenggang/geeps.git
If you use the Ubuntu 14.04 system, you can run the following commands (from geeps root directory) to install the dependencies:
./scripts/install-geeps-deps-ubuntu14.sh
./scripts/install-caffe-deps-ubuntu14.sh
Also, please make sure your CUDA library is installed in /usr/local/cuda.
After installing the dependencies, you can build GeePS by simply running this command from geeps root directory:
scons -j8
You can then build (our slightly modified) Caffe by first entering the apps/caffe directory and then running make -j8:
cd apps/caffe
make -j8
You can run Caffe distributedly across a cluster of machines with GeePS. In this section, we will show you the steps to run Caffe's CIFAR-10 example on two machines.
All commands in this section are executed from the apps/caffe directory:
cd apps/caffe
You will first need to prepare a machine file as examples/cifar10/2parts/machinefile, with each line being the host name of one machine. Since we use two machines in this example, this machine file should have two lines, such as:
host0
host1
We will use pdsh to launch commands on those machines with the ssh protocol, so please make sure that you can ssh to those machines without password.
When you have your machine file in ready, you can run the following command to download and prepare the CIFAR-10 dataset:
./data/cifar10/get_cifar10.sh
./examples/cifar10/2parts/create_cifar10_pdsh.sh
Our script will partition the datasets into two parts, one for each machine. You can then train an Inception network on it with this command:
./examples/cifar10/2parts/train_inception.sh
Please look at our wiki for more details. Happy training!
Henggang Cui, Hao Zhang, Gregory R. Ganger, Phillip B. Gibbons, and Eric P. Xing. GeePS: Scalable Deep Learning on Distributed GPUs with a GPU-Specialized Parameter Server. In ACM European Conference on Computer Systems, 2016 (EuroSys'16)