Run on AWS

In this tutorial, we will go through how to allocate a CPU/GPU instance in AWS and build it from scratch step-by-step.

Select and run an instance

We first need an AWS account, and then go the EC2 console after login in.

[Optional] You can select a region on the right upper corner that is close to your location to reduce the network latency. But also note that some regions may not have GPUs instances.

Then click "launch instance" to select the operation system and instance type.

We selected "Ubuntu 16.06":

and "p2.xlarge", which contains a single Nvidia K80 GPU. Note that there is a large number of instance, refer to ec2instances.info for detailed configurations and fees.

Note that we need to check the instance limits to guarantee that we can request the resource. If running out of limits, we can request more capacity by clicking the right link, which often takes about a single workday to process.

On the next step we increased the disk from 8 GB to 40 GB so we have enough space to install CUDA and store a reasonable size dataset. For large-scale datasets, we can "add new volume". Also you selected a very powerful GPU instance such as "p3.8xlarge", make sure you selected "Provisioned IOPS" in the volume type for better I/O performance.

Then we launched with other options as the default values. The last step before launching is choosing the ssh key, you may need to generate and store a key if you don't have one before.

After clicked "launch instances", we can check the status by clicking the instance ID link.

Once the status is green, we can right-click and select "connect" to get the access instruction.

For example, here we ssh to the instance with the provided address:

Install packages

Let's first update and install basic packages after we sshed to the instance.

sudo apt-get update && sudo apt-get install -y build-essential python-pip

Install CUDA

If you launched a CPU-only instance, you can skip this step.

Let go to Nvidia website to download a recent CUDA and install it by following its instructions.

For example, we can download CUDA 9.1 and install it:

wget https://developer.nvidia.com/compute/cuda/9.1/Prod/local_installers/cuda_9.1.85_387.26_linux
sudo sh cuda_9.1.85_387.26_linux

We need to answer a few questions during the installation.

accept/decline/quit: accept
Install NVIDIA Accelerated Graphics Driver for Linux-x86_64 387.26?
(y)es/(n)o/(q)uit: y
Do you want to install the OpenGL libraries?
(y)es/(n)o/(q)uit [ default is yes ]: y
Do you want to run nvidia-xconfig?
(y)es/(n)o/(q)uit [ default is no ]: n
Install the CUDA 9.1 Toolkit?
(y)es/(n)o/(q)uit: y
Enter Toolkit Location
 [ default is /usr/local/cuda-9.0 ]:
Do you want to install a symbolic link at /usr/local/cuda?
(y)es/(n)o/(q)uit: y
Install the CUDA 8.0 Samples?
(y)es/(n)o/(q)uit: n

After installation, we can check the GPU status by

nvidia-smi

Finally, add the CUDA library path into system:

echo "export LD_LIBRARY_PATH=\${LD_LIBRARY_PATH}:/usr/local/cuda/lib64" >>.bashrc

Install MXNet

Since we install CUDA 9.1, we can just install mxnet-cu91. Here we used the --pre flag to install the latest release.

pip install --pre mxnet-cu91

Once it's done, run python in terminal, and try to import mxnet for a sanity check.

import mxnet