--- layout: default title: GPU acceleration parent: Model-serving framework nav_order: 150 --- # GPU acceleration GPU acceleration is an experimental feature. For updates on the progress of GPU acceleration, or if you want to leave feedback that could help improve the feature, join the discussion in the [OpenSearch forum](https://forum.opensearch.org/). {: .warning} When running a natural language processing (NLP) model in your OpenSearch cluster with a machine learning (ML) node, you can achieve better performance on the ML node using graphics processing unit (GPU) acceleration. GPUs can work in tandem with the CPU of your cluster to speed up the model upload and training. ## Supported GPUs Currently, ML nodes following GPU instances: - [NVIDIA instances with CUDA 11.6](https://aws.amazon.com/nvidia/) - [AWS Inferentia](https://aws.amazon.com/machine-learning/inferentia/) If you need GPU power, you can provision GPU instances through [Amazon Elastic Compute Cloud (Amazon EC2)](https://aws.amazon.com/ec2/). For more information on how to provision a GPU instance, see [Recommended GPU Instances](https://docs.aws.amazon.com/dlami/latest/devguide/gpu.html). ## Supported images You can use GPU acceleration with both [Docker images](https://gitlab.com/nvidia/container-images/cuda/blob/master/doc/supported-tags.md) with CUDA 11.6 and [Amazon Machine Images (AMIs)](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/AMIs.html). ## PyTorch GPU-accelerated ML nodes require [PyTorch](https://pytorch.org/docs/stable/index.html) 1.12.1 work with ML models. ## Setting up a GPU-accelerated ML node Depending on the GPU, you can provision a GPU-accelerated ML node manually or by using automated initialization scripts. ### Preparing an NVIDIA ML node NVIDIA uses CUDA to increase node performance. In order to take advantage of CUDA, you need to make sure that your drivers include the `nvidia-uvm` kernel inside the `/dev` directory. To check for the kernel, enter `ls -al /dev | grep nvidia-uvm`. If the `nvidia-uvm` kernel does not exist, run `nvidia-uvm-init.sh`: ``` #!/bin/bash ## Script to initialize nvidia device nodes. ## https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#runfile-verifications /sbin/modprobe nvidia if [ "$?" -eq 0 ]; then # Count the number of NVIDIA controllers found. NVDEVS=`lspci | grep -i NVIDIA` N3D=`echo "$NVDEVS" | grep "3D controller" | wc -l` NVGA=`echo "$NVDEVS" | grep "VGA compatible controller" | wc -l` N=`expr $N3D + $NVGA - 1` for i in `seq 0 $N`; do mknod -m 666 /dev/nvidia$i c 195 $i done mknod -m 666 /dev/nvidiactl c 195 255 else exit 1 fi /sbin/modprobe nvidia-uvm if [ "$?" -eq 0 ]; then # Find out the major device number used by the nvidia-uvm driver D=`grep nvidia-uvm /proc/devices | awk '{print $1}'` mknod -m 666 /dev/nvidia-uvm c $D 0 mknod -m 666 /dev/nvidia-uvm-tools c $D 0 else exit 1 fi ``` After verifying that `nvidia-uvm` exists under `/dev`, you can start OpenSearch inside your cluster. ### Preparing AWS Inferentia ML node Depending on the Linux operating system running on AWS Inferentia, you can use the following commands and scripts to provision an ML node and run OpenSearch inside your cluster. To start, [download and install OpenSearch]({{site.url}}{{site.baseurl}}/install-and-configure/index/) on your cluster. Then export OpenSearch and set up your environment variables. This example exports OpenSearch into the directory `opensearch-2.5.0`, so `OPENSEARCH_HOME` = `opensearch-2.5.0`: ``` echo "export OPENSEARCH_HOME=~/opensearch-2.5.0" | tee -a ~/.bash_profile echo "export PYTORCH_VERSION=1.12.1" | tee -a ~/.bash_profile source ~/.bash_profile ``` Next, create a shell script file called `prepare_torch_neuron.sh`. You can copy and customize one of the following examples based on your Linux operating system: - [Ubuntu 20.04](#ubuntu-2004) - [Amazon Linux 2](#amazon-linux-2) After you've run the scripts, exit your current terminal and open a new terminal to start OpenSearch. GPU acceleration has only been tested on Ubuntu 20.04 and Amazon Linux 2. However, you can use other Linux operating systems. {: .note} #### Ubuntu 20.04 ``` . /etc/os-release sudo tee /etc/apt/sources.list.d/neuron.list > /dev/null <=2MB echo "-Xss2m" | tee -a $OPENSEARCH_HOME/config/jvm.options # Increase max file descriptors to 65535 echo "$(whoami) - nofile 65535" | sudo tee -a /etc/security/limits.conf # max virtual memory areas vm.max_map_count to 262144 sudo sysctl -w vm.max_map_count=262144 ``` #### Amazon Linux 2 ``` # Configure Linux for Neuron repository updates sudo tee /etc/yum.repos.d/neuron.repo > /dev/null <=2MB echo "-Xss2m" | tee -a $OPENSEARCH_HOME/config/jvm.options # Increase max file descriptors to 65535 echo "$(whoami) - nofile 65535" | sudo tee -a /etc/security/limits.conf # max virtual memory areas vm.max_map_count to 262144 sudo sysctl -w vm.max_map_count=262144 ``` When the script completes running, open a new terminal for the settings to take effect. Then, start OpenSearch. OpenSearch should now be running inside your GPU-accelerated cluster. However, if any errors occur during provisioning, you can install the GPU accelerator drivers manually. #### Prepare ML node manually If the previous two scripts do not provision your GPU-accelerated node properly, you can install the drivers for AWS Inferentia manually: 1. Deploy an AWS accelerator instance based on your chosen Linux operating system. For instructions, see [Deploy on AWS accelerator instance](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/frameworks/torch/torch-neuron/setup/pytorch-install.html#deploy-on-aws-ml-accelerator-instance). 2. Copy the Neuron library into OpenSearch. The following command uses a directory named `opensearch-2.5.0`: ``` OPENSEARCH_HOME=~/opensearch-2.5.0 ``` 3. Set the `PYTORCH_EXTRA_LIBRARY_PATH` path. In this example, we create a `pytorch` virtual environment in the OPENSEARCH_HOME folder: ``` PYTORCH_NEURON_LIB_PATH=~/pytorch_venv/lib/python3.7/site-packages/torch_neuron/lib/ mkdir -p $OPENSEARCH_HOME/lib/torch_neuron; cp -r $PYTORCH_NEURON_LIB_PATH/ $OPENSEARCH_HOME/lib/torch_neuron export PYTORCH_EXTRA_LIBRARY_PATH=$OPENSEARCH_HOME/lib/torch_neuron/lib/libtorchneuron.so ``` 4. (Optional) To monitor the GPU usage of your accelerator instance, install [Neuron tools](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/tools/index.html), which allows models to be used inside your instance: ``` # Install Neuron Tools sudo apt-get install aws-neuronx-tools -y ``` ``` # Add Neuron tools your PATH export PATH=/opt/aws/neuron/bin:$PATH ``` ``` # Test Neuron tools neuron-top ``` 5. To make sure you have enough memory to upload a model, increase the JVM stack size to `>+2MB`: ``` echo "-Xss2m" | sudo tee -a $OPENSEARCH_HOME/config/jvm.options ``` 6. Start OpenSearch. ## Troubleshooting Due to the amount of data required to work with ML models, you might encounter the following `max file descriptors` or `vm.max_map_count` errors when trying to run OpenSearch in a your cluster: ``` [1]: max file descriptors [8192] for opensearch process is too low, increase to at least [65535] [2]: max virtual memory areas vm.max_map_count [65530] is too low, increase to at least [262144] ``` To troubleshoot the max file descriptors error, run the following command: ``` echo "$(whoami) - nofile 65535" | sudo tee -a /etc/security/limits.conf ``` To fix the `vm.max_map_count` error, run this command to increase the count to `262114`: ``` sudo sysctl -w vm.max_map_count=262144 ``` ## Next steps If you want to try a GPU-accelerated cluster using AWS Inferentia with a pretrained HuggingFace model, see [Compiling and Deploying HuggingFace Pretrained BERT](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/bert_tutorial/tutorial_pretrained_bert.html).