SwishDL

Description

This repo contains scripts and instructions to create your own Cloud Layer for efficiently running and scaling your DL training jobs.

Recommended setup:

1. Use sharded versions of CIFAR10, and ImageNet that are already available at below locations.
   • CIFAR10: http://storage.googleapis.com/lpr-demo
   • ImageNet: http://storage.googleapis.com/lpr-imagenet
2. Create kubernetes cluster to meet your needs - GPU type (K80, P100, V100), Number of GPUs per node (1, 2, 4, 8).
3. Create a docker image of your trainer.
4. Use deploy script and launch your trainer (explained later).

Overview of code structure:

• dltrainer Contains Dockerfile and kubernetes deoployment file for PyTorch trainer.

• cache-server Contains Dockerfile and kubernetes deployment file for NGINX cache-server.

• ku script: Contains commands to setup kubernetes cluster on Google Cloud

• kuber-cluster-config.sh script: Contains parameters that can be configured to customize the kubernetes setup

Initial setup

Before proceeding, please ensure you have following packages installed locally; follow instructions available online.

1. Docker
2. Google Cloud SDK: After installing Google Cloud SDK, run gcloud init

Kubernetes cluster setup:

1. Set parameters in kube-cluster-config.sh
2. Call ./ku init. Please don't kill the execution inbetween. The command does the following:
   • Creates a kubernetes clusters
   • Creates a GPU node-pool with each node containing requested number of GPUs
   • Install cache-server and deploys it into the cluster

Deploying PyTorch trainer on Kubernetes cluster:

Change to dltrainer folder.

1. Three important files are
   • train.py: PyTorch trainer
   • model.py: Neural Network model
   • dataset.py: Dataset loader. This file also serves as an example to fetch data using dlinputs library.
2. Build the docker image
   docker build -t name_of_your_docker_image .
3. Tag docker image
   docker tag name_of_your_docker_image gcr.io/$GCLOUD_PROJECT_NAME/name_of_your_docker_image:v1
4. Upload docker image to a cloud repository.
   gcloud docker -- push gcr.io/$GCLOUD_PROJECT_NAME/name_of_your_docker_image:v1
5. Configure location of your image in trainer-deploy.yml

spec:
    template:
        spec:
          containers:
          - name: imagenet-training
            image: gcr.io/$GCLOUD_PROJECT_NAME/name_of_your_docker_image:v1

6. Configure other parameters required for your training job in trainer-deploy.yml - type of GPUs, number of GPUs.

apiVersion: batch/v1
kind: Job
metadata:
  name: trainer-job
spec:
  template:
    spec:
      containers:
      - name: imagenet-training
        image: gcr.io/$GCLOUD_PROJECT_NAME/name_of_your_docker_image:v1 # TODO Put location of your image on cloud repository
        command: ["python"]
        args:
        - "train.py"
        - "--devices"
        - "1" # TODO Set the number of GPUs required by your Job        
        resources:          
          limits:
            nvidia.com/gpu: 1 # TODO Set this number to same as the number of GPUs required by your Job
      nodeSelector:
        cloud.google.com/gke-accelerator: nvidia-tesla-k80 
      restartPolicy: Never
  backoffLimit: 4

7. Deploy your job
   kubectl create -f trainer-job.yml

Profiler Dashboard

The base docker image also comes with a built-in profiler, which tracks GPU, CPU and Network Utilization every 30seconds.

Below is a sample dashboard from training a network across 4 GPUs.