Create a GCE instance with 8 V100 GPUs

create_gce_instance.sh

export INSTANCE_NAME="xla-benchmark-8xV100"
export IMAGE_FAMILY="tf-1-12-cu100"
export PROJECT_NAME="<your project name>"

gcloud beta compute instances create $INSTANCE_NAME \
  --project=$PROJECT_NAME \
  --machine-type=n1-standard-64 \
  --maintenance-policy=TERMINATE \
  --accelerator=type=nvidia-tesla-v100,count=8 \
  --tags=http-server,https-server \
  --image-family=$IMAGE_FAMILY \
  --image-project=deeplearning-platform-release \
  --boot-disk-size=100GB \
  --boot-disk-type=pd-ssd \
  --local-ssd interface=nvme \
  --local-ssd interface=nvme \
  --local-ssd interface=nvme \
  --local-ssd interface=nvme \
  --metadata install-nvidia-driver=True

## Combines the 4 local nvme SSD drives into a single RAID 0 drive.
# Install raid management tool.
sudo apt-get update && sudo apt-get install mdadm --no-install-recommends

# Creates RAID 0 array.
sudo mdadm --create /dev/md0 --level=0 --raid-devices=4 \
/dev/nvme0n1 /dev/nvme0n2 /dev/nvme0n3 /dev/nvme0n4 

# Formats and mounts the array.
sudo mkfs.ext4 -F /dev/md0
sudo mkdir -p /data/imagenet
sudo mount /dev/md0 /data
sudo chmod a+w /data

# Installs custom TensorFlow 1.12 binary with AVX2.  Binary included on
# the image already has XLA but the custom binary is compiled with AVX2.
sudo pip install --force-reinstall https://storage.googleapis.com/tf-performance/tf_binary/tensorflow-1.12.0.a6d8ffa.AVX2.CUDA10-cp27-cp27mu-linux_x86_64.whl