Using Preemptible VMs and GPUs on GCP

Configuring preemptible VMs and GPUs for Kubeflow Pipelines on GCP

This document describes how to configure preemptible virtual machines(preemptible VMs)and GPUs on preemptible VM instances(preemptible GPUs)for your workflows running on Kubeflow Pipelines on Google Cloud Platform (GCP).

Introduction

Preemptible VMs are Compute Engine VMinstances that last a maximumof 24 hours and provide no availability guarantees. Thepricing of preemptible VMs islower than that of standard Compute Engine VMs.

GPUs attached to preemptible instances(preemptible GPUs)work like normal GPUs but persist only for the life of the instance.

Using preemptible VMs and GPUs can reduce costs on GCP.In addition to using preemptible VMs, your Google Kubernetes Engine (GKE)cluster can autoscale based on current workloads.

This guide assumes that you have already deployed Kubeflow Pipelines. If not,follow the guide to deploying Kubeflow on GCP.

Using preemptible VMs with Kubeflow Pipelines

In summary, the steps to schedule a pipeline to run on preemptibleVMs are asfollows:

  • Create anode poolin your cluster that contains preemptible VMs.
  • Configure your pipelines to run on the preemptible VMs. The following sections contain more detail about the above steps.

1. Create a node pool with preemptible VMs

Use the gcloud command tocreate a node pool.The following example includes placeholders to illustrate the importantconfigurations:

  1. gcloud container node-pools create PREEMPTIBLE_CPU_POOL \
  2. --cluster=CLUSTER_NAME \
  3. --enable-autoscaling --max-nodes=MAX_NODES --min-nodes=MIN_NODES \
  4. --preemptible \
  5. --node-taints=preemptible=true:NoSchedule \
  6. --service-account=DEPLOYMENT_NAME-vm@PROJECT_NAME.iam.gserviceaccount.com

Where:

  • PREEMPTIBLE_CPU_POOL is the name of the node pool.
  • CLUSTER_NAME is the name of the GKE cluster.
  • MAX_NODES and MIN_NODES are the maximum and minimum number of nodesfor the GKEautoscalingfunctionality.
  • DEPLOYMENT_NAME is the name of your Kubeflow deployment. If you usedthe CLI to deploy Kubeflow,this name is the value of the ${KFAPP} environment variable. If you usedthe deployment UI,this name is the value you specified as the deployment name.
  • PROJECT_NAME is the name of your GCP project.

Below is an example of the command:

  1. gcloud container node-pools create preemptible-cpu-pool \
  2. --cluster=user-4-18 \
  3. --enable-autoscaling --max-nodes=4 --min-nodes=0 \
  4. --preemptible \
  5. --node-taints=preemptible=true:NoSchedule \
  6. --service-account=user-4-18-vm@ml-pipeline-project.iam.gserviceaccount.com

2. Schedule your pipeline to run on the preemptible VMs

After configuring a node pool with preemptible VMs, you must configure yourpipelines to run on the preemptible VMs.

In the DSL code foryour pipeline, add the following to the ContainerOp instance:

  1. .apply(gcp.use_preemptible_nodepool())

The above function works for both methods of generating the ContainerOp:

Note:

  • Call .set_retry(#NUM_RETRY) on your ContainerOp to retrythe task after the task is preempted.
  • If you modified thenode taintwhen creating the node pool, pass the same node toleration to theuse_preemptible_nodepool() function.

For example:

  1. import kfp.dsl as dsl
  2. import kfp.gcp as gcp
  3. class FlipCoinOp(dsl.ContainerOp):
  4. """Flip a coin and output heads or tails randomly."""
  5. def __init__(self):
  6. super(FlipCoinOp, self).__init__(
  7. name='Flip',
  8. image='python:alpine3.6',
  9. command=['sh', '-c'],
  10. arguments=['python -c "import random; result = \'heads\' if random.randint(0,1) == 0 '
  11. 'else \'tails\'; print(result)" | tee /tmp/output'],
  12. file_outputs={'output': '/tmp/output'})
  13. @dsl.pipeline(
  14. name='pipeline flip coin',
  15. description='shows how to use dsl.Condition.'
  16. )
  17. def flipcoin():
  18. flip = FlipCoinOp().apply(gcp.use_preemptible_nodepool())
  19. if __name__ == '__main__':
  20. import kfp.compiler as compiler
  21. compiler.Compiler().compile(flipcoin, __file__ + '.zip')

Using preemptible GPUs with Kubeflow Pipelines

This guide assumes that you have already deployed Kubeflow Pipelines. Insummary, the steps to schedule a pipeline to run withpreemptible GPUsare as follows:

  • Make sure you have enough GPU quota.
  • Create a node pool in your GKE cluster that contains preemptible VMs withpreemptible GPUs.
  • Configure your pipelines to run on the preemptible VMs with preemptibleGPUs. The following sections contain more detail about the above steps.

1. Make sure you have enough GPU quota

Add GPU quota to your GCP project. The GCPdocumentation liststhe availability of GPUs across regions. To check the available quota forresources in your project, go to theQuotas page in the GCPConsole.

2. Create a node pool of preemptible VMs with preemptible GPUs

Use the gcloud command tocreate a node pool.The following example includes placeholders to illustrate the importantconfigurations:

  1. gcloud container node-pools create PREEMPTIBLE_GPU_POOL \
  2. --cluster=CLUSTER_NAME \
  3. --enable-autoscaling --max-nodes=MAX_NODES --min-nodes=MIN_NODES \
  4. --preemptible \
  5. --node-taints=preemptible=true:NoSchedule \
  6. --service-account=DEPLOYMENT_NAME-vm@PROJECT_NAME.iam.gserviceaccount.com \
  7. --accelerator=type=GPU_TYPE,count=GPU_COUNT

Where:

  • PREEMPTIBLE_GPU_POOL is the name of the node pool.
  • CLUSTER_NAME is the name of the GKE cluster.
  • MAX_NODES and MIN_NODES are the maximum and minimum number of nodesfor theGKE autoscalingfunctionality.
  • DEPLOYMENT_NAME is the name of your Kubeflow deployment. If you usedthe CLI to deploy Kubeflow,this name is the value of the ${KFAPP} environment variable. If you usedthe deployment UI,this name is the value you specified as the deployment name.
  • PROJECT_NAME is the name of your GCP project.
  • GPU_TYPE is the type ofGPU.
  • GPU_COUNT is the number of GPUs.

Below is an example of the command:

  1. gcloud container node-pools create preemptible-gpu-pool \
  2. --cluster=user-4-18 \
  3. --enable-autoscaling --max-nodes=4 --min-nodes=0 \
  4. --preemptible \
  5. --node-taints=preemptible=true:NoSchedule \
  6. --service-account=user-4-18-vm@ml-pipeline-project.iam.gserviceaccount.com \
  7. --accelerator=type=nvidia-tesla-t4,count=2

3. Schedule your pipeline to run on the preemptible VMs with preemptible GPUs

In the DSL code foryour pipeline, add the following to the ContainerOp instance:

  1. .apply(gcp.use_preemptible_nodepool()

The above function works for both methods of generating the ContainerOp:

Note:

  • Call .set_gpu_limit(#NUM_GPUs, GPU_VENDOR) on yourContainerOp to specify the GPU limit (for example, 1) and vendor (forexample, 'nvidia').
  • Call .set_retry(#NUM_RETRY) on your ContainerOp to retrythe task after the task is preempted.
  • If you modified thenode taintwhen creating the node pool, pass the same node toleration to theuse_preemptible_nodepool() function.

For example:

  1. import kfp.dsl as dsl
  2. import kfp.gcp as gcp
  3. class FlipCoinOp(dsl.ContainerOp):
  4. """Flip a coin and output heads or tails randomly."""
  5. def __init__(self):
  6. super(FlipCoinOp, self).__init__(
  7. name='Flip',
  8. image='python:alpine3.6',
  9. command=['sh', '-c'],
  10. arguments=['python -c "import random; result = \'heads\' if random.randint(0,1) == 0 '
  11. 'else \'tails\'; print(result)" | tee /tmp/output'],
  12. file_outputs={'output': '/tmp/output'})
  13. @dsl.pipeline(
  14. name='pipeline flip coin',
  15. description='shows how to use dsl.Condition.'
  16. )
  17. def flipcoin():
  18. flip = FlipCoinOp().set_gpu_limit(1, 'nvidia').apply(gcp.use_preemptible_nodepool())
  19. if __name__ == '__main__':
  20. import kfp.compiler as compiler
  21. compiler.Compiler().compile(flipcoin, __file__ + '.zip')

Comparison with Cloud AI Platform Training service

Cloud AI Platform Training is a GCPmachine learning (ML) training service that supports distributed training andhyperparameter tuning, and requires no complex GKE configuration. Cloud AIPlatform Training charges the Compute Engine costs only for the runtime of thejob.

The table below compares Cloud AI Platform Training with Kubeflow Pipelinesrunning preemptible VMs or GPUs:

Cloud AI Platform Training Kubeflow Pipelines with preemption
Configuration No GKE configuration Requires GKE configuration
Cost Compute Engine costs for the job lifetime Lower price with preemptible VMs/GPUs/TPUs
Accelerator Supports various VM types, GPUs, and CPUs Support various VM types, GPUs, and CPUs
Scalability Automates resource provisioning and supports distributed training Requires manual configuration such as GKE autoscaler and distributed training workflow
Features Out-of-box support for hyperparameter tuning Do-it-yourself hyperparameter tuning with Katib

Next steps