Here is a quick introduction to using preemptible VMs in Google Kubernetes Engine. Our GCP Support team is here to lend a hand with your queries and issues.

Did you know that a preemptible VM is actually a finite Compute Engine virtual machine instance? It lasts up to 24 hours and offers no availability guarantee. However, they are priced lower than Compute Engine VMs.

Their low cost makes them a feasible choice for quick batch jobs, testing, and application that are fault tolerant. Since the instance terminates after a fixed time, it works well for systems or applications that distribute processes across several instances in a cluster. In this article, we’ll look at how to use preemptible VMs to run fault-tolerant workloads in Google Kubernetes Engine.

How to use Preemptible VMs in Google Kubernetes Engine

The GPUs that are attached to preemptible instances function like normal GPUs but only during the duration of the instance.

Besides low cost, preemptible VMs enable Google Kubernetes Engine Cluster to autoscale easily according to the current workloads.

We can easily create a cluster with preemptible VMs via Google cloud with this command after replacing ClusterName with the actual cluster name:

gcloud container clusters create ClusterName \
--preemptible

Similarly, we can create a node pool with preemptible VMs with this command after replacing PoolName with the actual pool name:

gcloud container node-pools create PoolName \
--cluster=CLUSTER_NAME \
--preemptible

Here are some of the other tasks we can carry out on preemptible VMs:

• Schedule pods with nodeSelector

  This involves scheduling pods on preemptible VMs by using nodeSelector in the deployments.
• Add a node taint to a node pool that uses preemptible VMs

  Nodes using preemptible VMs have to be tainted in order for GKE to place Pods with matching toleration on those nodes. This is where the —node-taints flag comes in handy.

  Once the flag is added, only pods that can tolerate the node taint get scheduled to the node.

Our experts would like to point out that we have to create at least an additional node pool in the cluster that does not use preemptible VMs ahead of adding a GPU node pool that uses preemptible VMs. This ensures the Google Kubernetes Engine can place system components safely. If you are looking for help with preemptible VMs, drop us a comment.

[Need assistance with a different issue? Our team is just a click away.]

Conclusion

To sum up, our Support Engineers introduced us to preemptible VMs and how to use them in Google Kubernetes Engine.