If you need something for long-running, non- restartable jobs you dont use preemptible VMs Think answer is D.

Incorrect options are B. VerticalPodAutscaler scales PODS based on the app you deploy. For handle infrequently GPU access, you need infrequently GPU nodes VerticalAutscaler Pod deployed on a non GPU node it useless, [We cant have the node always have GPU for infrequent requests] C. Preemptible VMs cant last long D. For infrequent access, you don't want to have a permanent homogenous cluster. The correct option is "A" auto-provisioning = Attaches and deletes node pools to cluster based on the requirements. Hence creating a GPU node pool, and auto-scaling would be better https://cloud.google.com/kubernetes-engine/docs/how-to/node-auto-provisioning

Selected Answer: D

You are not able to add GPUs to existing node pools. This significantly impacts the viability of option A. My reasoning for D: A dedicated GPU node pool allows configuring those nodes with specific instance types, disk sizes, etc., ensuring the best fit for the long-running jobs. While it incurs some cost even with a minimum size of 1, it might still be more cost-efficient than full auto-provisioning if the jobs are infrequent but require a predictable baseline capacity. Separating GPU and non-GPU workloads can improve resource scheduling and prevent potential conflicts.

In my opinion, more sense has A, but then i read again and again the answer - "Enable node auto-provisioning" with GPU will not works due to limitation "You cannot add GPUs to existing node pools". If "A" was like "Create GPU node pool with enabled auto-provisioning" this will be correct answer, in in our case should be D

The most cost-effective option for your scenario would be **C. Create a node pool with preemptible VMs and GPUs attached to those VMs**. Preemptible VMs are Google Cloud's excess compute capacity. They are up to 80% cheaper than regular instances, making them a cost-effective choice for fault-tolerant workloads that do not require continuous availability³. However, please note that preemptible VMs are subject to availability and can be preempted if Google Cloud requires access to those resources, but they will be a good choice if the jobs can tolerate occasional preemptions³. While options A, B, and D could also be used in certain scenarios, they may not provide the same level of cost-effectiveness for long-running, non-restartable jobs that require GPUs⁵. Always consider the nature of your workloads and their tolerance for interruptions when choosing the right solution.

The ans is D,since they require gpu

The correct option is "D"

It is A Node auto-provisioning creates node pools based on the following information: CPU, memory, and ephemeral storage resource requests. GPU requests. Pending Pods' node affinities and label selectors. Pending Pods' node taints and tolerations. https://cloud.google.com/kubernetes-engine/docs/concepts/node-auto-provisioning

For me is A

Think answer is D.

Best option is A: https://cloud.google.com/kubernetes-engine/docs/concepts/node-auto-provisioning#how-it-works

I go with D

By default, GKE uses the E2 machine series unless any of the following conditions apply: The workload requests a feature that is not available in the E2 machine series. For example, if a GPU is requested by the workload, the N1 machine series is used for the new node pool. https://cloud.google.com/kubernetes-engine/docs/concepts/node-auto-provisioning

D. Create a node pool of instances with GPUs, and enable autoscaling on this node pool with a minimum size of 1.

Option A, which seems to be the most voted so far, is incorrect in my opinion. From Google's docs: "You cannot add GPUs to existing node pools." https://cloud.google.com/kubernetes-engine/docs/how-to/gpus#limitations So, answer A implies that GPUs are already used in the current cluster that needs to be autoscaled.

The correct answer should be A