An AI platform engineering team is designing a provisioning workflow on Google Kubernetes Engine (GKE) for large-scale, distributed training jobs that require multi-host TPU/GPU slices. The workload has the following operational requirements and constraints: Queuing & Atomic Scheduling: The batch jobs can tolerate waiting in a queue until all required accelerator nodes are available simultaneously (all-or-nothing provisioning). Interruption Resistance: Once scheduled, training runs can execute for up to seven days without experiencing preemption. Cost Optimization: The team needs discounted compute pricing without committing to one-year or multi-year capacity reservations. Which accelerator provisioning strategy should the team implement?

Configure GKE accelerator node pools using Dynamic Workload Scheduler (DWS) in Flex-start provisioning mode.: ### What Dynamic Workload Scheduler (DWS) Flex-start Is **Dynamic Workload Scheduler (DWS)** in **Flex-start** mode is a capacity provisioning mechanism designed specifically for AI/ML and batch computing workloads requiring scarce accelerators like **GPUs** and **TPUs**. Instead of failing immediately when resources are temporarily constrained, DWS queues resource requests and provisions the entire requested capacity simultaneously once available. ### How It Meets the Scenario Requirements * **Queuing & Atomic Scheduling**: DWS Flex-start automatically queues requests and guarantees an **all-or-nothing** atomic allocation. It ensures that multi-node distributed training jobs do not suffer from partial provisioning deadlocks where some worker nodes start while others remain pending. * **Interruption Resistance**: Once Flex-start successfully provisions the requested VM and accelerator capacity, the instances run without the threat of **Spot preemption** for a bounded runtime duration of up to **seven days**. * **Cost Optimization**: Flex-start provides discounted pricing compared to standard on-demand instances because workloads allow Google Cloud scheduling flexibility, all without requiring upfront financial lock-in. ### Key Advantages and Capabilities * **Gang/Atomic Allocation**: Eliminates idle compute charges that occur when distributed jobs wait for missing nodes in a slice. * **Bounded Execution Windows**: Allows batch training jobs to specify run durations up to 7 days with guaranteed non-preemptible execution for the allocated window. * **Native GKE Integration**: Integrates directly with Kubernetes Job APIs and Queued Resource controllers for seamless scheduling orchestration. * **Reduced Stockout Failures**: Converts immediate API quota/capacity rejections into orderly asynchronous queue fulfillment. ### Why This Is Optimal DWS Flex-start perfectly balances **cost efficiency**, **batch queuing**, and **run stability**, making it the recommended GCP architecture for high-performance accelerator workloads that need uninterrupted multi-day runs without fixed multi-year reservations.

Dynamic Workload Scheduler Flex Start Provisioning | Professional Cloud DevOps Engineer

Dynamic Workload Scheduler Flex Start Provisioning

Worked example. The correct answer is already marked and every option is explained below, so there is nothing to select here. To answer questions yourself, start the free trial.

An AI platform engineering team is designing a provisioning workflow on Google Kubernetes Engine (GKE) for large-scale, distributed training jobs that require multi-host TPU/GPU slices. The workload has the following operational requirements and constraints:

Queuing & Atomic Scheduling: The batch jobs can tolerate waiting in a queue until all required accelerator nodes are available simultaneously (all-or-nothing provisioning).
Interruption Resistance: Once scheduled, training runs can execute for up to seven days without experiencing preemption.
Cost Optimization: The team needs discounted compute pricing without committing to one-year or multi-year capacity reservations.

Which accelerator provisioning strategy should the team implement?

Configure GKE accelerator node pools using Dynamic Workload Scheduler (DWS) in Flex-start provisioning mode.

Purchase one-year or three-year TPU and GPU committed capacity reservations across all target zones.

Configure standard on-demand GKE node pools with Horizontal Pod Autoscaling and Cluster Autoscaler.

Provision GKE node pools backed by Spot VMs and manage worker restarts using Kubernetes Job retries.

Professional Cloud DevOps Engineer

Dynamic Workload Scheduler Flex Start Provisioning

Related questions