Workload Service Selection Flex Start Compute Engine

What Flex-start is

Flex-start is a Google Cloud compute consumption and provisioning model designed for short-duration workloads (lasting up to seven days) that require compute resources and accelerators like GPUs. Under this model, Compute Engine schedules and provisions the requested compute instances as soon as capacity becomes available in the selected zone, providing access without requiring manual reservation negotiations.

How it meets the requirements

• Execution Duration: Flex-start instances support run durations of up to seven days, perfectly accommodating jobs that require 4 to 6 days of continuous runtime.
• Fault Tolerance: Unlike Spot instances, Flex-start VMs are non-preemptible during their active run duration, guaranteeing that multi-day fine-tuning or inference tasks run without unexpected termination.
• Scheduling Flexibility: Flex-start operates on a best-effort, delayed-start scheduling queue, which fits workloads that can start whenever capacity is available.
• Cost Optimization: The provisioning model provides substantial discounts of up to 53% for vCPUs, memory, and GPUs compared to standard on-demand pricing on a pay-as-you-go (PAYG) basis, without requiring long-term upfront commitments.

Key advantages and capabilities

• Pay-as-you-go discounting: Delivers up to a 53% discount while billing only for the actual run duration of the provisioned instances.
• Orchestration integration: Flex-start natively integrates with Google Kubernetes Engine (GKE) and Compute Engine Managed Instance Groups (MIGs) using resize requests.
• Densely allocated hardware: Ensures optimal interconnectivity and low network latency for compute-intensive accelerator configurations.
• Zero over-provisioning waste: Automatically terminates instances at the conclusion of the specified run duration.

Why this is optimal

Flex-start directly matches the combination of flexible start times, multi-day non-preemptible execution, and pay-as-you-go cost reduction, outperforming Spot VMs (which risk preemption) and fixed calendar reservations (which charge for idle reservation capacity).