Professional Cloud ArchitectDesign and plan a cloud solution architectureHard
A research institution processes highly sensitive genomic data and uses custom-built, open-source machine learning models. They need a compute environment that provides fine-grained control over the underlying infrastructure, including specific GPU types and custom kernel modules. The workloads are long-running and require consistent performance, but the institution also wants to optimize costs by leveraging committed use discounts. Which Google Cloud compute option should you recommend?
- ACloud Run with custom container images
- BCompute Engine with custom machine types and GPUs
- CCloud Functions with custom runtimes
- DGKE Autopilot with GPU enabled
Show answer & explanationAnswer & explanation
Correct answer: B. Compute Engine with custom machine types and GPUs
Compute Engine provides the most granular control over VM instances, allowing selection of specific GPU types and installation of custom kernel modules. This level of customization, combined with committed use discounts for cost optimization, makes it ideal for specialized research workloads.
Why the other options are wrong
- A. Cloud Run is a serverless platform for stateless containers; it does not offer fine-grained control over specific GPU types or kernel modules and is not typically used for long-running, custom-kernel-dependent ML workloads.
- C. Cloud Functions are event-driven, short-lived, and don't provide the necessary control over infrastructure like specific GPUs or custom kernel modules for complex ML models.
- D. GKE Autopilot manages the underlying nodes, abstracting away control over specific GPU types and kernel modules, which is contrary to the 'fine-grained control' requirement.
Compute Engine with Customization
Google Cloud's Infrastructure-as-a-Service (IaaS) offering, allowing users to create and run virtual machines with extensive customization options for machine types, GPUs, storage, and operating systems.
- Provides the highest level of control over underlying infrastructure.
- Supports various GPU types and custom kernel modules for specialized workloads.
- Offers committed use discounts for significant cost savings on stable workloads.
- Requires manual management of VMs and operating systems.
Memory trick: Compute Engine delivers precise control for custom research.