Examkin

Soru

Zorluk: OrtaSelecting and Designing Compute and Application Platforms

A biomedical research enterprise needs to execute high-throughput genomic processing jobs on Google Cloud. The workload consists of non-HTTP containerized batch tasks that run between 15 and 45 minutes per execution, require GPU acceleration for parallel processing, and run unpredictably based on daily research submissions. The architecture team must select compute platforms that minimize operational management overhead while ensuring zero cost incurred when no jobs are executing. Which TWO compute architecture deployment strategies fulfill these requirements? (Select TWO.)

  1. Execute the containerized tasks using Cloud Batch configured with GPU-attached Spot VM instances that scale down to zero upon job completion.Cevap
  2. Deploy the processing workloads as Cloud Run jobs configured with GPU allocation and scale-to-zero execution limits.Cevap
  3. C
    Provision a persistent Google Kubernetes Engine (GKE) Standard cluster with a static multi-zone GPU node pool dedicated to queuing and running incoming jobs.
  4. D
    Purchase 3-year Committed Use Discounts (CUDs) for fixed Compute Engine GPU instance pools sized to handle maximum anticipated peak daily throughput.
  5. E
    Configure Compute Engine Managed Instance Groups (MIGs) containing GPU instances and set the autoscaler policy strictly to CPU utilization percentage.

Cevap

The optimal strategies are deploying the containerized workloads using Cloud Batch with GPU-attached Spot VMs and utilizing Cloud Run jobs configured with GPU resources.
For non-HTTP containerized workloads that run intermittently and require GPU acceleration, fully managed job-based serverless architectures such as Cloud Batch (with Spot VMs) and Cloud Run jobs offer optimal serverless execution. Both options eliminate baseline infrastructure costs by scaling down to zero when idle while taking on all infrastructure lifecycle management.

Adım Adım Çözüm

1
Analyze workload operational parameters and scaling requirements.
The workload is non-HTTP batch processing, GPU-dependent, long-running (15-45 minutes), and highly intermittent.
Serverless container options or dedicated batch schedulers that scale to zero and support GPUs are required to prevent paying for idle infrastructure.
2
Evaluate Google Cloud managed compute platforms against operational overhead limits.
Cloud Batch and Cloud Run jobs provide fully managed, serverless execution models for batch containers with native GPU support and zero baseline cost.
Both services manage underlying infrastructure lifecycles automatically, removing cluster administration tasks.
3
Eliminate options that introduce management overhead, idle cost, or improper scaling metrics.
Discard permanent GKE cluster setups, long-term CUD commitments for variable workloads, and CPU-only MIG autoscaling.
These alternatives increase operational complexity and introduce cost misallocations or improper autoscaling triggers.

Anahtar Kavram

Selecting Serverless Compute and Managed Batch Platforms for Intermittent GPU Workloads
Bu soruyu puanla

Konular

Selecting and Designing Compute and Application Platforms