MLS-C01 Question 285
Select 3You are managing an ML training workflow on Amazon SageMaker using GPU instances. After analyzing your usage, you notice that the GPU utilization is consistently below 30%, leading to higher costs without improved performance. Which actions should you take to rightsize your resources and optimize costs while maintaining the functionality of your workflow?
- A
Switch to a smaller instance type with fewer GPUs, such as moving from ml.p3.8xlarge to ml.p3.2xlarge.
- B
Enable Amazon SageMaker Managed Spot Training to take advantage of spot instances for cost savings.
- C
Reduce the volume size of the EBS storage attached to the instance to lower storage costs.
- D
Switch to a CPU-only instance for training, such as ml.c5.xlarge, to eliminate GPU costs.
- E
Analyze the training script for inefficiencies and optimize it to utilize the GPU more effectively.
Show answer and explanation
Correct answers: A, B, E
Explanation
In this scenario, the main issue is the low GPU utilization during ML training, leading to unnecessary costs. Rightsizing the instance (e.g., switching to a smaller GPU instance) and leveraging SageMaker Managed Spot Training are effective strategies to reduce costs while maintaining functionality. Additionally, optimizing the training script to better utilize the GPU can improve performance and cost efficiency. Reducing EBS storage and switching to a CPU-only instance do not appropriately address the problem.
- A. Correct.
Switching to a smaller instance type with fewer GPUs is a valid strategy to rightsize resources, especially when GPU utilization is low. This can help reduce costs while still providing enough capacity for the workload.
- B. Correct.
Using SageMaker Managed Spot Training allows you to take advantage of spot instances, which can significantly reduce costs compared to on-demand instances. This is a recommended cost-optimization strategy.
- C. Incorrect.
Reducing the volume size of EBS storage affects storage costs, but it does not address the core issue of low GPU utilization and is not directly related to rightsizing compute resources.
- D. Incorrect.
Switching to a CPU-only instance may not be appropriate for ML training workflows that require GPU acceleration. This could degrade performance and is not a viable option in this scenario.
- E. Correct.
Optimizing the training script can help improve GPU utilization and overall performance. This is a valid action to take when addressing low GPU utilization.