MLA-C01 Question 392
Single answerA company is running an Amazon SageMaker training job for a machine learning model using an ml.p3.16xlarge instance. The team notices that the GPU utilization for the instance is consistently low. How can the team optimize infrastructure usage and reduce costs while ensuring the training job completes successfully?
- A
Switch to a smaller instance type, such as ml.p3.2xlarge, and monitor GPU utilization.
- B
Enable SageMaker Managed Spot Training for the training job.
- C
Switch to an instance type with higher GPU capacity, such as ml.g5.24xlarge.
- D
Increase the batch size of the training job to improve GPU utilization.
Show answer and explanation
Correct answer: A
Explanation
Low GPU utilization on an ml.p3.16xlarge instance indicates that the selected instance type may be over-provisioned for the training job's requirements. Switching to a smaller instance type, such as ml.p3.2xlarge, allows the team to optimize infrastructure usage and reduce costs while still ensuring the training job completes successfully. Monitoring GPU utilization after switching is important to confirm the instance meets the job's needs.
- A. Correct.
Switching to a smaller instance type, such as ml.p3.2xlarge, can reduce costs while still meeting the training job's requirements if GPU utilization is low. This is a valid optimization strategy.
- B. Incorrect.
While SageMaker Managed Spot Training can reduce costs, it is not directly related to optimizing the GPU utilization issue in this scenario.
- C. Incorrect.
Switching to an instance with even higher GPU capacity, such as ml.g5.24xlarge, would increase costs unnecessarily and is not appropriate given the low GPU utilization.
- D. Incorrect.
Increasing the batch size might improve GPU utilization, but it does not directly optimize infrastructure costs. Additionally, it could lead to other issues, such as out-of-memory errors.