MLS-C01 Question 282
Select 2A data scientist is training a machine learning model on Amazon SageMaker using an ml.m5.4xlarge instance. After monitoring the training job, they observe that the CPU utilization is consistently below 10%. Additionally, the training job takes significantly longer than expected. Which of the following actions should the data scientist take to rightsize resources and optimize the training job?
- A
Switch to an ml.m5.large instance to reduce underutilized resources.
- B
Switch to a GPU-based instance, such as ml.p3.2xlarge, to accelerate the training process.
- C
Increase the Provisioned IOPS for the EBS volume attached to the instance to improve data throughput.
- D
Use Amazon SageMaker Automatic Model Tuning to find the best hyperparameter configuration.
- E
Switch to an ml.inf1.2xlarge instance for improved inference performance.
Show answer and explanation
Correct answers: A, B
Explanation
The observed low CPU utilization suggests that the current instance is overprovisioned for the training workload. Switching to a smaller CPU instance (ml.m5.large) can reduce costs, while switching to a GPU-based instance (ml.p3.2xlarge) can accelerate the training process if the workload supports GPU acceleration. Both options align with the goal of rightsizing resources effectively.
- A. Correct.
Switching to an ml.m5.large instance is a valid option because it uses fewer CPU resources, which aligns with the observed low CPU utilization. This can help reduce costs.
- B. Correct.
Switching to a GPU-based instance, such as ml.p3.2xlarge, is a valid option if the training job is computationally intensive and can benefit from parallel processing on GPUs. This can significantly reduce training time.
- C. Incorrect.
Increasing the Provisioned IOPS for the EBS volume might improve data throughput, but this is not directly related to the observed issue of low CPU utilization or the prolonged training time.
- D. Incorrect.
While using Amazon SageMaker Automatic Model Tuning is helpful for optimizing hyperparameters, it does not address the issue of underutilized resources or the prolonged training time.
- E. Incorrect.
Switching to an ml.inf1.2xlarge instance is not relevant in this case because it is designed for inference workloads, not training jobs.