MLA-C01 Question 208
Single answerA data scientist is training a machine learning model using Amazon SageMaker. The training job involves a large dataset stored in Amazon S3, and the model requires GPUs for faster training. The team notices that the training job is taking longer than expected, and GPU utilization is low. Upon investigation, the team finds that the model spends a significant amount of time fetching data from Amazon S3. What is the most effective way to resolve this issue?
- A
Use Amazon FSx for Lustre to cache the dataset for faster retrieval.
- B
Switch to a CPU-based instance type to save costs on unused GPU resources.
- C
Enable Amazon S3 Transfer Acceleration to reduce data transfer latency.
- D
Use Amazon SageMaker Pipe mode to stream data directly from Amazon S3.
Show answer and explanation
Correct answer: D
Explanation
Amazon SageMaker Pipe mode allows data to be streamed directly from Amazon S3 to the training instances in real-time, reducing the time the model spends waiting for data to be loaded. This increases GPU utilization and improves training efficiency for large datasets, making it the optimal solution for this scenario.
- A. Incorrect.
Amazon FSx for Lustre can be used to cache datasets for high-performance access; however, it is not the most efficient solution for streaming data in this SageMaker use case. SageMaker Pipe mode is more suitable for addressing the specific issue described.
- B. Incorrect.
Switching to a CPU-based instance type would save costs, but it does not address the underlying issue of low GPU utilization caused by data retrieval inefficiencies.
- C. Incorrect.
Amazon S3 Transfer Acceleration can improve data transfer speeds over long distances, but it does not resolve data streaming inefficiencies between Amazon S3 and SageMaker training instances.
- D. Correct.
Amazon SageMaker Pipe mode streams data directly from Amazon S3, allowing for lower memory usage and higher GPU utilization. This is the most effective solution to resolve the issue of low GPU utilization caused by data fetch delays.