Which solution will meet these requirements?
Modify the training configuration to use two ml.p2.xlarge instances.
Modify the training configuration to use Pipe input mode.
Modify the training configuration to use a single ml.p3.2xlarge instance.
Modify the training configuration to use Amazon Elastic File System (Amazon EFS) instead of Amazon S3 to store the input training data.
Explanations:
Using two ml.p2.xlarge instances will increase the cost and may not solve the storage issue since the bottleneck is related to input data size. Simply adding more instances won’t address the storage problem.
Using Pipe input mode allows for streaming data directly from Amazon S3 to the training job. This method avoids the need for storing all the data on the local disk of the instance, which resolves the storage limitation while optimizing performance.
The ml.p3.2xlarge instance is a more powerful instance with better GPU performance but does not address the storage problem related to input data size. It is not cost-effective for this specific issue, as it does not solve the root cause.
Using Amazon EFS for input data storage is not necessary and would add extra cost for an NFS solution. S3 is already a highly optimized and cost-effective storage solution for training data, and switching to EFS would not resolve the storage issue effectively.
I calculate that the answer is:
Modify the training configuration to use Pipe input mode.