SAP-C02 Practice Question 55: Training Data Shared Storage
AWS Certified Solutions Architect – Professional (SAP-C02) · Test 6 · question 5 of 10
A computer vision team trains models on 32 GPU instances at a time. The training corpus is 60 TB of images and annotation files that lives in Amazon S3, which the data engineering group treats as the authoritative copy and updates daily.
Today each training job begins by copying the subset it needs from S3 onto the instances' Amazon EBS volumes. That copy takes 80 to 100 minutes before a single training step runs, it is repeated on every instance, and the same objects are downloaded again for every experiment on the same data. The training framework expects a POSIX file system mounted at a common path on all 32 instances, and it reads small files randomly at high concurrency.
The team wants training to start almost immediately, wants aggregate read throughput measured in hundreds of gigabytes per second, and wants S3 to remain the source of truth. What should the solutions architect recommend?
Show answer and explanation
Answer:
FSx for Lustre is AWS's high-performance parallel file system for workloads where storage has to keep up with compute, and AWS names machine learning and HPC as its target use cases, citing sub-millisecond latencies, up to multiple terabytes per second of throughput, and millions of IOPS, which matches the small-file, high-concurrency read pattern described. The S3 data repository integration is what removes the copy step: when the file system is linked to an S3 bucket, it transparently presents S3 objects as files, importing the listing at creation and pulling object content as it is read, so training starts almost immediately and S3 remains the authoritative copy. It is POSIX-compliant and mounts at a common path on all 32 instances, so the framework needs no change, and AWS specifically notes that using FSx for Lustre avoids the repetitive download of the same objects for iterative jobs on the same dataset.
Option B removes the per-instance copy but not the bottleneck. Amazon EFS is a general-purpose NFS file system; it is not built for the multi-hundred-gigabyte-per-second aggregate throughput this job needs, and it makes a second authoritative copy of 60 TB that must be kept in step with the daily S3 updates.
Option C misuses block storage. EBS Multi-Attach shares one volume with a limited number of instances in a single Availability Zone and provides no shared file system semantics; without a cluster file system on top, 32 instances cannot safely mount it, and it still requires copying the corpus.
Option D optimizes the step that should be eliminated. Transfer Acceleration is aimed at long-distance internet uploads and downloads, not at transfers between EC2 and S3 in the same Region, so it would not meaningfully shorten the 80-to-100-minute wait.
Reference: What is Amazon FSx for Lustre?