NCP-AII Question 30
Select 3You are managing an AI infrastructure that processes large datasets for training deep learning models. The storage system has been experiencing high latency, impacting overall training performance. Which actions would optimize the storage system for AI workloads?
- A
Implement a tiered storage solution that uses NVMe for frequently accessed data and HDDs for cold data.
- B
Enable data compression on the storage system to reduce the amount of disk space used.
- C
Adopt a parallel file system such as Lustre or GPFS for better throughput in distributed training.
- D
Increase the number of CPU cores in the system to improve storage I/O operations.
- E
Utilize data prefetching techniques to load data into memory before it is needed.
Show answer and explanation
Correct answers: A, C, E
Explanation
Optimizing storage for AI workloads involves reducing latency and ensuring high throughput. Tiered storage solutions balance performance and cost by aligning the speed of storage media to the workload requirements. Parallel file systems ensure better storage scalability and throughput for distributed training. Data prefetching minimizes latency by ensuring data is available in memory before it is required. These techniques collectively improve storage performance for AI infrastructure.
- A. Correct.
Tiered storage optimizes performance and cost by matching the speed of storage media (e.g., NVMe for hot data and HDD for cold data) to the workload needs, making it suitable for AI workloads.
- B. Incorrect.
While data compression reduces disk space usage, it adds computational overhead and may increase latency during decompression, making it less effective for optimizing storage in performance-critical AI workloads.
- C. Correct.
Parallel file systems like Lustre or GPFS improve storage throughput and scalability, which are essential for distributed AI training where multiple nodes access the same data.
- D. Incorrect.
Increasing CPU cores does not directly impact storage performance, as storage latency is more influenced by the storage system's architecture and I/O optimization.
- E. Correct.
Data prefetching reduces storage access latency by proactively loading data into memory, particularly beneficial for sequential data access patterns in AI workloads.