AWS Delivers High-Performance NFS Access with S3 Files

Key Takeaways

• S3 Files offers NFS 4.1 access to S3 data.
• Latency under 1 ms for active data, terabytes per second throughput.
• High‑performance tier stores files <125 KB, default threshold configurable.
• Pricing: $0.30/GB‑month storage, $0.03/GB reads, $0.06/GB writes (US‑East‑1).
• Available in 34 AWS regions, targeting HPC and AI workloads.

Pulse Analysis

High‑performance computing (HPC) environments have long relied on POSIX‑compatible file systems such as NFS and Lustre because they need fast, low‑latency access to individual files. Object stores like Amazon S3 excel at durability and cost efficiency but historically required applications to use REST‑style GET/PUT calls, a mismatch for legacy Linux workloads. AWS’s launch of Amazon S3 Files bridges that gap by exposing S3 buckets through an NFS 4.1 interface, allowing existing HPC codes to read and write files directly without data migration or code rewrites.

The service promises sub‑millisecond latency for “active” data and aggregate read bandwidth measured in multiple terabytes per second. It achieves this with an automatic tiering layer: small, random‑access files (default < 125 KB) are cached on high‑performance storage, while larger sequential reads fall back to the underlying S3 object store. Although Lustre still delivers higher raw throughput for massive parallel writes, S3 Files provides sufficient speed for interactive AI training, shared‑memory analytics, and workloads that benefit from close‑to‑open consistency across many compute nodes.

From a business perspective, S3 Files simplifies data pipelines by eliminating the need to duplicate datasets between object and file storage. Customers pay only for the high‑performance tier ($0.30 per GB‑month in US‑East‑1) and for reads ($0.03/GB) and writes ($0.06/GB), making the model attractive for variable AI workloads. With availability in 34 AWS regions, the service positions Amazon to capture a larger share of the growing HPC‑AI market, where data‑centric agility often outweighs raw cost considerations.