GraceSSD: Per-File Reliability Hints for Error-Tolerant Machine Learning Storage

Abstract

We propose GraceSSD, a per-file storage abstraction that exposes the inherent error resilience of modern machine learning (ML) applications to the underlying flash device. GraceSSD lets an application attach a lightweight resilience hint to each file through the standard extended-attribute (xattr) interface, declaring how much bit-level imprecision of the file the application can tolerate and how strict its read-latency requirement is. These per-inode hints flow unmodified through the kernel storage stack to the SSD, where the controller relaxes the uncorrectable bit-error-rate (UBER) target per file, letting older flash blocks be read with fewer LDPC decoding retries. Our evaluation on MLPerf training workloads shows that GraceSSD prolongs SSD lifetime by 76.7% and improves end-to-end training throughput by 54.3% on average without sacrificing accuracy, requiring only a few lines of application code.