Practical MDS Array Codes with Balanced Repair Bandwidth, Disk I/O, and Computational Overhead
Year: 2026
Type: conference-paper
Abstract: Maximum distance separable (MDS) codes are widely deployed in distributed storage systems because they provide the maximum reliability for a given storage overhead. However, conventional MDS codes (e.g., Reed-Solomon codes) typically require high repair bandwidth to repair a single failed node. To address this issue, Dimakis et al. introduced minimum storage regenerating (MSR) codes, which minimize the repair bandwidth while maintaining the MDS property. Unfortunately, existing MSR constructions and known lower bounds indicate that high-rate MSR codes with exact repair require an exponential sub-packetization level. This, in turn, leads to numerous non-contiguous disk I/O operations and high computational overhead during repair, thereby limiting practical adoption. In this paper, we present a practical MDS array code construction that balances repair bandwidth, disk I/O efficiency, and computational overhead. The proposed code is obtained from a base MDS code by introducing only a few additional XOR operations and adopting a low sub-packetization level, thereby achieving sub-optimal repair bandwidth with low I/O overhead. Moreover, encoding and repair for single failures can be performed using standard decoding of the base MDS code, together with a small number of XOR operations, resulting in low computational overhead.
Authors: Guodong Li, Jingwen Chu, Sihuang Hu
Institution: Shandong University
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204156429