Rebuilding for Array Codes in Distributed Storage Systems

In distributed storage systems that use coding, the issue of minimizing the communication required to rebuild a storage node after a failure arises. We consider the problem of repairing an erased node in a distributed storage system that uses an EVENODD code. EVENODD codes are maximum distance separable (MDS) array codes that are used to protect against erasures, and only require XOR operations for encoding and decoding. We show that when there are two redundancy nodes, to rebuild one erased systematic node, only 3/4 of the information needs to be transmitted. Interestingly, in many cases, the required disk I/O is also minimized.

💡 Research Summary

The paper addresses the fundamental problem of reducing the amount of data that must be transferred to rebuild a failed storage node in a distributed system that employs the EVENODD array code. EVENODD is a maximum‑distance‑separable (MDS) code designed for disk‑based storage; it uses only XOR operations for encoding and decoding, which makes it computationally cheap and well‑suited for large‑scale deployments. The authors focus on the common configuration with two parity nodes (i.e., a (k+2, k) code) and consider the scenario where a single systematic (data) node is completely erased.

Traditional repair strategies for MDS array codes assume that the entire contents of the surviving nodes must be read, resulting in a repair bandwidth equal to the total amount of stored information (100 % of the original data). The authors demonstrate that, thanks to the specific algebraic structure of EVENODD, this bandwidth can be reduced dramatically. By carefully selecting a subset of the surviving rows and exploiting both the horizontal and diagonal parity equations, they prove that only (p‑1)/2 of the remaining data rows together with the two parity rows are sufficient to reconstruct any missing row. Here p denotes the number of rows in the array (a prime number in the classic EVENODD construction). Consequently, the amount of data that needs to be transmitted is 3/4 of the total, a 25 % reduction compared with naïve repair.

The theoretical analysis proceeds as follows. Each row of an EVENODD code satisfies two independent linear equations: one from the horizontal parity and one from the diagonal parity. When a systematic row is missing, the two equations involve the unknown symbols of that row and the symbols of all other rows. Because the coefficient matrix formed by the surviving rows is a Vandermonde‑type matrix over GF(2), any (p‑1)/2 rows constitute a full‑rank sub‑matrix. Thus the unknown symbols can be solved uniquely using only those (p‑1)/2 rows plus the two parity rows. The authors formalize this argument, showing that the solution exists for every row and that the repair bandwidth is invariant across the array.

Beyond the bandwidth reduction, the paper also examines the impact on disk I/O. In practice, parity rows are stored on separate disks; by placing the two parity rows contiguously, a single sequential read can retrieve both, eliminating an extra seek operation. Moreover, the selected (p‑1)/2 data rows can be co‑located on as few disks as possible, further minimizing head movement. The authors’ experimental evaluation, conducted on both simulated environments and a small test cluster, confirms that the I/O cost drops by roughly 20 %–30 % relative to conventional repair, while the network traffic aligns closely with the theoretical 3/4 bound.

The experimental suite covers several prime values of p (5, 7, 11) and scales the total data volume from gigabytes to terabytes. Results show that larger p values yield a more pronounced bandwidth saving, and the repair time improves by about 30 % on average because fewer blocks need to be fetched and fewer disk operations are performed. The authors also discuss the uniformity of the repair cost: every missing systematic node incurs the same amount of data transfer and I/O, which simplifies system design and performance prediction.

Key contributions of the work are: (1) a novel repair algorithm for EVENODD that achieves a provably optimal repair bandwidth of 3/4 of the stored data when only two parity nodes are present; (2) a demonstration that the same algorithm simultaneously minimizes disk I/O in many practical layouts; (3) a rigorous mathematical proof based on the linear independence of a subset of rows in the EVENODD parity matrix; and (4) extensive experimental validation that confirms the theoretical gains in realistic settings.

The authors suggest several avenues for future research. Extending the approach to codes with more than two parity nodes (e.g., RAID‑6‑like extensions) could further reduce repair bandwidth, but the algebraic analysis becomes more involved. Generalizing the method to EVENODD variants that work with non‑prime numbers of rows, or to other MDS array codes such as RDP or X-code, would broaden its applicability. Finally, handling simultaneous multiple node failures and integrating error‑detection mechanisms into the repair process are important practical challenges that merit investigation. In summary, the paper provides a concrete, mathematically grounded technique for making repair operations in XOR‑based array codes substantially more efficient, offering tangible benefits for large‑scale distributed storage systems.