[Paper Review] Optimal Repair Schemes for Some Families of Full-Length Reed-Solomon Codes

Son Hoang Dau, Olgica Milenković|arXiv (Cornell University)|Jan 15, 2017

Advanced Data Storage Technologies13 references7 citations

TL;DR

This paper presents optimal linear repair schemes for full-length Reed-Solomon codes over $\mathbb{F}_{q^t}$ with any redundancy $r \geq q^s$ and $1 \leq s < t$, using linearized polynomials to achieve a repair bandwidth of $(n-1)(t-s)\log_2 q$ bits. The scheme is optimal when $n = q^t$ and $r = q^s$, extending prior work that required $(t-s) \mid t$ and using subfield traces to reduce repair overhead.

ABSTRACT

Reed-Solomon codes have found many applications in practical storage systems, but were until recently considered unsuitable for distributed storage applications due to the widely-held belief that they have poor repair bandwidth. The work of Guruswami and Wootters (STOC'16) has shown that one can actually perform bandwidth-efficient linear repair with Reed-Solomon codes: When the codes are over the field $\mathbb{F}_{q^t}$ and the number of parities $r \geq q^s$, where $(t-s)$ divides $t$, there exists a linear scheme that achieves a repair bandwidth of $(n-1)(t-s)\log_2 q$ bits. We extend this result by showing the existence of such a linear repair scheme for every $1 \leq s &lt; t$. Moreover, our new schemes are optimal among all linear repair schemes for Reed-Solomon codes when $n = q^t$ and $r = q^s$. Additionally, we improve the lower bound on the repair bandwidth for Reed-Solomon codes, also established in the work of Guruswami and Wootters.

Motivation & Objective

• To close the gap in optimal repair bandwidth for full-length Reed-Solomon codes when redundancy $r = q^s$ and $(t-s) \nmid t$, where prior schemes were limited.
• To extend the Guruswami-Wootters repair framework beyond the trace-based method to work for all $1 \leq s < t$, not just when $t-s$ divides $t$.
• To construct linear repair schemes with minimal repair bandwidth for Reed-Solomon codes with full length $n = q^t$ and redundancy $r = q^s$, proving optimality.
• To improve the lower bound on repair bandwidth for Reed-Solomon codes, showing it matches the achievable bandwidth in all tested cases.

Proposed method

• The method uses linearized polynomials $L_W(x) = \prod_{w \in W}(x - w)$, where $W$ is an $\mathbb{F}_q$-subspace of dimension $s$ in $\mathbb{F}_{q^t}$, to generate dual codewords for repair.
• Check polynomials $g_i(x) = L_W(u_i(x - \alpha^*)) / (x - \alpha^*)$ are constructed using an $\mathbb{F}_q$-basis $\{u_1, \dots, u_t\}$ of $\mathbb{F}_{q^t}$, ensuring they are valid checks for the Reed-Solomon code.
• The repair process recovers a failed symbol $f(\alpha^*)$ by downloading $g_i(\alpha)$ values from $n-1$ surviving nodes, each contributing $t-s$ bits over $\mathbb{F}_q$, yielding total bandwidth $(n-1)(t-s)\log_2 q$ bits.
• The rank of the evaluation of $g_i(x)$ at $\alpha \neq \alpha^*$ is bounded by $t-s$ over $\mathbb{F}_q$, ensuring bandwidth efficiency.
• The scheme is proven optimal when $n = q^t$ and $r = q^s$, using a lower bound on repair bandwidth that improves upon Guruswami and Wootters' bound.
• Theoretical and numerical evidence supports that the improved lower bound matches the optimal repair bandwidth across all full-length Reed-Solomon codes.

Experimental results

Research questions

• RQ1 Can optimal repair bandwidth be achieved for full-length Reed-Solomon codes with $r = q^s$ and $s < t$ when $(t-s) \nmid t$?
• RQ2 Is it possible to generalize the Guruswami-Wootters repair scheme beyond trace-based methods to use linearized polynomials with distinct roots for all $1 \leq s < t$?
• RQ3 What is the tightest possible lower bound on the repair bandwidth for Reed-Solomon codes, and does it match the achievable bandwidth in full-length cases?
• RQ4 Can a repair scheme be constructed that is optimal for all $r \geq q^s$ and $1 \leq s < t$, not just for $r = q^s$?
• RQ5 Does the use of linearized polynomials instead of trace polynomials yield better or equivalent repair efficiency in terms of bandwidth and subpacketization?

Key findings

• The repair bandwidth of $(n-1)(t-s)\log_2 q$ bits is achievable for all full-length Reed-Solomon codes with $n = q^t$ and $r \geq q^s$, for any $1 \leq s < t$, using linearized polynomials.
• The proposed repair scheme is optimal among all linear repair schemes when $n = q^t$ and $r = q^s$, matching the theoretical lower bound.
• The improved lower bound on repair bandwidth for Reed-Solomon codes matches the achievable bandwidth in all tested cases, suggesting it is tight.
• The use of linearized polynomials with distinct roots allows repair schemes to be constructed for all $s < t$, removing the prior restriction that $(t-s) \mid t$.
• The repair scheme achieves optimal bandwidth with subpacketization level $\log_q n$, significantly lower than exponential subpacketization in other MSR codes.
• Theoretical and numerical results indicate that the improved lower bound is tight and matches the optimal repair bandwidth for all full-length Reed-Solomon codes.