[Paper Review] On Locally Recoverable (LRC) Codes

Mario Blaum|arXiv (Cornell University)|Dec 18, 2015

Advanced Data Storage Technologies12 references3 citations

TL;DR

This paper presents simple, explicit constructions of optimal locally recoverable (LRC) codes using parity-check matrices based on extended Reed-Solomon codes. It achieves optimal minimum distance $ d = \ u + g + 1 $ with field size $ q \geq mn $, and reduces field size to $ q \geq n $ when $ g \leq \ell + 1 $, enabling efficient implementations with XOR operations and rotations over small fields.

ABSTRACT

We present simple constructions of optimal erasure-correcting LRC codes by exhibiting their parity-check matrices. When the number of local parities in a parity group plus the number of global parities is smaller than the size of the parity group, the constructed codes are optimal with a field of size at least the length of the code. We can reduce the size of the field to at least the size of the parity groups when the number of global parities equals the number of local parities in a parity group plus one.

Motivation & Objective

• To design simple, explicit constructions of optimal LRC codes that support fast local recovery for single-device failures in distributed storage.
• To minimize the required field size for LRC codes while maintaining optimal minimum distance and global erasure correction capability.
• To provide a framework for constructing codes that balance local repair efficiency with global fault tolerance in practical storage systems.
• To extend known constructions to cases where $ \ell + g < n $, simplifying the design compared to prior complex methods.

Proposed method

• Constructs parity-check matrices using extended Reed-Solomon (ERS) codes for local and global parities.
• Uses a block-structured parity-check matrix where local parities are derived from RS codes with parameters $ (n, \ell; 0,0), (n, \ell; 0,r), \dots $.
• Introduces a global parity submatrix based on an extended RS code of length $ mn - 1 $, with parameters $ (mn-1, g; \ell, 0) $.
• Applies XOR operations and vector rotations over fields defined by irreducible polynomials $ 1 + x + \cdots + x^{p-1} $ for efficient implementation.
• Ensures the resulting code achieves minimum distance $ d = \ell + g + 1 $ by proving the parity-check matrix yields an MDS code after XOR operations.
• Reduces field size to $ q \geq n $ when $ g \leq \ell + 1 $, leveraging structural properties of the parity-check matrix.

Experimental results

Research questions

• RQ1 Can optimal LRC codes be constructed with a simple, explicit parity-check matrix structure?
• RQ2 What is the minimal field size required to achieve optimal minimum distance in LRC codes with local and global parities?
• RQ3 How can field size be reduced when $ g \leq \ell + 1 $ without sacrificing optimality?
• RQ4 Can extended Reed-Solomon codes be used to construct optimal LRC codes with efficient XOR-based decoding?
• RQ5 How does the proposed construction compare in performance and field size to existing PMDS or SD code constructions?

Key findings

• The constructed LRC code achieves the Singleton-type bound with equality, resulting in optimal minimum distance $ d = \ell + g + 1 $.
• For $ \ell + g < n $, the code is optimal and requires a field size of at least $ q = mn $.
• When $ g \leq \ell + 1 $, the field size can be reduced to $ q \geq n $, significantly improving practicality.
• The parity-check matrix structure allows for efficient implementation using only XOR operations and vector rotations over small fields.
• An example construction for $ (m,n;\ell,g) = (2,8;1,4) $ over $ GF(16) $ achieves $ d = 6 $, matching the optimal bound.
• The method generalizes to larger arrays, such as $ 6 \times 5 $ with $ \ell = g = 2 $, yielding $ d = 5 $, and remains optimal under the given constraints.