[Paper Review] Explicit Construction of MBR Codes for Clustered Distributed Storage.

Jy-yong Sohn, Jaekyun Moon|arXiv (Cornell University)|Jan 8, 2018

Advanced Data Storage Technologies3 citations

TL;DR

This paper presents an explicit construction of minimum-bandwidth-regenerating (MBR) codes for clustered distributed storage systems (DSS), generalizing prior nonclustered MBR codes. The scheme enables exact regeneration of any failed node with minimal repair bandwidth, achieving capacity-achieving performance in multi-rack data center environments with arbitrary system parameters.

ABSTRACT

Capacity of the distributed storage system (DSS) is often discussed in the context of tradeoff between storage overhead and repair bandwidth. This paper considers capacity-achieving coding for the clustered form of distributed storage that reflects practical storage networks. The suggested coding scheme is shown to exactly regenerate the arbitrary failed node with minimum required bandwidth, i.e., the proposed scheme is a minimum-bandwidth-regenerating (MBR) code of clustered DSSs with general parameter setting. The proposed code is a generalization of the existing MBR code designed for nonclustered DSSs. This code can be implemented in data centers with multiple racks (clusters), depending on the desired system parameters.

Motivation & Objective

• To address the capacity-achieving coding problem in clustered distributed storage systems reflecting real-world data center architectures.
• To extend existing MBR codes—originally designed for nonclustered DSS—into a clustered setting with general system parameters.
• To achieve exact regeneration of any failed node with minimum repair bandwidth in a multi-rack storage environment.
• To provide an explicit construction method that is implementable in practical data center networks.

Proposed method

• The proposed code generalizes existing nonclustered MBR codes to support clustered DSS with arbitrary parameters.
• It employs a structured coding scheme that maintains the MBR property under the clustered storage model.
• The construction ensures that any single failed node can be regenerated with minimal bandwidth, matching the theoretical minimum.
• The code design accounts for inter-rack communication costs by modeling the cluster topology explicitly.
• It uses a linear-algebraic framework to define encoding and repair processes that preserve optimal repair efficiency.
• The scheme is explicitly constructible, enabling practical deployment in data centers with multiple racks.

Experimental results

Research questions

• RQ1 How can MBR codes be extended to clustered distributed storage systems with multiple racks?
• RQ2 What is the optimal repair bandwidth achievable in a clustered DSS with general system parameters?
• RQ3 Can an explicit construction be designed that achieves exact regeneration with minimum bandwidth in a clustered setting?
• RQ4 How does the proposed code maintain capacity-achieving performance in a multi-rack environment?

Key findings

• The proposed MBR code achieves exact regeneration of any failed node with minimum repair bandwidth, confirming its optimality in the clustered DSS model.
• The code is explicitly constructible, enabling practical implementation in data centers with multiple racks.
• The construction generalizes prior nonclustered MBR codes to support arbitrary system parameters in a clustered setting.
• The scheme maintains capacity-achieving performance, matching the theoretical tradeoff limit for clustered DSS.
• The method ensures minimal inter-rack communication during node repair, aligning with real-world data center network topologies.
• The code structure supports efficient encoding and repair operations under realistic storage cluster constraints.