Scalable constructions of fractional repetition codes in distributed storage systems

In distributed storage systems built using commodity hardware, it is necessary to have data redundancy in order to ensure system reliability. In such systems, it is also often desirable to be able to quickly repair storage nodes that fail. We consider a scheme–introduced by El Rouayheb and Ramchandran–which uses combinatorial block design in order to design storage systems that enable efficient (and exact) node repair. In this work, we investigate systems where node sizes may be much larger than replication degrees, and explicitly provide algorithms for constructing these storage designs. Our designs, which are related to projective geometries, are based on the construction of bipartite cage graphs (with girth 6) and the concept of mutually-orthogonal Latin squares. Via these constructions, we can guarantee that the resulting designs require the fewest number of storage nodes for the given parameters, and can further show that these systems can be easily expanded without need for frequent reconfiguration.

💡 Research Summary

The paper addresses the problem of designing distributed storage systems that combine high data reliability with fast node repair, especially in scenarios where each storage node can hold many data chunks while each chunk is replicated only a few times. Building on the fractional repetition code framework introduced by El Rouayheb and Ramchandran, the authors focus on the asymmetric case where the node size (the number of chunks stored per node) is much larger than the replication degree (the number of copies of each chunk).

The core of the construction is a combinatorial interpretation of Steiner systems as bipartite incidence graphs. In such a graph, one side (X) represents storage nodes (blocks) and the other side (Y) represents data chunks (elements). A block contains k elements, each element belongs to l blocks, and the absence of 4‑cycles guarantees that no pair of elements appears in more than one block – the defining property of a Steiner system with t = 2. The authors prove a Moore‑type lower bound (Lemma 1) on the number of vertices required for a bipartite graph with girth at least six (no 4‑cycles). When the bound is met, the graph is a cage and the corresponding storage design is optimal in the sense of using the fewest possible nodes and chunks for the given parameters.

To construct such optimal graphs, the paper leverages two well‑known combinatorial objects: (i) mutually orthogonal Latin squares (MOLS) and (ii) finite projective geometries. For any prime power q, a set of q MOLS of order q exists. By arranging the connections between the second and third layers of the bipartite graph according to these Latin squares, the authors guarantee that no 4‑cycle is introduced. The resulting regular bipartite cage has degree q + 1 on both sides and contains q² + q + 1 vertices on each side, which matches the point‑line incidence structure of the projective plane PG(2, q).

Algorithm 1 details the step‑by‑step construction for the regular case (k = l = q + 1). Starting from a single Y‑vertex, the algorithm creates a star of X‑vertices, expands each X‑vertex to q new Y‑vertices, and then uses the MOLS to connect the remaining X‑vertices to the new Y‑vertices without forming 4‑cycles. The resulting graph is bipartite, girth‑6, and meets the Moore bound, thus providing the smallest possible storage system for the chosen replication degree and node size.

The most significant contribution lies in the scalable extension of this construction. By treating the regular cage as a base case (n = 1) and recursively applying the same MOLS‑based expansion, the authors obtain a family of designs where each node can store up to

  qⁿ + qⁿ⁻¹ + … + q + 1

chunks, while each chunk is still replicated exactly q + 1 times. Importantly, the expansion does not require moving any existing data; new nodes and new chunks are simply added, preserving the Steiner property at every stage. This property directly addresses practical concerns in large‑scale cloud storage, where capacity must be increased without costly data reshuffling or downtime.

From a systems perspective, the designs support exact repair: when a node fails, the replacement node can be reconstructed by contacting exactly k = q + 1 surviving nodes, each providing a single chunk. Because the data is stored uncoded, the same nodes can also serve as compute resources, enabling data‑local processing—a valuable feature for big‑data analytics workloads.

The paper situates its work within a broad literature: it contrasts with network‑coding approaches that target functional repair, with prior fractional repetition codes that rely on random or small‑scale constructions, and with block‑design based RAID or parity‑distribution schemes. By providing a deterministic, mathematically optimal, and easily extensible construction, the authors bridge the gap between theoretical coding bounds and practical storage system engineering.

In summary, the authors present a novel, graph‑theoretic method for building fractional repetition codes that achieve the theoretical minimum number of storage nodes and data chunks for a given replication degree and node capacity. The method is based on girth‑6 bipartite cage graphs constructed via mutually orthogonal Latin squares and finite projective geometries, and it supports seamless, data‑preserving expansion. This work advances the state of the art in reliable, repair‑efficient distributed storage and opens avenues for applying similar combinatorial techniques to error‑correcting codes, RAID designs, and other distributed data‑management problems.