[Paper Review] Storage codes -- coding rate and repair locality

Henk D. L. Hollmann|arXiv (Cornell University)|Jan 18, 2013

Advanced Data Storage Technologies19 references4 citations

TL;DR

This paper establishes new information-theoretic lower bounds on storage overhead in distributed storage codes under functional repair, showing that when repair locality $ r $ is fixed, the maximum coding rate $ R $ is bounded by $ r/(r+1) $ if $ \alpha = \beta $, and by $ 1/2 $ if $ \alpha = r\beta $. These bounds are tight and generalize known limits for MDS and exact-repair codes.

ABSTRACT

The {\em repair locality} of a distributed storage code is the maximum number of nodes that ever needs to be contacted during the repair of a failed node. Having small repair locality is desirable, since it is proportional to the number of disk accesses during repair. However, recent publications show that small repair locality comes with a penalty in terms of code distance or storage overhead if exact repair is required. Here, we first review some of the main results on storage codes under various repair regimes and discuss the recent work on possible (information-theoretical) trade-offs between repair locality and other code parameters like storage overhead and code distance, under the exact repair regime. Then we present some new information theoretical lower bounds on the storage overhead as a function of the repair locality, valid for all common coding and repair models. In particular, we show that if each of the $n$ nodes in a distributed storage system has storage capacity $\ga$ and if, at any time, a failed node can be {\em functionally} repaired by contacting {\em some} set of $r$ nodes (which may depend on the actual state of the system) and downloading an amount $\gb$ of data from each, then in the extreme cases where $\ga=\gb$ or $\ga = r\gb$, the maximal coding rate is at most $r/(r+1)$ or 1/2, respectively (that is, the excess storage overhead is at least $1/r$ or 1, respectively).

Motivation & Objective

• To investigate the fundamental trade-off between coding rate and repair locality in distributed storage systems under functional repair.
• To derive information-theoretic lower bounds on storage overhead as a function of repair locality $ r $, valid across common coding and repair models.
• To close the gap between known constructions and theoretical limits by proving tight bounds in extreme cases of $ \alpha = \beta $ and $ \alpha = r\beta $.
• To generalize the cutset bound framework to allow adaptive repair set selection during node repair, modeling a game-theoretic interaction between KILLER and BUILDER.
• To provide tight upper bounds on the maximum achievable coding rate $ R = m/(n\alpha) $ under functional repair with variable repair set choices.

Proposed method

• Formalizes the repair process as a game between KILLER (who kills nodes) and BUILDER (who creates new nodes by contacting $ r $ live nodes), modeling the information flow network.
• Applies max-flow min-cut arguments on the evolving information flow graph to derive upper bounds on the maximum amount of storable information $ m $.
• Uses the cutset bound framework from [4] but extends it to allow dynamic repair set selection, where the set of $ r $ helper nodes can be chosen adaptively during repair.
• Derives bounds on the maximum coding rate $ R = m/(n\alpha) $ under two extreme cases: $ \alpha = \beta $ and $ \alpha = r\beta $, using game-theoretic analysis.
• Proves that in the $ \alpha = \beta $ case, $ R \leq r/(r+1) $, and in the $ \alpha = r\beta $ case, $ R \leq 1/2 $, with equality achievable via known constructions.
• Generalizes results to $ r = 2 $, deriving a tighter bound $ R \leq (\alpha + \beta)/(3\alpha) $, with a precise expression for $ m \leq q\alpha + (q-e)\beta $ when $ n = 3q - e $.

Experimental results

Research questions

• RQ1 What is the maximum achievable coding rate $ R $ in a distributed storage system with functional repair and repair locality $ r $, when the storage per node $ \alpha $ and repair bandwidth per helper $ \beta $ are constrained?
• RQ2 How does the coding rate $ R $ scale under the extreme cases $ \alpha = \beta $ and $ \alpha = r\beta $, and are these bounds tight?
• RQ3 Can the cutset bound be generalized to allow adaptive repair set selection (i.e., choosing $ r $ helpers dynamically), and what impact does this have on rate bounds?
• RQ4 Is there a fundamental trade-off between repair locality $ r $, storage overhead, and code distance in functional repair systems?
• RQ5 Can information-theoretic bounds be derived that are valid across all common coding and repair models, including functional repair with variable helper sets?

Key findings

• For the case $ \alpha = \beta $, the maximum coding rate is bounded by $ R \leq r/(r+1) $, and this bound is tight, achieved by exact-repair MDS codes with $ n = r + 1 $.
• For the case $ \alpha = r\beta $, the maximum coding rate is bounded by $ R \leq 1/2 $, and this bound is tight, achieved by exact-repair-by-transfer linear codes.
• When $ r = 2 $, the coding rate satisfies $ R \leq (\alpha + \beta)/(3\alpha) $, and for $ n = 3q - e $ with $ e \in \{0,1,2\} $, the maximum stored information is $ m \leq q\alpha + (q - e)\beta $.
• The bounds are information-theoretically tight and hold for all functional repair models, regardless of whether the repair set is fixed or adaptively chosen.
• The results generalize previous work by showing that the trade-off between repair locality and coding rate is fundamentally limited by information flow network structure, even under adaptive helper selection.
• The game-theoretic model of KILLER and BUILDER confirms that the derived bounds represent the worst-case capacity of any information flow network under optimal adversarial node failure and adaptive repair.