Adaptive Chunking & Erasure Coding

• Adaptive Chunking and Erasure Coding are integrated techniques that dynamically adjust data segmentation and redundancy to balance service delay and throughput.
• They employ Maximum Distance Separable codes and backlog-threshold adaptations to mitigate random erasures while managing resource utilization.
• Empirical evaluations in cloud storage and networked systems demonstrate throughput improvements up to 3x and delay reductions of 76%-85% under variable loads.

Adaptive chunking and erasure coding constitute an integrated set of techniques for minimizing delay and optimizing throughput in systems subject to random erasures and service latency, most notably in cloud storage, erasure channels, and networked communication. By dynamically selecting both chunk sizes and erasure code rates in response to varying workload and channel conditions, these methods enable near-optimal trade-offs between service delay, system throughput, resource utilization, and reliability.

1. Mathematical Model and Principles of Adaptive Chunking and Erasure Coding

At the core, objects (files, packets, status updates) are divided into $k$ chunks, each of which is processed—stored, transmitted, or delivered—in parallel. A $(n,k)$ Maximum Distance Separable (MDS) code, typically a Reed–Solomon or random linear code, expands these $k$ chunks into $n$ coded symbols such that any $k$ of the $n$ suffice for reconstruction. This introduces a redundancy ratio $r=n/k$.

Chunk size $B$ is generally set as the object size divided by $k$; i.e., $B=J/k$ for object of size $(n,k)$0. The service time for each chunk may include both a fixed overhead and a random (typically exponential) component, modeled as $(n,k)$1 with $(n,k)$2 and $(n,k)$3 typically affine in $(n,k)$4 (Liang et al., 2013, Liang et al., 2014, Liang et al., 2013).

For each user request, the mean service delay $(n,k)$5 and queueing delay $(n,k)$6 are key determinants of performance: $(n,k)$7

$(n,k)$8

where $(n,k)$9 is expected system usage per request and $k$0 is the number of parallel threads or connections.

Increasing $k$1 reduces $k$2 via the statistical diversity of chunk-completion times but increases resource usage and hence system congestion, raising $k$3 (Liang et al., 2014). The selection of $k$4 must therefore be dynamically adapted according to arrival rates and queue backlogs.

2. Adaptive Algorithms and Backlog Thresholding

Backlog-driven adaptation is the principal mechanism for dynamic selection of chunk size and code rate. Queue length $k$5, which reflects the time-varying load and congestion state, serves as the feedback variable.

The TOFEC (Throughput Optimal FEC Cloud) policy formalizes the adaptive mapping $k$6 for each request class. For fixed system parameters, first-order optimality conditions yield strictly decreasing functions $k$7, ensuring that under light load, high $k$8 and $k$9 (many small chunks and high redundancy) are chosen to minimize service delay, while under heavy load, $n$0 and $n$1 are reduced to preserve throughput and prevent backlog growth (Liang et al., 2013, Liang et al., 2014, Liang et al., 2013).

Precomputed threshold sequences $n$2 and $n$3 for each class $n$4 enable stateless, efficient run-time adaptation. On each request arrival, a smoothed queue estimate $n$5 is compared to these thresholds to select the appropriate $n$6 pair: $n$7

$n$8

with admission control ensuring $n$9 to prevent connection explosion.

Tracing the queue-adaptive selection across workloads shows that TOFEC interpolation tracks the lower envelope of all fixed-code throughput-delay curves, achieving empirical mean delay improvements up to $k$0 at low load and preserving full system capacity (up to $k$1 more requests supported) under heavy load (Liang et al., 2013, Liang et al., 2014, Liang et al., 2013).

3. Incremental Redundancy and Block Size Optimization

In erasure channels and hybrid-ARQ communication, adaptive chunking and erasure coding appear in the selection of incremental block sizes for transmissions with feedback. Each message of length $k$2 is encoded into $k$3 symbols, sent in $k$4 chunks. After every chunk, decoding is attempted. Sequential differential optimization (SDO) is employed to choose chunk end-points $k$5 to minimize expected transmission cost $k$6, the total number of symbols sent until successful decoding (Heidarzadeh et al., 2018).

Key findings include the asymptotic decoupling of code overhead (Erdös–Borwein constant $k$7) and channel erasures (factor $k$8), forming

$k$9

which fully characterizes average blocklength. Smooth CDF approximations (normal/log-normal) allow recursive computation of optimal chunk boundaries, yielding code rates and block sizes that maximize throughput $n$0 in delay-sensitive settings.

This optimization provides design rules for practical block-ACK and feedback-limited systems, and yields methods to maintain near-optimal throughput-delay tradeoffs by bridging random coding and channel statistics (Heidarzadeh et al., 2018).

4. Adaptive Coding for Networked and Broadcast Systems

In multi-user or multi-hop network scenarios, adaptive chunking and coding extends to batched network coding, random linear network coding with feedback, and information freshness/AoI metrics.

In batched network coding, data is chunked into batches, which are adaptively recoded at each hop according to the batch's incoming rank (degree of freedom). The expected rank after recoding and transmission through an erasure channel, $n$1, is a concave function in the number of recoded packets, enabling a finite-dimensional concave optimization (water-level "almost-deterministic" allocation) for resource allocation and throughput optimality (Yin et al., 2021). This ensures per-batch redundancy adapts to real-time erasure estimates with minimal randomness in resource allocation and strong robustness to parameter estimation errors.

In broadcast erasure channels with per-symbol feedback, adaptive schemes split updates into $n$2 symbols, use rateless random linear coding for the strong user, and invoke mixed coding during periods of user desynchronization. This maintains low Age of Information (AoI) for both users, turning the otherwise exponential AoI growth in the weak user into linear scaling in $n$3 (Feng et al., 2019).

Adaptive causal RLNC (AC-RLNC) for point-to-point and networked systems applies a two-stage adaptation: a priori FEC based on observed channel erasure rates and a posteriori FEC triggered by feedback, both tuned via a threshold on redundancy-vs-throughput tradeoff. This methodology achieves $n$4 of channel capacity with tightly bounded delay, outperforming non-adaptive ARQ baselines, particularly in bursty or high-latency scenarios (Cohen et al., 2019).

5. Performance Evaluation and Empirical Validation

Empirical studies, particularly using Amazon S3 traces, consistently demonstrate the efficacy of adaptive chunking and erasure coding in realistic cloud environments. Key observations include:

• Aggressive chunking and redundancy (high $n$5) reduce mean and high-percentile delays by up to $n$6 under light workloads (Liang et al., 2013).
• Static, non-adaptive strategies optimized for either throughput or latency yield suboptimal tradeoffs: the former suffer high tail delays, while the latter collapse system capacity at loads exceeding $n$7 of the no-coding regime.
• Load-adaptive algorithms such as TOFEC, and its simpler greedy variants, provide the best average delays while preserving full rate regions, though only threshold-based adaptive methods match the static optimum across all percentiles (Liang et al., 2014, Liang et al., 2013, Liang et al., 2013).
• Under abrupt load changes, backlog-driven adaptation algorithms reconverge to new optima within $n$810 seconds, dramatically outperforming static codes in clearing backlog (Liang et al., 2013).

6. System Design Guidelines, Limitations, and Extensions

Key design considerations include:

• Selection of chunk size $n$9: must balance per-chunk overheads and parallelism; empirical evidence supports $r=n/k$00.5–1MB as effective for S3 (Liang et al., 2013).
• Maximum code length $r=n/k$1: restrict to avoid excessive connection load at low arrivals; practical $r=n/k$2–$r=n/k$3.
• Theoretical models generally approximate queues as M/M/1 or M/G/1, assuming i.i.d. chunk delays—a reasonable fit for measured S3 distributions (correlation $r=n/k$4).
• Adaptive schemes assume fast and reliable measurement of queue lengths or feedback; error-tolerant algorithms exist for the tuning of per-batch redundancy in network coding (Yin et al., 2021).
• Extensions include joint $r=n/k$5 adaptation, deadline-awareness, multi-class thresholds, and integration with cost or energy objectives.

Limitations surface when chunk service times are highly correlated, which degrades diversity gains, or where feedback is unavailable or unreliable, requiring primarily a priori adaptation modes. Practical systems realize the full benefits only when storage and compute resources can support the requisite parallelism and fast scheduling.

---

References:

• "TOFEC: Achieving Optimal Throughput-Delay Trade-off of Cloud Storage Using Erasure Codes" (Liang et al., 2013)
• "On Throughput-Delay Optimal Access to Storage Clouds via Load Adaptive Coding and Chunking" (Liang et al., 2014)
• "FAST CLOUD: Pushing the Envelope on Delay Performance of Cloud Storage with Coding" (Liang et al., 2013)
• "A Systematic Approach to Incremental Redundancy over Erasure Channels" (Heidarzadeh et al., 2018)
• "Adaptive Coding for Information Freshness in a Two-user Broadcast Erasure Channel" (Feng et al., 2019)
• "A Unified Adaptive Recoding Framework for Batched Network Coding" (Yin et al., 2021)
• "Adaptive Causal Network Coding with Feedback" (Cohen et al., 2019)