Evaluating GossipSub for Data Availability Sampling Under Ethereum Consensus Deadlines

Abstract

This thesis investigates how GossipSub configuration choices influence data dissemination for a Data-Availability Sampling (DAS) workload under strict consensus time bounds. Using a configurable PeerSim-based simulator, we model a FullDAS-like setting in which a block producer erasure-codes a blob into a 2D extended matrix of share segments, parti- tions row/column segments into topics (custody-style sharding), and executes a two-phase workflow: seeding, where share segments are disseminated over topic meshes, and sampling, where validators must retrieve uniformly random share segments within a T_DAS = 4 s deadline. We systematically vary topic granularity (TOPICS), segmentation (segment amount, SA), replication (K-copies), bandwidth caps, and omission fault rate α, and measure phase success rates, completion-time distributions (with emphasis on tail latency), bandwidth consumption, and duplication overhead. The results show that segmentation and replication dominate performance and overhead: increasing SA from coarse to moderate values reduces duplication with diminishing returns beyond SA ≈ 1-16, while larger K increases redundancy and overhead and mainly provide a robustness margin under adverse conditions. Seeding completes quickly and remains resilient for SA≥ 4 even at high omission, whereas sampling is tail-latency dominated and degrades more sharply as α increases, leading to widespread deadline misses near α = 0.5. Based on these findings, we adopt TOPICS= 256, SA= 8, K = 4, and a conservative per-node bandwidth cap of 60 Mbit/s for faulted multi-slot experiments to isolate GossipSub dynamics from bandwidth saturation.