Not All Reads Are Conflicts: A Write-Only Analysis of the Sui Blockchain

2026-07-29

AI Summary

This work addresses the performance limitations of the Sui blockchain, which stem from workload contention. Conventional read-write conflict graph approaches overestimate actual serialization dependencies. To remedy this, the authors propose a write-only (W-only) analysis model that introduces the W-only conflict graph as a lower bound on contention. By integrating write-set analysis, union-find-based object clustering, and empirical mainnet data, the study precisely characterizes genuine write serialization events. Findings reveal that excluding read-only dependencies eliminates the hub-and-spoke topology commonly assumed, indicating that parallelism potential has been overestimated by 30–40%. Moreover, 10–50% of transaction value flows through serial paths, exposing significant ordering risk. The results demonstrate that Sui’s contention topology is highly assortative and dominated by dense clusters, fundamentally revising prior understanding of its parallel execution capacity.

Abstract

Sui's object-centric data model enables parallel transaction execution, but realised performance is fundamentally bounded by workload contention. Prior empirical analyses of Sui have relied on "read+write" (R+W) conflict graphs inherited from account-based blockchains. Because Sui's engine serialises only on mutable shared access, R+W graphs contain spurious edges, bounding contention from above. In this paper, we adopt a complementary "write-set-only" (W-only) model in which every edge represents a real write-serialisation event, providing a lower bound on contention. Together, the two models bracket Sui's true execution-dependency structure. Applying the W-only analysis to Sui mainnet data through 2025 yields three primary findings. First, removing read-only dependencies (notably the system clock) causes previously reported "hub-and-spoke" structures to collapse. The remaining contention topology is highly assortative and clique-dominated, with the W-only bound shaving roughly $30$--$40\%$ off the R+W estimate of Sui's optimal-parallelism headroom. Second, via union-find object grouping, we isolate DeepBook (Sui's native central limit order book). While it dominates contention by volume, its underlying logic does not impose disproportionate sequential bottlenecks. Finally, we quantify the economic cost of contention, showing that $10$--$50\%$ of the network's USD-denominated value flows through sequentially constrained execution paths, exposing it to potential ordering effects.

Problem

• contention
• conflict graph
• parallel transaction execution
• write-set-only
• execution dependency

Innovation

• write-set-only analysis
• contention modeling
• object-centric blockchain
• execution dependency
• parallelism bound