Paper 2026/796

Pufferfish: DAG-Guided Speculative Execution for Low-Latency BFT SMR

Jianting Zhang, Purdue University
Alberto Sonnino, Mysten Labs / University College London (UCL)
Lefteris Kokoris-Kogias, Mysten Labs
Aniket Kate, Mysten Labs / Purdue University
Abstract

Modern Byzantine fault-tolerant state machine replication (BFT SMR) systems increasingly decouple data dissemination from transaction ordering, but typically execute transactions only after consensus finalizes their order. This execution-after-consensus design serializes ordering and execution on the confirmation path, so transactions incur both costs even under normal network conditions. Decoupling enables dissemination to continue during ordering stalls, but amplifies the same bottleneck: disseminated transactions accumulate into an execution backlog that replicas must drain after ordering recovers. This paper presents Pufferfish, a BFT SMR design that uses speculative execution to move execution off the post-consensus critical path. Replicas speculatively execute transactions according to a predicted pre-commit order derived from a DAG-based consensus substrate. To make this work reusable, Pufferfish combines adaptive probabilistic speculation, which predicts the most likely final order from the local DAG view, with commit-aware snapshots, which preserve selected prefix states to reduce replay when speculation is wrong. In the common case, Pufferfish overlaps execution with consensus to reduce confirmation latency. When ordering is delayed, the same mechanism converts otherwise idle time into useful execution work, limits backlog growth, and shortens recovery. We implement Pufferfish and evaluate it in a geo-distributed AWS deployment across five regions. At 80,000 transactions per second (TPS) with 10 replicas, Pufferfish reduces p99 confirmation latency from 681ms to 430ms (36.9%) relative to an execution-after-consensus baseline. With 49 replicas at 25,000 TPS near the baseline's saturation point, Pufferfish achieves a p99 latency of 306 ms, compared with 724 ms for the baseline. Under induced consecutive ordering stalls, it reduces latency from 17.1 s to 9.8 s, a 42.7% decrease.

Metadata
Available format(s)
PDF
Category
Implementation
Publication info
Preprint.
Keywords
Byzantine Fault-tolerantState Machine ReplicationBlockchainPartial Synchrony
Contact author(s)
zhan4674 @ purdue edu
alberto @ mystenlabs com
lefteris @ mystenlabs com
aniket @ purdue edu
History
2026-11-11: last of 3 revisions
2026-04-22: received
See all versions
Short URL
https://ia.cr/2026/796
License
Creative Commons Attribution
CC BY

BibTeX

@misc{cryptoeprint:2026/796,
      author = {Jianting Zhang and Alberto Sonnino and Lefteris Kokoris-Kogias and Aniket Kate},
      title = {Pufferfish: {DAG}-Guided Speculative Execution for Low-Latency {BFT} {SMR}},
      howpublished = {Cryptology {ePrint} Archive, Paper 2026/796},
      year = {2026},
      url = {https://eprint.iacr.org/2026/796}
}
Note: In order to protect the privacy of readers, eprint.iacr.org does not use cookies or embedded third party content.