Duet: Co-Optimizing P2P Message Propagation and Rotating-Leader Consensus

arXiv:2607.22209 · cs.DC · Submitted 2026-07-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: We're looking at the paper: "Duet: Co-Optimizing P2P Message Propagation and Rotating-Leader Consensus".

Jane: Here's the gist of the paper. In blockchain systems, peer-to-peer (P2P) overlay networks play a crucial role in providing reliable, scalable and efficient message-delivery services to upper layers.

Tom: With this support, we propose three improvements to rotating-leader consensus protocols and their underlying P2P networks: (1) accelerating leader rotation; (2) introducing a reliable-broadcast paradigm that employs tree-based dissemination in the normal case and falls back to gossip only when necessary; and (3) constructing latency-aware dissemination trees.

Jane: The results demonstrate that, compared with gossip-based dissemination over the same topology, our prototype improves peak throughput by up to 7.26×.

Tom: System Context and Motivation Blockchain systems are typically deployed on a peer-to-peer (P2P) overlay network where P2P networks provide reliable, scalable and efficient message transmission services to upper layers Such features are essential for achieving scalability and simplifying the design, implementation, and maintenance of the upper-layer consensus protocol.

Title and authors: Jane: So, we’re looking at "Duet: Co-Optimizing P2P Message Propagation and Rotating-Leader Consensus," which sounds pretty technical, but the core idea seems to be tying the message flow directly into how leaders change in a rotating-leader consensus system.

Tom: Exactly. The authors are showing how you can improve both the leader rotation and the underlying P2P network efficiency simultaneously, rather than just tweaking one piece in isolation.

Lu: They focus on three main improvements: first, accelerating leader rotation; second, introducing a reliable broadcast paradigm that uses tree-based dissemination mostly, falling back to gossip only when it’s truly necessary; and third, building these latency-aware dissemination trees.

Meng: Tree-based dissemination sounds good for efficiency but sounds complicated to implement when things go wrong or the network topology changes frequently. What's the catch there?

Tom: Well, they address that by abstracting the network configuration as a Network Reconfiguration primitive, which means they record things like edge set and node set on-chain. This lets the system handle network evolution smoothly because consensus is required on updated information about those structures.

Jane: That sounds like a big piece of work for maintaining state consistency across the whole system, which is crucial when you're dealing with decentralized networks. So, how do they actually accelerate that leader rotation part?

Lu: For accelerating leader rotation, they look at the expected latency of a random cycle in the P2P network and use a greedy heuristic to construct this proposer ring. This aims to minimize total latency by iteratively selecting unvisited nodes closest to the current one.

Tom: So they’re not just picking any next node; they're using proximity based on what the network structure tells them about delays, which sounds much smarter than a simple round-robin approach.

The paper's summary: Jane: Okay, so to wrap up where we are, the paper shows that by combining these ideas—accelerating rotation and using a smarter way to spread messages—they can significantly improve performance compared to just relying on gossip.

Tom: They introduce a specific way of spreading proposals: in normal conditions, they use a tree-based broadcast, which is much faster than the usual gossip method, but they have this safety net where it falls back to gossip only when it's absolutely needed.

Lu: The construction of these latency-aware dissemination trees is key here. They select spanning trees based on network topology and latencies with a greedy algorithm that considers propagation delays, bandwidth constraints, and prioritizing upcoming proposers to minimize the maximum delay.

Meng: That sounds like they are trying to build the fastest possible route for a message, but I wonder how much overhead that tree construction itself adds when the network is already under stress.

Tom: They show empirical results on Amazon EC2 where their prototype, Duet, achieves up to seven point two six times higher throughput compared to gossip-based dissemination over the same network topology.

Jane: seven point two six times! That’s a huge jump in performance when you’re just looking at message delivery speed, and they found it especially pronounced when they combined the tree dissemination with that greedy proposer order.

Lu: They also noted that Duet shows better handling of skewed geographic placements compared to their baseline setups, which is important because real-world blockchain deployments often have uneven node distribution across the globe.

The paper's improvements: Tom: So, let’s talk about those three improvements again in more detail. First, we have accelerating leader rotation through that greedy heuristic we mentioned earlier.

Jane: And second is this reliable broadcast paradigm, which is using tree-based dissemination as the primary method and only resorting to gossip when it's necessary for safety or liveness.

Lu: And third, they specifically constructed latency-aware dissemination trees by solving a Traveling Salesman Problem on a weighted undirected graph to find cycles that minimize total weight.

Meng: Solving the Traveling Salesman Problem sounds computationally intensive, so how do they keep that construction fast enough to run in real-time within the consensus cycle?

Tom: They use a greedy algorithm for the tree construction, which considers propagation delays and bandwidth constraints while prioritizing upcoming proposers to minimize that maximum delay.

Jane: And this leads into what they call latency-aware dissemination trees, which are designed to reduce the maximum possible delay for message delivery within the network structure itself.

Lu: They also built in resilience by showing that the tree phase timer is critical for recovery; you have to wait out that tree timer before retransmission starts if internal relays fail.

Tom: And when a leader fails, they showed Duet extends the wait time beyond what vanilla Tendermint does by having the node wait out one tree timer before entering the propose-timer path itself.

Conclusion: Jane: So, to wrap this up, Duet is essentially a cross-layer design that optimizes rotating-leader consensus by using geographic locality for proposer ordering and cutting out redundant proposal transmissions in normal operation.

Tom: It successfully integrates these ideas into Tendermint and libp2p to achieve those significant throughput improvements we saw on Amazon EC2, proving the co-design approach works.

Lu: The framework they propose is promising because it’s a structure that can be extended to multileader or leaderless consensus protocols, which is a big area for future research.

Meng: From an engineering standpoint, the sensitivity of the gossip retransmission stage to how long that tree-phase timer runs means you have to tune things very carefully for optimal performance in production environments.

Lalam: I’m confident in this direction because the results show a seven point two six times throughput increase over gossip, and it works well under skewed geographic conditions, which is a real-world constraint.

Jane: It really shows how important it is to think about the network topology when you design consensus protocols, not just the logic of voting itself.

Tom: So we’ve seen how recording network data on-chain lets us build these smarter dissemination trees that beat standard gossip methods by a lot, and that’s what Duet does.

Shanghai Jiao Tong University · Shanghai Academy of Future Internet Technology · Beijing Academy of Blockchain and Edge Computing

cs.DC

Submitted: 2026-07-24

Updated: 2026-11-12

Code: https://github.com/Decentralized-Computing-Lab/Duet

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: In blockchain systems, peer-to-peer (P2P) overlay networks play a crucial role in providing reliable, scalable and efficient message-delivery services to upper layers.

Key concepts

Network Reconfiguration (NR) Primitive
This abstraction records network topologies and latencies directly on the blockchain. It allows the system to handle changes in the P2P network smoothly by requiring consensus on updated information, such as which nodes are connected or their current latency values, ensuring the protocol adapts to network evolution.
Tree-Based Dissemination
This is a highly efficient method for sending messages where information spreads like a tree structure. In normal operation, it's used to broadcast proposals quickly. The system constructs these trees based on network topology and latencies to minimize the maximum delay in reaching all necessary nodes.
Latency-Aware Dissemination Trees
These trees are constructed specifically to account for the time it takes for messages to travel across different network paths. The construction uses a greedy algorithm that prioritizes edges based on propagation delays and bandwidth constraints, aiming to minimize the overall maximum delay for proposal delivery.
Duet Prototype
Duet is a cross-layer design that co-optimizes rotating-leader consensus with P2P message propagation. It integrates three main improvements—accelerated leader rotation, reliable broadcast using trees, and latency awareness—into existing frameworks like Tendermint and libp2p to significantly boost performance.

Terminology

Summary

In blockchain systems, peer-to-peer (P2P) overlay networks play a crucial role in providing reliable, scalable and efficient message-delivery services to upper layers. p. 0

Where the paper says it

Duet: Co-Optimizing P2P Message Propagation and Rotating-Leader Consensus Yifeng Ye Shanghai Jiao Tong University China Rongji Huang Shanghai Academy of Future Internet Technology China Gerui Wang Beijing Academy of Blockchain and Edge Computing China Mingchao Wan Beijing Academy of Blockchain and Edge Computing China Yuxing Duan Beijing Academy of Blockchain and Edge Computing China Jingjing Zhang Fudan University China Shengyun Liu Shanghai Jiao Tong University China Abstract In blockchain systems, peer-to-peer (P2P) overlay networks play a crucial role in providing reliable, scalable and efficient message-delivery services to upper layers. However, the consensus layer and the underlying P2P network remain mutually opaque in existing blockchains, waiving the opportunity for further improvement.

With this support, we propose three improvements to rotating-leader consensus protocols and their underlying P2P networks: (1) accelerating leader rotation; (2) introducing a reliable-broadcast paradigm that employs tree-based dissemination in the normal case and falls back to gossip only when necessary; and (3) constructing latency-aware dissemination trees. p. 0
Where the paper says it

We therefore leverage this abstraction to record network topologies and latencies in a trusted and coordinated manner. With this support, we propose three improvements to rotating-leader consensus protocols and their underlying P2P networks: (1) accelerating leader rotation; (2) introducing a reliable-broadcast paradigm that employs tree-based dissemination in the normal case and falls back to gossip only when necessary; and (3) constructing latency-aware dissemination trees.

The results demonstrate that, compared with gossip-based dissemination over the same topology, our prototype improves peak throughput by up to 7. p. 0
Where the paper says it

We integrate the above ideas into Tendermint and libp2p, and conduct empirical evaluation on Amazon EC2 platform using up to 300 nodes distributed across 10 regions. The results demonstrate that, compared with gossip-based dissemination over the same topology, our prototype improves peak throughput by up to 7.

26×.

System Context and Motivation

Blockchain systems are typically deployed on a peer-to-peer (P2P) overlay network where P2P networks provide reliable, scalable and efficient message transmission services to upper layers Such features are essential for achieving scalability and simplifying the design, implementation, and maintenance of the upper-layer consensus protocol. Although most consensus protocols depend on timing assumptions about message delivery to ensure safety and/or liveness [25], the consensus layer and the underlying network are typically designed independently. p. 0

Where the paper says it

Such features are essential for achieving scalability and simplifying the design, implementation, and maintenance of the upper-layer consensus protocol, which reaches agreement on the order of blocks among nodes. Although most consensus protocols depend on timing assumptions about message delivery to ensure safety and/or liveness [25], the consensus layer and the underlying network are typically designed independently.

In contrast to other P2P applications, blockchain can naturally be abstracted as a state machine. p. 0
Where the paper says it

However, the consensus layer and the underlying P2P network remain mutually opaque in existing blockchains, waiving the opportunity for further improvement. In contrast to other P2P applications, blockchain can naturally be abstracted as a state machine.

Key Proposed Optimizations

The work introduces three specific improvements integrated into Tendermint and libp2p

  1. Accelerating leader rotation. p. 0

    Where the paper says it

    We therefore leverage this abstraction to record network topologies and latencies in a trusted and coordinated manner. With this support, we propose three improvements to rotating-leader consensus protocols and their underlying P2P networks: (1) accelerating leader rotation; (2) introducing a reliable-broadcast paradigm that employs tree-based dissemination in the normal case and falls back to gossip only when necessary; and (3) constructing latency-aware dissemination trees.

  2. Introducing a reliable-broadcast paradigm that employs tree-based dissemination in the normal case and falls back to gossip only when necessary. p. 0

    Where the paper says it

    We therefore leverage this abstraction to record network topologies and latencies in a trusted and coordinated manner. With this support, we propose three improvements to rotating-leader consensus protocols and their underlying P2P networks: (1) accelerating leader rotation; (2) introducing a reliable-broadcast paradigm that employs tree-based dissemination in the normal case and falls back to gossip only when necessary; and (3) constructing latency-aware dissemination trees.

  3. Constructing latency-aware dissemination trees. p. 0

    Where the paper says it

    We therefore leverage this abstraction to record network topologies and latencies in a trusted and coordinated manner. With this support, we propose three improvements to rotating-leader consensus protocols and their underlying P2P networks: (1) accelerating leader rotation; (2) introducing a reliable-broadcast paradigm that employs tree-based dissemination in the normal case and falls back to gossip only when necessary; and (3) constructing latency-aware dissemination trees.

Design of Network Abstraction

The paper abstracts network configuration as a Network Reconfiguration (NR) primitive, which involves recording network topologies and latencies on-chain This abstraction allows the system to handle network evolution smoothly by requiring consensus on updated information like edge set, node set, and evolution view identifier.

Optimization of Proposer Sequence

The authors address the overhead of rotating proposers by considering the expected latency of a random cycle in a P2P network They employ a greedy heuristic to construct the proposer ring, which aims to minimize total latency by iteratively selecting unvisited nodes closest to the current one.

Efficient Message Dissemination

The core innovation lies in splitting proposal dissemination into two stages: a highly-efficient tree-based broadcast stage and a gossip-based re-transmission stage The tree construction involves selecting spanning trees based on network topology and latencies to minimize maximum delay, using a greedy algorithm that considers propagation delays, bandwidth constraints, and prioritizing upcoming proposers.

Performance Evaluation Results

Empirical evaluation on Amazon EC2 demonstrated that Duet achieves up to 7.26× higher throughput than gossip-based dissemination over the same topology. p. 1

Where the paper says it

The results demonstrate that Duet achieves up to 7. 26× higher throughput than gossip-based dissemination over the same topology.

The performance gains are most pronounced when combining the tree dissemination with a greedy proposer order, reaching 21. p. 8
Where the paper says it

The modest drop from 𝑁 = 100 to 𝑁 = 300 likely reflects diminishing returns from deeper pipelining and increased gossip-based vote traffic as the number of nodes grows. 3 Ablation study To isolate how dissemination and proposer order affect performance, we combine each of three proposal dissemination methods {Gossip, K-ary tree, Duet tree} with each of two proposer orders {random, greedy}.

09k tps under these optimal conditions. p. 8
Where the paper says it

00k tps with 4196 ms latency, compared with 21. 09k tps and 3080 ms for Duet.

Duet also shows better exploitation of skewed geographic placements compared to baselines, maintaining a clear performance advantage in Ethereum-like deployments. p. 9
Where the paper says it

Full per-region counts are listed in Appendix B. Figure 10 shows that Duet retains a clear performance advantage under this skewed placement.

Resilience and Limitations

The analysis of resilience under failures indicates that the tree phase timer is critical for recovery, as it must be waited out before retransmission begins when internal relays fail. Leader failure experiments show that Duet extends the wait time beyond vanilla Tendermint by first waiting out one tree timer before entering the propose-timer path. p. 6

Where the paper says it

We assume 𝑡𝑚𝑟𝑟,tree is set to 2𝛿 and 𝑡𝑚𝑟𝑟,propose is set to 2Δ + 2𝛿, where 𝛿 ≤ Δ. In contrast to vanilla Tendermint, Duet introduces an additional 2×𝑡𝑚𝑟𝑟,tree = 4𝛿 waiting time per round to handle situations where correct nodes need to advance to the next round (e.

The gossip-based retransmission stage is highly sensitive to the duration of the tree-phase timer, requiring careful tuning for optimal performance. p. 9
Where the paper says it

09k tps, exceeding Gossip under the same greedy order by 6. 4 Timer Sensitivity The gossip-based retransmission stage of Duet is highly sensitive to the duration of the tree-phase timer.

Future Directions

The prototype is promising for further integration into multileader or leaderless consensus protocols. The authors suggest extending these ideas to structured overlays and exploring in-network consensus acceleration using programmable network devices These extensions aim to leverage the benefits of topology awareness and consensus semantics in more complex environments. p. 1

Where the paper says it

Simply recording the underlying network topology and latencies on-chain allows us to leverage this information to improve consensus and message routing in multiple ways: • We can optimize the proposer sequence and accelerate proposer rotation, thereby improving the throughput of RLC protocols; • By cleverly leveraging votes in the consensus protocol as acknowledgments, we can first employ an efficient (but not reliable) tree-based dissemination scheme to broadcast proposals and resort to gossip-based dissemination only as a fallback; and, • We can construct an efficient dissemination tree to minimize the cost of disseminating messages. We instantiate the above ideas by integrating them into a pipelined variant of Tendermint [12] and libp2p [4], a classical RLC protocol built upon P2P networks and a modular P2P networking framework adopted by many decentralized applications, respectively.

Conclusion

Duet is a cross-layer design for improving rotatingleader consensus by optimizing proposer sequence via geographic locality and eliminating redundant proposal transmissions in the normal case. It successfully integrates these ideas into Tendermint and libp2p, achieving significant throughput improvements over existing dissemination methods. p. 0

Where the paper says it

With this support, we propose three improvements to rotating-leader consensus protocols and their underlying P2P networks: (1) accelerating leader rotation; (2) introducing a reliable-broadcast paradigm that employs tree-based dissemination in the normal case and falls back to gossip only when necessary; and (3) constructing latency-aware dissemination trees. We integrate the above ideas into Tendermint and libp2p, and conduct empirical evaluation on Amazon EC2 platform using up to 300 nodes distributed across 10 regions.

The work provides a framework that can be extended to multileader or leaderless consensus protocols. p. 1
Where the paper says it

Simply recording the underlying network topology and latencies on-chain allows us to leverage this information to improve consensus and message routing in multiple ways: • We can optimize the proposer sequence and accelerate proposer rotation, thereby improving the throughput of RLC protocols; • By cleverly leveraging votes in the consensus protocol as acknowledgments, we can first employ an efficient (but not reliable) tree-based dissemination scheme to broadcast proposals and resort to gossip-based dissemination only as a fallback; and, • We can construct an efficient dissemination tree to minimize the cost of disseminating messages. We instantiate the above ideas by integrating them into a pipelined variant of Tendermint [12] and libp2p [4], a classical RLC protocol built upon P2P networks and a modular P2P networking framework adopted by many decentralized applications, respectively.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Accelerate leader rotation by recording network topologies and latencies on-chain, which allows for optimizing proposer sequence and accelerating proposer rotation to improve throughput of rotating-leader consensus protocols.

  2. Introduce a reliable broadcast paradigm that employs tree-based dissemination in the normal case and falls back to gossip only when necessary, which is achieved by constructing latency-aware dissemination trees.

  3. Construct latency-aware dissemination trees by solving a Traveling Salesman Problem on a weighted undirected graph to find cycles that minimize total weight, and then employing a greedy algorithm to select edges based on propagation delays and bandwidth constraints.

  4. Integrate these ideas into Tendermint and libp2p to create Duet, which allows for the optimization of message propagation in rotating-leader consensus protocols by co-designing them with P2P networks.

  5. Implement mechanisms that allow nodes to wait for sufficient time before proceeding to the prevote phase when a new height is entered, specifically by setting the duration of the propose timer to be sufficiently long, such that "the duration of trr propose of height h+1 must be sufficiently long to accommodate trr tree of height h, i.e., trr propose >= 2Δ + trr tree."

  6. Allow nodes to resume progress more slowly during a failed-leader window by first waiting out one tree timer before entering the propose-timer path, which is itself lengthened to cover the preceding height's tree timer.

Abstract

In blockchain systems, peer-to-peer (P2P) overlay networks play a crucial role in providing reliable, scalable and efficient message-delivery services to upper layers. However, the consensus layer and the underlying P2P network remain mutually opaque in existing blockchains, waiving the opportunity for further improvement. In contrast to other P2P applications, blockchain can naturally be abstracted as a state machine. We therefore leverage this abstraction to record network topologies and latencies in a trusted and coordinated manner. With this support, we propose three improvements to rotating-leader consensus protocols and their underlying P2P networks: (1) accelerating leader rotation; (2) introducing a reliable-broadcast paradigm that employs tree-based dissemination in the normal case and falls back to gossip only when necessary; and (3) constructing latency-aware dissemination trees. We integrate the above ideas into Tendermint and libp2p, and conduct empirical evaluation on Amazon EC2 platform using up to 300 nodes distributed across 10 regions. The results demonstrate that, compared with gossip-based dissemination over the same topology, our prototype improves peak throughput by up to 7.26 times.

Sources

Related papers