Skip to content
Crypto.world

Research

Decentralized network congestion control for DAG-based distributed ledger system

Blockchain networks have two built-in brakes on spam: competitive PoW and per-transaction fees. Blockless DAG ledgers such as IOTA deliberately drop both to achieve high throughput and near-instant…

Decentralized network congestion control for DAG-based distributed ledger system
arXiv
2609.09961
Published
2026-09-09
Authors
Mayank Pandey, Rachit Agarwal, Sandeep Kumar Shukla, Nishchal Kumar Verma

AI summary

Overview

Research area: Congestion control for decentralized networks, specifically transaction-spam mitigation in directed acyclic graph (DAG)-based distributed ledger technology (DLT). The work sits at the intersection of distributed systems, consensus/PoW design, and game theory. Published on arXiv under cs.CR (Cryptography and Security), arXiv:2609.09961v1, 09 Sep 2026, CC BY 4.0, by Mayank Pandey, Rachit Agarwal, Sandeep Kumar Shukla and Nishchal Kumar Verma (IIT Kanpur).

Technical level: Intermediate. The core idea is intuitive, but the paper formalizes it with quadratic reputation functions, exponential PoW cost scaling, and a non-cooperative game with a Nash-equilibrium argument.

Scope (one sentence): The paper proposes and analytically validates a node-specific, behavior-based variable proof-of-work difficulty rule that penalizes transaction spamming in blockless DAG ledgers without transaction fees and without sacrificing throughput.

What This Paper Is About

Blockchain networks have two built-in brakes on spam: competitive PoW and per-transaction fees. Blockless DAG ledgers such as IOTA deliberately drop both to achieve high throughput and near-instant confirmation, which leaves them open to a single well-resourced node (or a colluding group) flooding the ledger with transactions and crowding out everyone else. The authors ask how to discourage that spamming using only the difficulty of the small "proof of attachment" PoW that every DAG transaction already requires, so that every participant keeps an equal opportunity to transact.

Key Contributions

  1. A congestion-control model for DAG-based DLT networks that assigns each node a variable, node-specific PoW difficulty derived from two reputation components: the share of tokens the node transfers and the number of transactions it issues in the preceding period. The model aims to give all stakeholders equal opportunity regardless of computational resources and to penalize nodes that exceed the prescribed transaction count.
  2. A demonstration that this difficulty schedule makes the prescribed behaviour rational, by formulating transaction issuance as a non-cooperative game and proving the existence of a Nash equilibrium.
  3. A claim of continuity of prescribed behaviour when the network's user composition changes, established through a uniform change in the Nash equilibrium when nodes join or leave and the network capacity Λ is updated.
  4. A resource-penalty characterization showing that exceeding the fair-share transaction limit by even one transaction imposes a disproportionately larger computational burden in the following period.

Main Findings

  • Fair-share point is the local minimum: For 0 ≤ w_i(CT) ≤ Λ, the ratio of assigned difficulty to transaction frequency, d_i(CT)/w_i(CT), attains its local minimum at w_i(CT) = K_cross, the average per-user transaction capacity. A rational node therefore has an incentive to stay at or below K_cross.
  • Cost of over-issuing is disproportionate: Increasing difficulty by 1 doubles the average computational requirement and decreasing by 1 halves it, since Υ_i(t) ∝ 2^{d_i(t)}. The paper's equation (10) states that the resource requirement for issuing one transaction above K_cross is approximately 2^(K_cross²) times the requirement at K_cross.
  • Network capacity is partitioned, not auctioned: Capacity is modelled as Λ = n·K_cross + K_add, with K_add << K_cross treated as residual system capacity. The parameters K_0, K_1 and K_2 of the quadratic reputation function are chosen so that K_cross is the only viable (non-negative) root.
  • Spending reputation is bounded: The token-transfer component f_i^B(CT) lies in [0, 1], equals 1 when the node transferred no tokens in the period, and equals 0 only in the unlikely case where the node accounted for all token transfers in the network.
  • DAG ledgers are more exposed than blockchains: Because nodes add and validate transactions cooperatively rather than competing for blocks, and because fees are absent, DAG-based DLT lacks the equivalent of blockchain's spam deterrents and needs a dedicated mechanism.
  • IOTA's existing adaptive PoW is insufficient: The paper states that IOTA's variable difficulty, d_i(t) = d_0 + floor(γ_i × a_i(t)) with γ_i tied to mana, is formulated to prevent spamming only, and that its methodology is not sufficient to completely deter users from spamming the Tangle.
  • Simulation results are not reported in the available excerpt: Section 5 is described as containing numerical examples and simulation, but the supplied content is truncated before those results, so specific benchmark values, node counts or throughput figures are not reported here.

Methodology in Plain English

The network is modelled as two coupled graphs: a bidirectional graph of participating nodes and a DAG-shaped graph of transactions, where each new transaction must attach to a pre-decided number of existing unverified transactions ("tips"), taken as 2 in this work. Time is split into uniform periods of length T, and each node's reputation for the coming period is computed from its activity in the previous one.

Reputation has two parts. The first measures how much of the network's token flow the node accounted for, which rewards real spending but caps the reward. The second counts how many transactions the node issued and is deliberately shaped as a downward-opening quadratic: it rises as a node becomes active, peaks at the fair-share level K_cross, and then falls. The assigned PoW difficulty combines a base level d_0, a linear term from the derivative of the frequency-reputation function (applied up to K_cross), and a quadratic penalty term scaled by the spending reputation (applied beyond K_cross), rounded with floor and ceiling operations.

Because difficulty enters the brute-force hash search exponentially, that schedule translates into a computational cost that climbs gently up to the fair share and then explodes. The authors then treat each node's choice of how many transactions to issue as a non-cooperative game and show that a Nash equilibrium exists at the prescribed behaviour, and that the equilibrium shifts uniformly when the set of nodes changes, so newcomers and leavers do not break the incentive structure.

Why This Matters

Research impact. Congestion control for centralized communication systems is a mature field, while congestion control for DLT networks — and for DAG-based networks in particular — is described by the authors as relatively recent and underexplored. The paper reframes spam prevention as an incentive-design problem rather than a fee-market problem, and offers a proof-based argument rather than only a heuristic.

Real-world applications:

  • Fee-less machine-to-machine micropayments, including transactions from low-resource devices such as IoT sensors, which the paper notes are isolated by blockchain-style competitive consensus.
  • Networks under load from organic growth in the user base, where more transaction-generating points appear without any malicious intent.
  • Smart-contract platforms experiencing transaction surges when a new token launches and assets are added.
  • Defence against colluding groups of high-resource nodes attempting to get invalid transactions such as double-spends verified by flooding the ledger.

Industry relevance. The proposal targets the design of the consensus layer of DAG-based ledgers (the paper references IOTA, Graphchain, Phantom, Meshcash, Spectre, G-IOTA, E-IOTA and Hashgraph) and claims applicability to other PoW-based DLT networks with modifications, which matters to protocol designers who want spam resistance without reintroducing fees or sacrificing the throughput that makes DAG ledgers attractive.

Future Directions

  • Empirical validation: the paper positions Section 5 as numerical examples and simulation, but the excerpt does not report the parameter values, node counts or measured outcomes, so reproducing the simulation under stated network sizes is an open task.
  • Choosing K_0, K_1 and K_2 in practice: the paper constrains them so that K_cross is the valid root, but how these should be set and re-tuned for real, fluctuating capacity Λ is not resolved in the supplied content.
  • Adversarial behaviour beyond spam: the model addresses transaction flooding, while collusion aimed at double-spending and the interaction with tip-selection manipulation are identified as concerns rather than fully modelled.
  • Portability to non-DAG and non-PoW networks: the authors claim the model is viable for other PoW-based DLT networks with modifications, but those modifications and their game-theoretic guarantees remain to be worked out.

Target Audience

Protocol designers and researchers working on DAG-based ledgers, IOTA-style tip-selection systems, or fee-less transaction networks; graduate students and academics in distributed systems and blockchain security who want a game-theoretic treatment of spam resistance; and engineers at DLT projects evaluating non-fee-based congestion-control mechanisms. Readers should be comfortable with PoW difficulty notation, basic graph formalisms, and Nash equilibrium concepts.

Authors’ abstract

We propose a variable and behavior-based node-specific proof-of-work (PoW) model for a directed acyclic graph (DAG)-based distributed ledger technology (DLT) network to mitigate decentralized network congestion control. Network congestion control for centralized communication systems is an established field of study, with detailed and continuous research being done on the subject. However, attention to congestion control in decentralized networks is relatively recent and…

The full abstract on arXiv
Read the original paper