Improving Quality of Service for Users of Leaderless DAG-Based Distributed Ledgers

Andrew Cullen, Lianna Zhao, Luigi Vigneri, Robert Shorten

TL;DR

This paper tackles the challenge of ensuring a good user experience in DAG-based distributed ledgers, like IOTA, by proposing a novel user-node interaction mechanism. It introduces a system where nodes advertise their Quality of Service (QoS) and users probabilistically select nodes based on this information, leading to more balanced traffic and improved transaction speeds.

What Problem Does This Paper Address?

The widespread adoption of distributed ledger technologies (DLTs) hinges on their usability, especially for users who don't want to run full node infrastructure. While solutions exist for traditional blockchain architectures, leaderless DAG-based DLTs (like IOTA) present unique challenges. These systems lack a shared mempool and allow parallel block creation, making existing user-node interaction mechanisms unsuitable. This leads to variable transaction completion times, unpredictable fees, and potential for user exploitation. The paper aims to bridge this gap by designing a mechanism that ensures a consistently good Quality of Service (QoS) for users and fair resource utilization for nodes in these DAG-based DLTs.

Methods

The core of the paper's contribution lies in its proposed user-node interaction mechanism. This mechanism operates on the principle of stochastic node selection.

  • Node-side: Nodes continuously monitor and advertise their expected Quality of Service (QoS), primarily focusing on transaction delay and fees. They use an AIMD (Additive Increase Multiplicative Decrease) algorithm to dynamically adjust their transaction issuing rates based on local congestion, and this rate influences the advertised delay. For fees, a simple proportional controller is used to increase fees when delays exceed a setpoint.
  • User-side: Users receive these QoS indicators from nodes. They then construct a probability distribution over the available nodes, where the probability of selecting a node is inversely proportional to its advertised delay and fees (and directly proportional to its reputation). This probabilistic selection acts as a distributed load balancer.

The paper also introduces and compares several user selection policies:

  • URNS (Uniform Random Node Selection): A naive baseline where users pick nodes randomly.
  • RBNS (Reputation-Based Node Selection): Users select nodes based on their known reputation.
  • DBNS (Delay-Based Node Selection): Users select nodes based on a combination of reputation and advertised delay.
  • DBNS+ (Delay and Fee-Based Node Selection): The most sophisticated policy, incorporating both delay and fees, and a user-defined cost threshold to manage demand during congestion.

Datasets

Not discussed in the paper. The authors developed a Python simulator based on IOTA's DAG-based distributed ledger for their experiments.

Key Insights and Ideas

  • DAGs require a different approach: The fundamental difference in DAGs (parallel processing, no shared mempool) necessitates a departure from traditional blockchain user-node interaction models. Users must actively choose nodes, making load balancing crucial.
  • QoS indicators are key: Nodes advertising their expected delay and fees provide essential feedback for users to make informed decisions, enabling a distributed load balancing system.
  • Probabilistic selection is effective: By having users probabilistically select nodes based on QoS, traffic can be balanced across the network, preventing individual nodes from becoming overwhelmed and ensuring better overall service.
  • Fees as a congestion control mechanism: Incorporating transaction fees into the user's decision-making process (DBNS+) is crucial for managing demand during high congestion, preventing unbounded delays.
  • User experience is not just about average delay: While average delay is important, the probability of experiencing a very bad delay (e.g., >20 seconds) is a critical factor for user adoption, and DBNS+ significantly reduces this probability.

Key Results

The simulations demonstrated a clear progression in performance across the tested policies:

  • URNS performed poorly, with low-reputation nodes being overwhelmed and high-reputation nodes underutilized, leading to high and variable delays.
  • RBNS showed improvement by favoring reputable nodes, but still suffered from load imbalances.
  • DBNS further improved performance by incorporating delay feedback, leading to more balanced traffic and reduced fluctuations.
  • DBNS+ emerged as the most effective policy, particularly under high load (120% capacity). It stabilized delays by increasing fees in response to congestion, effectively managing user demand and significantly reducing the probability of experiencing excessively long delays (e.g., 0.01 probability of delay > 20 seconds at 120% capacity, compared to 0.32 for URNS).

Limitations and Future Work

The authors acknowledge several limitations:

  • Simplifying assumptions: The model assumes users have full knowledge of node statistics, which might be unfeasible in large networks.
  • Limited scope of QoS: The primary focus was on delay and fees, while other QoS aspects were not deeply explored.
  • No adversarial behavior analysis: While the mechanism is designed to be robust, a full analysis of adversarial behavior (e.g., spamming attacks) was deferred.

Future work suggested includes:

  • Investigating user behavior when they only query a subset of nodes.
  • Exploring more complex node behaviors, such as collaboration and fee sharing.
  • Conducting more extensive simulations with a wider range of parameters.
  • Performing detailed attack analyses for the proposed policies.

Summary updated

BibTeX
@article{cullen2023improving,
  author = {Andrew Cullen and Lianna Zhao and Luigi Vigneri and Robert Shorten},
  title = {{Improving Quality of Service for Users of Leaderless DAG-Based Distributed Ledgers}},
  journal = {Distributed Ledger Technol. Res. Pract},
  year = {2023},
  doi = {10.1145/3604942},
}

Loading related content...

Figure