Result Summary

<aside> 💡

The table below provides a quick validation summary of the main experiment results before going through the detailed setup and scenarios.

</aside>

Feature/Experiment	Status	Summary
Delivery delay regression	✅ Validated	No obvious regression observed with QUIC or YAMUX.
Priority queue behavior	✅ Validated	Medium/Low queues drop on overflow; High queue overflow disconnects peers.
Control traffic reliability	✅ Validated	High-priority control traffic remains active while Low queue is under pressure.
Slow peer penalty	✅ Validated	Penalty is applied on drops and decays after traffic stops.
Parameter sensitivity	✅ Validated	`decay = 0.2` looks safer for bursts; `decay = 0.9` looks more aggressive.
Score impact with weight `-0.05`	✅ Validated	Recoverable bursts cause only temporary score disturbance.
Uplink-limited sender behavior	⚠️ Issue found	Sender-side uplink bottlenecks can falsely penalize healthy remote peers.

Context

This page presents the experiment scenarios and results for testing the new priority queue features introduced.

The tested PR is:

https://github.com/vacp2p/nim-libp2p/pull/2429

The main feature to evaluate is the new outbound PubSub priority queue system. With this change, outbound PubSub traffic is split into three priority queues:

Queue	Traffic type	Expected behavior
High priority	Control messages and subscriptions	If the queue overflows, the peer is disconnected
Medium priority	Local publishes	If the queue overflows, messages are dropped
Low priority	Relayed messages and IWANT-response traffic	If the queue overflows, messages are dropped

A slowPeerPenalty mechanism is also added for peers that repeatedly overflow Medium or Low queues. This penalty can eventually lead to pruning, depending on the value of slowPeerPenaltyWeight, because this parameter controls how much the penalty affects the global peer score.

The related parameters are important for interpreting the results:

Parameter	Role
`slowPeerPenaltyWeight`	Controls how strongly the slow peer penalty affects the global peer score
`slowPeerPenaltyThreshold`	Defines the threshold after which the penalty starts affecting the score
`slowPeerPenaltyDecay`	Controls how quickly the penalty is reduced over time

In other words, the penalty does not affect the global score before it passes slowPeerPenaltyThreshold. Once it passes the threshold, the impact on the score depends on slowPeerPenaltyWeight. Then, over time, slowPeerPenaltyDecay controls how fast the accumulated penalty is removed.

With that in mind, and based on the validation request, the main objectives of these experiments are to:

<aside> 💡

validate that there is no delivery-delay regression with QUIC and YAMUX;
validate that control messages and subscription traffic remain reliable;
validate that queues behave as expected:
- High-priority queue overflow causes peer disconnection;
- Medium and Low queue overflow cause message drops;
- with the upstream default slowPeerPenaltyWeight = 0.0, Medium/Low overflow should not affect the global peer score;
- with slowPeerPenaltyWeight = -0.05, repeated Medium/Low overflow can affect the global score and can eventually lead to pruning;
validate that slowPeerPenalty is applied and decays over time;
check whether the values used for slowPeerPenaltyWeight, slowPeerPenaltyThreshold, and slowPeerPenaltyDecay are too aggressive. </aside>

For the purpose of these tests, a new test node and a new publisher were added.

The code and Dockerfiles are available here: https://github.com/vacp2p/dst-libp2p-test-node/tree/Mamoutou/gossip_queues