Adaptive Event Dissemination for Peer-to-Peer Multiplayer Online Games

In this paper we show that gossip algorithms may be effectively used to disseminate game events in Peer-to-Peer (P2P) Multiplayer Online Games (MOGs). Game events are disseminated through an overlay network. The proposed scheme exploits the typical behavior of players to tune the data dissemination. In fact, it is well known that users playing a MOG typically generate game events at a rate that can be approximated using some (game dependent) probability distribution. Hence, as soon as a given node experiences a reception rate, for messages coming from a given peer, which is lower than expected, it can send a stimulus to the neighbor that usually forwards these messages, asking it to increase its dissemination probability. Three variants of this approach will be studied. According to the first one, upon reception of a stimulus from a neighbor, a peer increases its dissemination probability towards that node irrespectively from the sender. In the second protocol a peer increases only the dissemination probability for a given sender towards all its neighbors. Finally, the third protocol takes into consideration both the sender and the neighbor in order to decide how to increase the dissemination probability. We performed extensive simulations to assess the efficacy of the proposed scheme, and based on the simulation results we compare the different dissemination protocols. The results confirm that adaptive gossip schemes are indeed effective and deserve further investigation.

💡 Research Summary

The paper addresses the problem of efficiently disseminating game events in peer‑to‑peer (P2P) multiplayer online games (MOGs). Traditional approaches such as static spanning trees are brittle in highly dynamic P2P overlays, while pure broadcast floods generate excessive redundant traffic. The authors propose an adaptive gossip scheme that exploits the statistical regularity of player‑generated events: each player’s inter‑event time follows a game‑specific probability distribution, allowing a node to estimate the expected reception rate from each peer.

The core idea is a stimulus‑based feedback loop. Whenever a node p observes that the rate of events arriving from a particular peer q falls below a threshold (α·ω, where ω is the expected generation rate), p sends a “stimulus” message to q. Upon receiving a stimulus, q temporarily raises its forwarding probability toward p (or toward a specific set of neighbors) by a fixed increment σ; this increase then decays linearly to the baseline value υ₀ after a configurable interval Δ. The decay prevents the system from converging to a pure broadcast (probability = 1) and limits the overhead of persistent high‑probability forwarding.

Three variants of the adaptive gossip protocol are defined:

1. Receiver‑oriented stimulus – The stimulus raises a single per‑neighbor forwarding probability υₙ for all senders. This is the simplest implementation; any stimulus from any source causes the neighbor to become more aggressive in forwarding all messages to the requesting node.

2. Generator‑oriented stimulus – Each source node s has its own forwarding probability γₛ. A stimulus from a receiver only increases γₛ, i.e., the probability that messages originated by s are forwarded, regardless of the destination. This provides finer granularity than the first variant while keeping state proportional to the number of peers.

3. Generator‑and‑receiver‑oriented stimulus – The most detailed scheme maintains a matrix of probabilities γₛ,ₙ for each (source, destination) pair. A stimulus affects only the entry corresponding to the specific source whose events are under‑delivered to the specific neighbor that issued the stimulus. This yields the highest precision but incurs O(N²) state overhead.

All three protocols share a common gossip core (Algorithm 1): when a node generates or receives a new event, it forwards the event to each neighbor (except the one it received it from) with probability given by the appropriate threshold (υₙ, γₛ, or γₛ,ₙ). Duplicate events are filtered using an LRU cache.

The authors evaluate the protocols via extensive simulations. They consider both scale‑free and uniform‑degree overlay topologies, varying the number of nodes, average degree, and event generation rates. Performance metrics include total transmitted messages (traffic overhead), average delivery latency, and delivery ratio (percentage of nodes that receive each event). Results show that adaptive gossip reduces traffic by 20‑35 % compared with a non‑adaptive baseline while maintaining delivery ratios above 95 %. The third variant (matrix‑based) achieves the best trade‑off between latency and overhead but requires roughly 1.5 × the memory of the second variant. In scale‑free networks the scheme also mitigates hub overload by dynamically increasing forwarding probabilities of under‑served peers.

The paper concludes that adaptive gossip, driven by simple stimulus feedback, effectively balances responsiveness, scalability, and bandwidth consumption in P2P MOGs. It highlights several open issues: (i) automatic tuning of stimulus parameters (σ, Δ, α) in live environments, (ii) robustness to real‑world network variability (packet loss, jitter), (iii) security concerns such as malicious stimulus injection, and (iv) the need for empirical validation on actual game deployments. Nonetheless, the work provides a solid foundation for decentralized, self‑adjusting event dissemination mechanisms in large‑scale online games.