Bandwidth requirement and state consistency in three multiplayer game architectures

Joseph D. Pellegrino, Constantinos DovrolisNETGAMES 2003 · DOI

TL;DR

This paper tackles the challenge of balancing bandwidth efficiency and state consistency in multiplayer games. The authors propose a novel "Peer-to-Peer with Central Arbiter" (PP-CA) architecture that aims to combine the low latency of peer-to-peer (P2P) with the simpler consistency management of client-server (CS) models, showing it significantly reduces server bandwidth requirements compared to traditional CS architectures.

What Problem Does This Paper Address?

Multiplayer games, especially fast-paced ones like first-person shooters, face a fundamental trade-off between keeping all players' game states synchronized (state consistency) and minimizing the amount of data that needs to be sent over the network (bandwidth requirement). Traditional Client-Server (CS) architectures centralize game state management on a server, which simplifies consistency but can lead to a server bottleneck as the number of players grows, due to its quadratic bandwidth scaling. Peer-to-Peer (P2P) architectures distribute this load, removing the server bottleneck but introducing complex distributed protocols for consistency and potentially higher bandwidth demands on individual players. This paper seeks a middle ground, aiming for the scalability of P2P while retaining simpler consistency management.

Methods

The authors analyze and compare three architectures:

  • Client-Server (CS): A central server manages game state and resolves inconsistencies. Players send updates to the server, which then broadcasts the global state back to all players.
  • Peer-to-Peer (P2P): Each player manages its own game state and communicates directly with all other players. Inconsistencies are resolved through distributed agreement protocols.
  • Peer-to-Peer with Central Arbiter (PP-CA): This is the proposed novel architecture. Players communicate directly with each other (like P2P) but also send their updates to a central arbiter. The arbiter simulates the global state and detects inconsistencies. Crucially, the arbiter only intervenes (sends corrected updates) when an inconsistency is detected, unlike the CS server which constantly broadcasts global state.

The analysis focuses on two key metrics: bandwidth requirement (at the server and players) and inconsistency resolution latency (specifically, the Maximum Inconsistency Period - MIP). They also conducted experiments using a modified open-source game, BZFlag, to measure these metrics.

Datasets

The experiments were conducted using the open-source 3D tank battle game "BZFlag". The authors modified the game to implement variations of the CS, P2P, and PP-CA architectures. Measurements were taken on Pentium-III PCs connected via an Ethernet switch, with network delays between player machines being minimal (less than a millisecond). Packet capture was performed using "tcpdump".

Key Insights and Ideas

  • CS Bandwidth Bottleneck: The CS architecture's bandwidth requirement at the server scales quadratically with the number of players, making it inherently unscalable for large player counts.
  • P2P Complexity: While P2P removes the server bottleneck, it introduces complex distributed protocols for state consistency and can place a higher bandwidth burden on individual players.
  • PP-CA: The Sweet Spot: The PP-CA architecture offers a compelling compromise. By having a central arbiter listen to all updates but only act when inconsistencies arise, it drastically reduces the arbiter's bandwidth requirement compared to a CS server.
  • Inconsistency Frequency Matters: The effectiveness of PP-CA hinges on the frequency of inconsistencies. If inconsistencies are rare, the arbiter's bandwidth is very low. If they are frequent, the arbiter's bandwidth can approach that of a CS server.
  • Latency Trade-offs: PP-CA inherits the low player-to-player communication latency of P2P, while its inconsistency resolution latency is comparable to CS, depending on the round-trip times.

Key Results

  • Bandwidth: The PP-CA architecture shows significantly lower server traffic compared to CS, with its traffic scaling linearly with the number of players, even under a 50% artificial inconsistency rate. In contrast, CS server traffic scales quadratically.
  • Inconsistency Resolution Latency (MIP): For a 4-player CS game, the MIP was estimated at around 27ms, with actual measurements showing most inconsistencies resolved within 25ms. For a 2-player PP-CA game, the MIP was estimated at roughly 29ms, with actual measurements showing resolution within 28ms. These figures suggest PP-CA can achieve comparable or slightly higher inconsistency resolution latency than CS, depending on the specific configuration.

Limitations and Future Work

  • CPU Load: The paper acknowledges that while PP-CA reduces network bandwidth requirements for the central entity, it does not reduce its processing load. The central arbiter still needs to simulate the entire game state, meaning PP-CA might not be scalable if the bottleneck is CPU power rather than network bandwidth.
  • Inconsistency Rate: The performance of PP-CA is highly dependent on the frequency of state inconsistencies. The paper explores a 50% artificial inconsistency rate, but real-world game scenarios might vary.
  • Future Work: The authors suggest that further research could explore how to optimize the PP-CA architecture for different types of games and how to manage the CPU load of the central arbiter more effectively. They also hint at the potential for PP-CA to facilitate game provider monitoring for accounting or pricing, which could be a direction for future exploration.

Summary updated

BibTeX
@inproceedings{pellegrino2003bandwidth,
  author = {Joseph D. Pellegrino and Constantinos Dovrolis},
  title = {{Bandwidth requirement and state consistency in three multiplayer game architectures}},
  booktitle = {NETGAMES},
  year = {2003},
  doi = {10.1145/963900.963905},
}
Figure