A Study of Network Stack Latency for Game Servers

P. Emmerich, D. Raumer, F. Wohlfart and G. Carle, "A Study of Network Stack Latency for Game Servers", in NetGames '14, Nagoya, Japan, 2014, p. 6.

This paper outlines what steps a server-client model for gaming servers should take to reduce the latency of their games. Emmerich et. al. note that the biggest share of delay often comes from the network between the server and the player and that this delay is often within tens of milliseconds. If a developer uses the regular network stack available from their OS, the delay can shoot into the hundreds of milliseconds.

Emmerich et. al. demonstrate that the following practices can result in additional, undesired latency:

Can gamers detect cloud delay?

K. Raaen, R. Eg and C. Griwodz, "Can Gamers Detect Cloud Delay?", in NetGames '14, Nagoya, Japan, 2014, p. 3.

The main goal of this paper was to empirically establish thresholds for detectable motor-visual delay in gaming. Prior studies were conducted as highly controlled tests; as a result it was hard to find a general case as we would want for gaming. Kjetil et. al. found that by designing an experiment where the test subject could adjust their motor-visual delay until it was no longer detectable, the mean range settled on was between 50-91ms. They also found a single test subject could discern differences up to 24ms, which shows that outside of the most performant players of any game developers need only aim for latencies near that 50ms mark.

The Revolution of StarCraft Network Traffic

C. Lee, "The revolution of StarCraft network traffic", in NetGames '12, 2012.

In this paper the author sought to use prior research on StarCraft's usage of network traffic alongside captured packets of StarCraft II, its sequel, to highlight the changes in network traffic in the RTS genre of games. RTS games, or Real-Time Strategy games, involve two to eight players in a match where each player creates and micromanages an army to take down the opposing player's base. These games, like fighting games, require lower latencies as the smallest of errors can result in a lost army and hence a lost game. In this paper Lee found that while the original StarCraft, released in 1998, used a peer-to-peer architecture for its network traffic its successor in Starcraft II transitioned to a server-client model to handle traffic. The game communicates with Battle.net servers in the background using both TCP and UDP, but when the game is in progress only UDP packets are sent between players and their respective server. They also found that StarCraft's original peer-to-peer model utilized more outgoing bandwidth as the number of players in the game increased whereas with StarCraft II's server-client model the outgoing bandwidth from any individual player was constant.