Game Server Distribution Project

1
Game Server Distribution Project

• Paras Shelawala
• John Tam
• Will Liu
• Stephen Yam

2
Game Requirements

• Real-Time Processing
• Many Small Packets
• Many Users

3
Peer-to-Peer Architecture

• All systems on LAN
• All synchronized
• Not dependent on single point of failure
• Limited number of users
• Broadcast packets

Ethernet
4
Client-Server Architecture

• Centralized processing of game state
• Prevalent in modern games such as Quake-3,
  Half-life, etc.
• Performance dependent on Internet conditions (WAN
  network)

5
Bottlenecks in Client-Server Architecture

• Computational
• All processing occurs on a single system
• Network
• Connection speed relies on slowest link between a
  client and the server
• Server network is limited by bandwidth

6
Solution to Computational Bottlenecks

• Distribute server load over a cluster of machines
• Connection speed within cluster must ensure
  synchronization between all machines
• Maximize processing on the client side

7
Solution to Computational Bottlenecks

• This solution has
• several difficulties

8
Difficulties withDistributing Server Load

• Need an intelligent and efficient method to
  distribute load across multiple servers
• Minimize communication between servers
• Fault-tolerance within cluster
• Synchronization of servers must be maintained
• Difficulties are Game Specific

9
Possible Solutionsto Network Bottleneck

• More significant than computational bottlenecks
• Cause of poor game play due to packet delay/loss
  consequences
• Cause of this problem is a temporary slow link
  somewhere across the network

10
Protocol Issues

• TCP is too slow too much overhead for real-time
  games
• UDP is fast, but has no guarantee of packet
  delivery
• Goal is to ensure a stable stream of packets
  between client and server using UDP

11
Experiments

• Tested Quake-III Arena
• Blah data
• Conclusion
• Quake-III is low-bandwidth, but has a high
  packet-generation frequency

12
Solutions toNetwork Bottleneck

• Use a remote repeating node which creates an
  alternate path for which packets can be
  transmitted/received
• Requires fast and stable connection between
  repeater and server cluster

13
Solutions toNetwork Bottleneck

• Separate clusters of servers with fast
  inter-connection
• Clients re-assess connection status to all
  clusters, determines if hand-off to another
  cluster would improve performance
• Both clients and servers keep track of
  performance data and determine best configuration

14
Solutions toNetwork Bottleneck
15
Suggestions/Comments