Multicast Tree Reconfiguration in Distributed Interactive Applications

1
Multicast Tree Reconfiguration in
DistributedInteractive Applications

• Pål Halvorsen1,2, Knut-Helge Vik1 and Carsten
  Griwodz1,2
• 1Department of Informatics, University of Oslo,
  Norway
• 2Simula Research Laboratory, Norway

2
Game environment
Real-World Proximity
Virtual World Proximity

• Typical massive multiplayer online games today
• Central server-based
• Experience high latency
• Physical world and virtual world locality are
  unrelated

3
Game environment
Asia
Europe
North America

• Example 1 hour trace of one region of Anarchy
  Online
• Here Most action in Europe
• At other times of day the center of action shifts
• Average latency could be reduced considerably

4
Reducing the worst-case latency

• For average area of interest find a node that
  improves average latency
• Let this leader node handle state on behalf of
  the server
• Variation of the central server approach
• Remaining problem
• Network utilization
• This leader node is probably less powerful than
  the server
• Games traffic is not adaptive

S
L
Transfer state
5
Non-adaptive games traffic

• Games traffic UDP or TCP
• UDP is not adaptive
• TCP games traffic is not either !!!

Number of packetsper round-trip time

• Games connections are so thin that TCPs
  congestion control does not apply
• We should conserve network resources

6
Reducing tree cost

• Tree structure saves resources
• Alleviates the communication overhead
• Best results Minimum Spanning Tree (MST) or
  Steiner Minimum Tree (SMT)
• Tree computation necessary for each join and
  leave
• MST and SMT computations are
• Too slow
• Usually centralized
• Need fast heuristics

S
L
Transfer state
Lowerworst-casedelay
L
n
New node enters
7
Various Join operations
all LEAVE operations remain in the graph until
degree is 2 or less
8
Effects of Join operations

• Tested on several topologies generated using
  BRITE
• Here several iterations of groups with
  Zipf-distributed popularities

Time for entering a group
Sum of all edge delays

• When groups are small
• complex algorithms are faster than simple ones
• complex algorithms provide better results

9
Effects of Join operations

• Tested on several topologies generated using
  BRITE
• Here several iterations of groups with
  Zipf-distributed popularities

Time for entering a group
Sum of all edge delays

• When groups are large and constantly changing
• connect best takes too long without any
  performance gain
• cost of more complex algorithms decreases quickly

10
Effects of Join operations

• Tested on several topologies generated using
  BRITE
• Here several iterations of groups with
  Zipf-distributed popularities

Time for entering a group
Sum of all edge delays

• When groups are very large and constantly
  changing
• nearly all nodes remain in the graph because we
• do allow nodes to leave only when they 2 or
  fewer neighbors

11
Conclusions

• Join operation by itself is not sufficient to
  define a tree
• Fast join operation is preferable for small
  groups
• Cost-conscious join operations are preferable for
  large groups
• Currently investigating
• Minimum Spanning Tree and Steiner Tree Heuristics
• Goal is to evaluate some that areDistributed and
  Dynamic
• Compare then with simple Join approaches
• Future work
• Resilience through pre-defined backup paths