Supporting Spectators in Online Multiplayer Games

1
Supporting Spectators in Online Multiplayer Games

• Ashwin Bharambe (CMU)
• Venkat Padmanabhan (Microsoft Research)
• Srinivasan Seshan (CMU)
• HotNets 2004

2
Spectators in Online Games

• Why?
• natural extension of TV model for physical games
• greater participation, opportunity to improve
  skills
• advertising opportunities
• Spectating
• enabling users to watch the proceedings of a game
• spectators gtgt players
• Cheering
• enabling spectators to cheer the players
• audio-based cheering may be most natural
• cheers are heard by players and other spectators

3
(No Transcript)
4
Terminology

• Virtual game world divided into game views
• Each game view consists of a set of entities
• Each entity has several attributes
• An update of the game view comprises a frame

5
Spectating versus A/V streaming
raw user input
Computation
entity-level updates
bitmaps (like A/V)
Bandwidth
6
Spectating versus A/V streaming

• Spectating
• 1 disaggregation into entities
• 2 persistence of game state
• 3 multiple camera views
• 4 flexible bandwidth adaptation
• A/V streaming
• bitmaps rather than entities
• limited flexibility

7
1 Disaggregation

• Each entity has a distinct identity
• Spectating stream comprises a stream of updates
  for the individual entities
• Differential importance of entities and
  attributes
• race car gt debris gt billboard
• position update gt scratch or dent
• Enables flexible loss recovery and bandwidth
  adaptation

8
2 Persistence

• Game state is persistent
• entities continue to exist until destroyed
• Selective patching
• mask lost update for an entity by reusing or
  extrapolating state from previous update
• Reliable updates
• e.g., entity creation and deletion

9
3 Multiple Camera Views

• Easy to generate
• e.g., each players view, birds eye view
• predefined views or spectators choice
• Overlapping views
• shared substreams to save bandwidth
• Switching between views

10
4 Bandwidth Adaptation

• Adaptation important for multiple reasons
• bandwidth heterogeneity
• congestion
• Disaggregation permits flexible strategies that
  conform to good behavior
• temporal subsampling
• reduced update frequency
• interpolation can mask artifacts to an extent
• spatial subsampling
• reduced field of view
• subsampling can be based on location, importance,
  etc.

11
Spectating Design Outline

• Assumption server-based game
• popular architecture for low-latency action games
• Game-specific spectating client
• typically derived from the players version
• includes game-specific optimizations (e.g., fast
  physics and rendering engines)
• Generic communication library
• end-host-based overlay multicast
• multi-tree architecture for resilience (CoopNet)

12
Spectating System Architecture
Server
Spectators
Players
13
Design Elements

• 1 Distributed delta encoding
• 2 Reliable message delivery
• 3 Bandwidth adaptation

14
1 Delta Encoding

• Typically only a small number of entities change
  from one frame to the next
• Delta encoding for server-player communication
• server only sends updates for entities that
  changed
• significant bandwidth savings (Quake 3 Mbps ?
  200 Kbps)
• more efficient than frame differencing in A/V
  context
• Reference frame for delta encoding
• most recent frame is typically the best
• sometimes older frame works better (e.g.,
  periodic events)

15
Distributed Delta Encoding

• Key challenge
• little direct server-spectator communication
• each spectators state might be different
• Solution distributed delta encoding
• hop-by-hop delta encoding
• parent computes delta with frames in common with
  child
• disaggregation makes this cheap compared to say
  hop-by-hop transcoding
• optimistic delta encoding
• assume that most recent frame is available at
  child
• recompute appropriate delta if assumption is
  invalid

16
2 Reliable Message Delivery

• Persistence ? certain missing updates can have
  lingering effect
• entity creation deletion
• e.g., bumper breaking off car
• Reliable message delivery
• certain updates marked as reliable
• not subsumed by later updates when computing
  delta
• Optimization
• later update can cancel earlier one
• e.g., entity creation followed by deletion
• avoids flooding new spectator with useless updates

17
3 Bandwidth Adaptation

• Goal flexibility while conforming to good
  behavior
• Temporal subsampling
• receive updates at reduced frequency
• interpolation can mask jerkiness for certain
  attributes
• delta encoding becomes less effective
• Spatial subsampling
• receive updates for only a subset of entities
• spatial ? location, view, importance, etc.
• spectating stream quality undiminished for chosen
  subset
• Temporal spatial subsampling for max flexibility

18
Mechanisms

• Striping
• multiple distribution trees
• different substream transmitted down each tree
• Filtering
• single logical tree
• parent transmit filtered stream to child
• Key tradeoff
• neither mechanism by itself is efficient
• pure striping ? lots of p2p relationships to be
  maintained
• pure filtering ? cumbersome
• Proposal
• temporal striping spatial filtering

19
Spatial Striping
S
S
S
S
1
1
2
1
1
20
Temporal Striping
S
S
S
S
1
1
2
1
1
21
Filtering
S
1
2
22
Temporal Striping Spatial Filtering
S
S
S
S
1
1
1
1
3
2
2
2
2
23
Cheering

• Audio is most natural medium for cheering
• Key differences compared to A/V conferencing
• minimal floor control (simultaneous cheers)
• less stringent delay requirements
• different combination of cheers for each player
• Alternatives
• fake cheers based on 1-bit from each spectator
• hard to accommodate diversity in language,
  accent, phrases
• composite cheering stream with in-network
  aggregation
• Can we recreate the stadium experience?

24
Open Issues

• Spectating can enable cheating
• not an issue for certain games (e.g., sports)
• delay can blunt cheating potential for some games
  (e.g., first-person shooter)
• but this might make cheering less effective
• cheating is a real possibility for slow, strategy
  games
• but perhaps spectating is not so interesting for
  these anyway
• Metric for measuring spectating quality
• entities rather than bits are what count
• classical PSNR needs to be rethought

25
Conclusion

• Spectating cheering present an interesting set
  of challenges
• Commonality and differences compared to classical
  applications (e.g., A/V streaming)
• Unique opportunities for optimization