Priority Round-Robin Scheduling for Very Large Virtual Environments

1
Priority Round-Robin Schedulingfor Very Large
Virtual Environments

• Chris Faisstnauer,
• Dieter Schmalstieg, Werner Purgathofer
• Vienna University of Technology

2
Introduction

• Virtual Environments with large amounts of
  elements
• Competition for limited resources (bottleneck)
• Graphics pipeline (rendering)
• Processing power (simulation)
• Network bandwidth (updates)
• Selection of element subset
• Reduce absolute number
• Traditional scheduling for objects
• Degradation of system performance
• Approximation must be made

3
Goal

• Development of a generic scheduling algorithm
• Employ it as stand-alone scheduling
• Combine with element-reduction methods
• Graceful degradation (find best approximation)
• Output sensitive
• Immune to starvation
• Enforce use of priorities
• Priorities based on freely definable error metric
• E.g. fast entities ? frequent updates
• slow entities ? seldom updates

4
Related Work

• Short-term scheduling
• independent processes
• allocate processor time
• optimize system behavior
• First Come-First Served (FCFS) / Round-Robin
  (RR)
• execute in order of submission (no priorities)
• output sensitive, immune to starvation
• Multi-Level Feedback Queue
• levels with decreasing priorities
• risk of starvation vs. constant monitoring

5
Basic Priority Round-Robin 1/3

• Elements compete for resources ? accumulate error
• Error modeled as error metric
• Assign each element Error Per Unit (EPU)
• Goal minimize cumulative error
• No traditional sorting
• Approximate sorting in multiple levels (FIFO)
• Elements assigned to level according EPU
• Level priority reflects scheduling frequency
• Combines advantage of Round-Robin full sorting

6
Basic Priority Round-Robin 2/3
i0 i1 i2
Repetition counti NrElementsi
NrLevels Predicted error ErrorPerUnit
Repetition Count
Selected elements A,C,G - B,D,G - A,E,G -
B,F,G
7
Basic Priority Round-Robin 3/3

• Assignment of elements to levels
• Minimum overall error
• Average Error Per Unit ? variable size levels
• Dynamic VE ? dynamic error distribution
• Varying traversal rate (level i)

ni number of elements in level
i tri traversal rate of level i level number
of levels
8
Optimum Traversal Rate 1/2
level number of levels ni nr of elements in
level i avi average EPU of level
i rci repetition count of level
i tri traversal rate of level i lerri level
error err cumulative error
9
Optimum Traversal Rate 2/2
level number of levels ni nr of elements in
level i avi average EPU of level
i rci repetition count of level
i tri traversal rate of level i lerri level
error err cumulative error
10
Evaluation

• Client-server system
• Server hosts simulator (translates elements in
  2D)
• Client visualizes scene (needs position updates)
• Subset of elements position can be updated
• Select subset using PRR-scheduling
• Visual error distance object position on server
  / client
• Evaluation of PRR (Priority Round-Robin)
• Comparison PRR vs. plain RR
• Comparison DRPRR vs. plain DR

11
Testbed (Video)
12
Example 1

• Scheduling 1000 out of 10000 simulated cars (10)
• Velocities (in units) 500 cars - velocity ?
  9,10
• 1500 cars - velocity ? 3,4
• 8000 cars - velocity ? 0.1,0.5

Overall error of PRR is 73 lower than RR
13
Example 2

• Scheduling 1000 out of 10000 simulated cars (10)
• Velocities (in units) 10000 cars - velocity ?
  1,10

Overall error of PRR is 7.5 lower than RR

14
Example 3

• Scheduling 1000 out of 10000 simulated cars (10)
• Simulating cars using Dead Reckoning (? 25 cars
  above threshold)
• 500 cars - velocity?9,10 - angle offset
  ?19?,20? every 10 steps
• 1500 cars - velocity?3,4 - angle offset
  ?4?,5? every 10 steps
• 8000 cars - velocity?0.1,0.5 - angle offset
  ?1?,2? every 10 steps

Overall error of DRPRR is 63 lower than DR
15
Example 4

• Quake Deathmatch scheduling 2 out of 9 players
• Position / velocity of entities given by recorded
  demo

Overall error of PRR is 48 lower than RR
16
Conclusions

• Enhance (plain) RR ? Priority Round-Robin (PRR)
• Enforcement of priorities
• Output sensitive
• Immune to starvation
• Freely definable error metric
• PRR is a suitable substitute for RR in most cases
• Minimize overall visual error in VE
• Combine Priority Round-Robin and Dead Reckoning
• Combine Priority Round-Robin and visibility
  techniques

17
Future Work

• Measure for object activity
• Use of visibility information
• Temporal Bounding Volumes (TBV)
• Temporally invariant Bounding Volumes (tiBV)
• Evaluate motion data from large Virtual
  Environments
• E.g. Everquest, Ultima Online
• Scheduling humanoid avatars
• Multiple Levels Of Detail (LOD)

18
OS vs. VE - Scheduling

• Operating systems
• Scheduled once
• Priorities scheduling order
• Small number of elements
• Constant monitoring
• Variable amount resources

• Virtual Environments
• Scheduled repeatedly
• Priorities scheduling frequency
• (Very) large number elements
• Output sensitive
• Constant amount resources

• Optimize system parameters
• Enforce priorities
• Minimize risk starvation