JSIMS Common Component Simulation Engine Performance Analysis

1
JSIMS Common Component Simulation Engine
Performance Analysis

• Craig Lammers
• Dr. Jeffrey S. Steinman
• RAM Laboratories, Inc.

2
Outline

• JSIMS Background
• Hierarchical Grids
• Performance Results

3
JSIMS Background

• New architecture in early 2000
• HLA Federation approach
• SPEEDES-based Common Component Simulation Engine
  (CCSE)
• Rapid Phase 1 development June 2000 - April 2001
• Critical to quickly implement functional
  capabilities in CCSE
• Unify and document SPEEDES APIs
• Distributed Simulation Management Services Layer
• HLA Gateway
• JSIMS-API Compatibility Layer (JCL) for backward
  compatibility
• Tiger teams were formed to shake out the overall
  architecture
• Difficult decision was made in May 2001
• Focus on either performance or on
  Checkpoint/Restart
• Checkpoint/Restart took priority - performance
  efforts deferred

4
JSIMS BackgroundContinued

• Performance Optimizations in Late 2002 - 2003
• Motion algorithms and range filtering
• Persistence and Checkpoint/Restart optimizations
• HLA Gateway time management optimizations
• Interest management and data distribution
  optimizations
• Specialized support for JSIMS federates (NASM)
• Program Cancelled
• But maintained until the Analysis of Alternatives
  report
• Hierarchical Grids (HiGrids) in Early 2004
• Overlap computation was a known bottleneck in
  system
• HiGrid-based interest management was planned for
  Phase 2, but was deferred until 2004
• Completed in June 2004
• Results show remarkable performance improvements

5
HiGrids Background

• Functionality
• Provide overlap computations between published
  filterable attributes and subscription filters
• Limited implementation in 1999
• Integrated with earlier SPEEDES object proxy
  system
• High performance results were presented at SIW
• JSIMS required a different HiGrid structure
• Much more capable implementation in 2004
• Works with more elaborate Federation Object
  system
• Supports a wide variety of data types
• Fully integrated and optimized with persistence
• Handles multiple filters from the same subscriber

6
1999 Performance Results Computations

7
HiGrid Class Diagram
HiGrid


HiGridDim
1
1
LatLonDim
Publisher Tree
Subscriber Tree
AltDim
IntervalDim


PubSubData
PubSubData
NumericDim
EnumDim


PubSubValue
PubSubValue
StringDim
BitFieldDim
ObjHandleDim
FoIdDim
8
Effects of Dimension Branching
9
Multiresolution Interval Dimension
10
Multiresolution Latitude Longitude Composition
11
Spherical Interest Region Defines Bounds on
Latitude and Longitude
12
HiGrid Performance

• Goals
• Compare performance to brute force approach
• Examine performance sensitivity to various
  conditions
• Test Environments
• Standalone testing of the HiGrid data structure
• Integrated testing of HiGrids with CCSE interest
  management
• JSIMS Federation testing began but was not
  completed

13
Benchmarks Compare HiGrids Against Brute-Force

• Standalone
• HiGrids with no publisher/subscriber overlaps
• Steady state case when publishers modify
  filterable attributes and subscribers modify
  filters
• Sensitivity to HiGrid dimension ordering
• Impact of persistence
• Integrated with CCSE interest management
• Data storage overlap computations for subscribe
  all cases
• Range filtering between moving entities and
  sensors

14
Standalone HiGrid Performance
15
Standalone HiGrid FilteringNo Overlaps
16
Standalone HiGrid FilteringModifying Publisher
Data
17
Standalone HiGrid FilteringDimension Ordering
18
Standalone HiGrid FilteringPersistence
19
HiGrid Performance when integrated with CCSE
Interest Management
20
HiGrid Data Storage Overlap Computations with
CCSE Filtering Disabled
21
CCSE Range Filtering with HiGrids
22
Conclusions

• Standalone testing
• For 20,000 publishers and subscribers with no
  overlaps, HiGrids outperformed brute force method
  by a factor of 100
• For 2,000 publishers and subscribers with 11
  overlaps, HiGrids perform at least as good as the
  brute force method
• Dimension ordering can have a large effect on
  performance
• Persistence adds a small amount of overhead to
  HiGrids
• CCSE integrated testing
• HiGrids perform up to 4 times faster than brute
  force interest management when subscribing to all
  publishers
• HiGrid performance for range filtering with 1
  discoveries scales well and achieves
  significantly better performance over brute force
• HiGrid interest management was successful in
  improving the performance of JSIMS