Software Architectures for Reducing Priority Inversion and Nondeterminism in Realtime Object Request

1
Software Architectures for Reducing Priority
Inversion and Non-determinism in Real-time Object
Request Brokers
Zhengping Wu September 23rd, 2002 CS 851 OS
Middleware
2
Outline

• Overview
• Comparison between different architectures
• Experiments
• Conclusion
• Discussion

3
Trends

• Middleware Integration frameworks
• QoS-enabled components Open systems
• Standardization and leveraging of real-time COTS
  hardware and software

4
Real-time ORB Endsystem
5
General Performance Overhead

• Network
• Communication protocol implementations
• ORB transport protocol implementations
• ORB core implementations

6
Optimizing Real-time ORB Endsystem
7
Priority Inversion and Non-determinisms Sources

• OS I/O subsystem
• ORB core
• Connection
• Concurrency
• Object adapter

8
A Multiplexed Connection Architecture
9

• Layered
• CORBA
• Request
• Demultiplexing

10
Alternative ORB Core Connection Architectures

• Multiplexed connection architectures
• Active connection architecture
• Leader/followers connection architecture
• Non-multiplexed connection architecture

11

• Active
• Connection
• Multiplexing
• Architecture

12

• Leader/Follower
• Connection
• Multiplexing
• Architecture

13

• Non-multiplexed
• Connection
• Architecture

14
Alternative ORB Core Concurrency Architectures

• The worker thread pool architecture
• The leader/follower thread pool architecture
• Threading framework architecture
• The reactor-per-thread-priority architecture

15

• Server-side
• Worker
• Thread
• Pool
• Concurrency
• Architecture

16

• Server-side
• Leader/Follower
• Concurrency
• Architecture

17

• Server-side
• Thread
• Framework
• Concurrency
• Architecture

18

• Server-side
• Reactor-per-
• Thread-Priority
• Concurrency
• Architecture

19
Experiments on Solaris

• Systems to be compared
• CORBAplus
• miniCOOL
• MT-Orbix
• TAO

20
ORB Endsystem Benchmarking Testbed
21
Comparative Latency
22
Comparative Jitter
23
Whitebox Result for CORBAplus
24
Whitebox Results for miniCOOL
25
Whitebox Results for MT-orbix
26
Whitebox Results for TAO
27
User-level Locking Overhead
28
Kernel-level Locking Overhead
29
Experiments on Chorus ClassiX

• Systems to be compared
• miniCOOL
• TAO

30
Comparative Latency for Local Configuration
31
Comparative Latency for Remote Configuration
32
Comparative Jitter
33
Conclusion

• Avoid dynamic connection establishment
• Minimize dynamic memory management
• Avoid multiplexing requests of different
  priorities over a shared connection
• Be flexible, efficient, and predictable
• Avoid reimplementing OS mechanisms
• Be guided by empirical performance benchmarks

34
Discussion

• Try every connection architecture on client-side
  with all kinds of concurrency architectures on
  server-side
• Is TAO the ideal architecture?