Algon Reusable Distributed Algorithms in Java

1
Algon Reusable Distributed Algorithms in Java

• Karen Renaud

Judith Bishop Basil Worrall Johnny Lo
2
Rationale

• Programming distributed applications is hard
• Extra concerns must be dealt with
• Non-determinism
• Contention
• Synchronisation difficulties

3
Distributed Algorithms

• Many different algorithms exist for
• Distributed mutual exclusion
• Distributed termination
• Distributed deadlock detection
• Few programmers understand
• how the algorithms work, or
• how to distinguish between them.

4
Ideally

• Developers of distributed systems should be able
  to
• experiment with different algorithms
• choose the best one for their particular system
• easily compare performance and efficiency

5
However

• Algorithms are often so bound up with the code
  that it is very difficult to substitute another
  algorithm
• Few programmers have the skills to compare and
  evaluate algorithms
• Programmers often do not have time for this!

6
Palliatives

• Systems exist which illustrate the functioning of
  algorithms Ben-Ari Worrall
• Libraries of algorithms are provided. Guerraoui
• But ...
• Little assistance is given to users in choosing
  and evaluating algorithms

7
Separating out the Distributed Algorithm

• Simplify the programmers task by providing
• a library of interchangeable algorithms
• a tool for easily substituting one algorithm for
  another without recompiling code
• support for critical evaluation of the algorithm
  performance

8
Example
Application
Classic Algorithm
9
Distributing the Algorithm
Application
Classic Algorithm
Distributed Classic Algorithm
Distributed Readers and Writers
10
Algon Framework
Application
11
Distributed nodes with Algon
Application
Application
Application
Algon Layer
Scheduler
Scheduler
Scheduler
Algorithm
Algorithm
Algorithm
12
How to Achieve Separation?

• How to separate algorithmic code from the
  application effectively ?
• How to make such algorithms easily
  interchangeable ?
• Determine the unique requirements of the
  distributed layer for each family of algorithms ?

13
Example Incorporating Algon to achieve
Distributed Mutual Exclusion
Application
read
write
In a centralised site a shared resource is
often controlled by a type of monitor which
grants Read or write access
14
Example Incorporating Algon
Replace it here with the Algon layer
Application
read
Algon Layer
write
Site 2
Site 3
15
The Nerve Centre The Mutual Exclusion Scheduler
acquireRead() acquireWrite() releaseRead() release
Write() check()
Scheduler

• Tailored to the needs of the distributed
• mutual exclusion family of algorithms
• Communicates with distributed algorithm
• Easily Configurable without recompilation

16
Class Diagram Clean interfaces, clear separation
Reader Writer
17
Class Diagram Clean interfaces, clear separation
Reader Writer
18
Setting up
App
19
Setting up
20
Relaxing mutual exclusion for Readers/Writers
Reader Writer
Algorithm
21
Example Distributed Deadlock Detection
22
Example Incorporating Algon to achieve
Distributed Deadlock Detection
Application
Detect deadlock
Algon Layer
Site 2
Site 3
23
The Nerve Centre The Deadlock Detection
Scheduler
detectDeadlock()
Scheduler

• Tailored to the needs of the distributed
• deadlock detection family of algorithms
• Communicates with distributed algorithm
• Easily Configurable without recompilation

24
Class Diagram Clean interfaces, clear separation
Scheduler
25
Setting up resources
Middle Ware Layer
App
Resources
26
Setting up
App
Middleware
27
Setting up
28
Detect Deadlock
Application
Other Site
Other Site
Algorithm
29
Evaluation Metrics

• Dynamic Extensibility
• Non-invasive configurability
• Scalability
• Flexibility
• Coupling
• Maintainability
• Substitutability
• Imperturbability
• Comprehensibility

30
Future Work

• Identifying metrics for measuring algorithm
  performance
• Execution time?
• Number of messages exchanged?
• Tailorability eg. Relaxed or stringent mutual
  exclusion
• System load

31
Future Work

• Measuring algorithm performance
• Portraying the performance and efficacy of
  different algorithms
• Supporting user choice of algorithms
• Fault Tolerance
• Adding distribution changes the failure semantics
• Algon may introduce new types of errors

32
In Conclusion.

• Algon provides
• A layered framework
• Various algorithms, wrappers, interfaces in Java
• Schedulers to handle communication
• Algon facilitates the separation of application
  code from algorithmic code
• Algon addresses the issue of complexity in
  distributed systems

33
Fin