Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group...

10
Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group Leaders: Krys Pawlikowski 1 , Don McNickle 2 , Greg Ewing 1 1 Department of Computer Science and Software Engineering 2 Department of Management Group Members: Will Gittoes, Yumi Masuri, Adriaan Schmidt, Mirko Eickhoff

Transcript of Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group...

Page 1: Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group Leaders: Krys Pawlikowski 1, Don McNickle 2, Greg Ewing.

Simulation Research GroupA Controller for Distributed Discrete-Event Simulation: Akaroa2

Group Leaders:

Krys Pawlikowski1, Don McNickle2, Greg Ewing1

1 Department of Computer Science and Software Engineering 2 Department of Management

Group Members:  

Will Gittoes, Yumi Masuri, Adriaan Schmidt, Mirko Eickhoff

Page 2: Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group Leaders: Krys Pawlikowski 1, Don McNickle 2, Greg Ewing.

2

Distributed Discrete-Event Simulation

Two Different Approaches:

• Break problem up into bits that can (nearly) be done in parallel – e.g. Timewarp

• Difficult to do, high communication costs, modest speed-up

• Speed up simulation by launching multiple replications in parallel – e.g. Akaroa2

• Easy to launch multiple replications, low communication cost, nearly linear speed-up

Page 3: Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group Leaders: Krys Pawlikowski 1, Don McNickle 2, Greg Ewing.

3

Akaroa2

• Launches identical replications of a simulation on a network

• Written in C++, works under most OS

• Only one extra line in the simulation program required

• User specifies the level of accuracy required, then the simulation runs until that is achieved

• Downloaded 1700 times from www.akaroa.canterbury.ac.nz

Page 4: Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group Leaders: Krys Pawlikowski 1, Don McNickle 2, Greg Ewing.

4

Akaroa2 on PlanetLab

Simulating this problem

Page 5: Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group Leaders: Krys Pawlikowski 1, Don McNickle 2, Greg Ewing.

5

Problems that Akaroa2 addresses

Problems Solutions

Speed Even simple systems may

need millions of events

Appropriate Accuracy for Decision-Making

Runs are either wastefully long, or not accurate enough

Multiple replications in parallel

Automated run-length control

Automated statistical analysis of results

Page 6: Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group Leaders: Krys Pawlikowski 1, Don McNickle 2, Greg Ewing.

6

Akaroa2: Architecture

Page 7: Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group Leaders: Krys Pawlikowski 1, Don McNickle 2, Greg Ewing.

7

Akaroa2: Interface

Page 8: Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group Leaders: Krys Pawlikowski 1, Don McNickle 2, Greg Ewing.

8

Akaroa2: Features

• Indifferent to fast or slow processors

• High fault tolerance: A dead processor is no different to a slow one

• Nearly linear speed-up: n processors produce results in 1/n of the time

• Extensively tested simulation features: RNG’s, transient analysis, confidence interval estimation

Page 9: Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group Leaders: Krys Pawlikowski 1, Don McNickle 2, Greg Ewing.

9

Akaroa2: Use on PlanetLab

• Requires very low bandwidth

• After set-up the communication is entirely one way and does not need acknowledgement

• In principle can be distributed over any set of hosts on the Internet – e.g. PlanetLab

No difference in simulation speed

Distributing Akaroa2 on PlanetLab Farzana, Ewing, Pawlikowski, Yamada, 2009

Page 10: Simulation Research Group A Controller for Distributed Discrete-Event Simulation: Akaroa2 Group Leaders: Krys Pawlikowski 1, Don McNickle 2, Greg Ewing.

10

Extension to Other Data Streams

Much of the expertise built up over the last 20 years with Akaroa2 can be transferred to non-simulation data:

• A reliable, fault-tolerant architecture

• Analysis of streams of data from multiple sources

• Analysis methods for correlated streams of data