ns-3 performance analysis and development of effective ... · synchronization lookahead concept:...

NS-3 performance analysis and development of effective load

balancing algorithms

Olga Lesnova & Eugene Kalishenko

Open Source & Linux Lab

http://osll.fruct.org

FRUCT 13 26.04.2013

Outline

FRUCT 13 26.04.2013 2

Networks modeled by NS-3 can be large...Or very large...Or very very very large...

When you are modeling very very large networks and want to get some statistics you can wait results for a very very long time...

Basic idea

FRUCT 13 26.04.2013

Main ideaLocal events in each subnetwork can be performed independently

Main problemSynchronize subnetworks

3

Synchronization algorithms

FRUCT 13 26.04.2013

Optimistic- Desynchronization happens rare

- Periodically interrupting and fixing synchronization errors

Effective | Difficult to implement

Conservative- Avoid desynchronization

Easy to implement | Not so effective

4

Objectives

Analyze existing parallel implementations performance

Effective network clustering

Effective load balancing

5FRUCT 13 26.04.2013

Algorithms

FRUCT 13 26.04.2013 6

SynchronizationLookahead concept: Event in one node can affect

another node with delay not less then delay in channel between them

FRUCT 13 26.04.2013 7

In Multithreaded version lookahead:- Local for each node- Minimal delay between node and its neighbors- For node A: min(5, 6, 1, 9) = 1

In Distributed version lookahead:- Global- Minimal delay between subnetworks- Min(3, 4, 5, 6, 7, 2) = 2

Multithreaded algorithm

FRUCT 13 26.04.2013 8

Distributed algorithm

FRUCT 13 26.04.2013 9

Analysis

Hardware:

Default & Multithreaded.: Intel(R) Core(TM) i5-2320 CPU @ 3.00GHz X 4

Distributed: Cluster 2 X Intel(R) Core(TM) i5-2320 CPU @ 3.00GHz X 41 - http://www.ssfnet.org/Exchange/gallery/baseline/index.html

FRUCT 13 26.04.2013 10

DARPA NMS Campus Network

2 / 202 / 42

2 / 644 / 20

4 / 424 / 64

8 / 208 / 42

8 / 64

0

1000

2000

3000

4000

5000

6000run time

Default Multithreaded Distributed

CN/LAN

time,

s

[1]

Profiling

FRUCT 13 26.04.2013 11

Distributed versionProfiled with Intel Cluster Studio

Multithreaded version Profiled with Intel Parallel studio

MotivationWhy effective network clustering is important?

FRUCT 13 26.04.2013 12

Good

clustering

Bad

clustering

Static load balancing

FRUCT 13 26.04.2013 13

Weighted graph

FRUCT 13 26.04.2013 14

Dynamic load balancing

FRUCT 13 26.04.2013 15

Results & Future Work

FRUCT 13 26.04.2013

Results- Multithreaded & Distributed algorithms analyzed

- Complete simulators profiling performed

- Static load balancing for Distributed Simulator developed

Future Work- Dynamic load balancing for Distributed Simulator

16

Questions & Answers

Lesnova Olga lesnovaolga@gmail.com

Open Source & Linux Lab, http://osll.fruct.org, osll@fruct.org

FRUCT 13 26.04.2013