ModelNet: Scalability and Accuracy in a Large-Scale Network Emulator

Ken Yocum, Kevin Walsh, Amin Vahdat, Priya Mahadevan, Dejan Kostic, Jeff Chase, and David Becker

Duke University, http://issg.cs.duke.edu

Goal: develop and explore scalability techniques for distributed service evaluation

ModelNet: Wide-area Emulation

•simulation– slow, must modify applications

•live deployment– limited in size, waiting for failures, not reproducible

•Emulation– static, small scale, hard to configure

•ModelNet: – accurate: error<1ms @ 95% cpu utilization– multiple techniques trade scalability for accuracy– Run unmodified applications on unmodified OS’s– reproducible results (recreate flash crowds, network partitions, etc)

– Supports overlays, wireless networking, content distribution networks, peer-to-peer systems . . .

faultscross traffic

1.) Create

bandwidthlatency

loss

4.) Bind

5.) Run

Map virtual to physical

trade accuracy for scalability

observe

target topology

ModelNet Performance

0

10000

20000

30000

40000

50000

60000

70000

80000

90000

0 20 40 60 80 100 120

1 hop2 hop3 hop4 hop

10Mbit flows

packets

/secon

d

FreeBSD 4.5, Architecture X Single Core Capacity (10vn’s/edgenode)

•Per-Hop cost = 2.8 µs•Per-packet cost = 8 µs•Distillation matters•packet sizes (1500 bytes), full mesh

ModelNet Validation Chord CFS

clients

3.) Assign

transits

stubs

topological cuts

core A

core B

core C

•Assign links to cores.

•balance load•reduce inter-core

traffic– replicate pipes– observe/adjust

•Graph partitioning problem– Spectral Analysis– Kernighan & Lin

ip

ipfwsrc:10.0.0.0/8 dst:10.0.0.0/8

app1 app2 app3 app4

Unmodified KernelLinux/FreeBSD/Solarisroute 10.1.0.0/16 -> core

interception DLLw.x.y.z -> 10.0.0.0/8

100 Mb/s

1 Gb/s

Network Switch

network interface

network interface

core ycore x

kernel

cross traffic emulation

modelnet emulated link

delay loss

emulated link...

routing matrix to

from

route lookup

client coreedge nodes emulation hardware

2.) Distilladaptive mesh refinement

preserve first hop

walk_in = 1

preserve corewalk_out = 1

clients

transits

stubs

Install ModelNetDataBase

configuration

balance load &minimize hardware footprint

0

0.2

0.4

0.6

0.8

1

1.2

0 100 200 300 400 500

CFS 8-KB

MN 8-KB

CFS 64-KB

MN 64-KB

CFS 1126-KB

MN 1126-KB

cu

mu

lati

ve p

erc

en

tag

e o

f tr

an

sfe

rs

KB/second

•Emulated RON•reproduce end-to-end effects•uses single edge node

•multiple cores•payload caching•ship description of packet, not data