is an office without wires feasible? sharad agarwal jakob eriksson, victor bahl, jitu padhye

14
Is an Office Without Wires Feasible? Sharad Agarwal Jakob Eriksson, Victor Bahl, Jitu Padhye

Post on 22-Dec-2015

217 views

Category:

Documents


0 download

TRANSCRIPT

Is an Office Without Wires Feasible?

Sharad Agarwal

Jakob Eriksson, Victor Bahl, Jitu Padhye

2.JUNE.2006 22

All-Wireless Office• No wires• No switches• No APs

2.JUNE.2006 33

All-Wireless Office• Not large corporation• Small offices

– 10-100 PCs– Rapid deployment– Short-term office– Low-cost solution

• Not replacement for wire

• Looking for good performance– how long a user waits for a transaction– small additional delay

2.JUNE.2006 4

All-Wireless Office• Office PCs

– Two 802.11 interfaces• simultaneous xmit & rcv on non-interfering channels• frequency diversity; range-rate tradeoff

• Office servers– mail, domain controllers, code repositories– proxies with wires

• Mesh routing– A lot of prior work

• Routing protocols• Link quality metrics

2.JUNE.2006 55

Questions

• What additional delay penalty will a mesh network impose– In typical office configurations– With typical office traffic

• How should an administrator pick :– Wireless hardware– IEEE 802.11 band– Routing metric– User-server placement– Spatial reuse, hidden terminal

2.JUNE.2006 66

Don’t We Already Know?• Typical evaluation

– Select sender, receiver at random; 1 TCP flow, 2 mins– Repeat 100 times, calculate median

0

500

1000

1500

2000

2500

3000

3500

WCETT ETX Shortest Path

Th

rou

gh

pu

t (K

bp

s)

Single Radio Two Radios

2.JUNE.2006 77

Methodology• Capture traffic from 11 office users

– Packet level capture insufficient• TCP effects in wireless, multihop very different

– Socket level is best: open, send, receive, close

• Replay on mesh testbed among office users– MCL by Draves, Padhye, Zill @ MSR (2004)

• DSR-like routing with virtual link layer optimizations• Link metrics: hop, RTT, PKTPAIR, ETX, WCETT

– Assign users, application servers to testbed

• Examine several design choices• Not disrupt actual users

2.JUNE.2006 8

0100020003000400050006000700080009000

Email + PFS DomainController

Proxies Source Depot Other Internal OtherExternal

MB

Tra

ns

ferr

ed

Received

Sent

8

Captured Traffic• Very diverse traffic types, sizes, concurrency• Map each type to 1-2 mesh machines for replay• Non-winsock traffic not captured

– Get user RPC; miss SMB, NBT (almost all IDS for us)

2.JUNE.2006 99

Replay Model• Concurrent

sessions• Session

– connect to disconnect

– multiple transactions; not concurrent

• Transaction– 1 send, 0+

receives

• Response time– start of send– end of last

receive of the transaction

2.JUNE.2006 1010

Mesh Testbed

DN

UP

EL 32

~ 32 m

~ 76 m

226

227

225

219

215

220217

218

216

210

214211

209

208

207205

206

203

201202

204

  Central Distant Extreme

User 01 203 203 203

User 02 205 205 205

User 03 207 206 206

User 04 208 208 208,207

User 05 209 209 209,210

User 06 211 211 211,214

User 07 226 215 215,216

User 08 225 217 217,202

User 09 218 218 218,204

User 10 227 219 219,225

User 11 204 220 220,226

Domain Controllers 214,215 204,226 201,227

Source Depots 217 227 201,227

Exchange 220 202 201,227

Proxies 216,219 201,225 201,227

ConfigNetgear

WGNetgear WAB/G

ORiNOCO Proxim

Transmit Power

RTS / CTS

A a-56 a-36   100%  

B a-56 a-36   100% on

C a-56 g-10   100%  

D   g-10 a-56 100%  

E   g-10 a-56 50%  

F   g-10 a-56 12.50%  

DN

UP

EL 32

~ 32 m

~ 76 m

226

227

225

219

215

220217

218

216

210

214211

209

208

207205

206

203

201202

204

DN

UP

EL 32

~ 32 m

~ 76 m

226

227

225

219

215

220217

218

216

210

214211

209

208

207205

206

203

201202

204

DN

UP

EL 32

~ 32 m

~ 76 m

226

227

225

219

215

220217

218

216

210

214211

209

208

207205

206

203

201202

204

2.JUNE.2006 1111

Light Load, Central Placement

2.JUNE.2006 1212

Heavy Load, Distant Placement

2.JUNE.2006 1313

Summary of Results• Results are unusual

– Captured traffic is very different than synthetic– Prior work’s throughput results not very helpful

• Many configurations – median delay <20ms– 802.11 hardware had upto 2.5x difference– 802.11 band had upto 2x difference– Server placement had upto 3x difference– No benefit of spatial reuse, hidden node avoidance– 2 routing metrics bad, 3 good & very similar

• "Feasibility Study of Mesh Networks for All-Wireless Offices", in ACM Mobisys, June 2006

2.JUNE.2006 1414

Open Issues / Limitations

• 1 testbed, 1 set of user traces– but many configurations, different time periods

• Performance can be improved further– cross interference detection & adaptation– gateway balancing

• Skipped some real world issues– fairness– security / DoS

• Jamming, routing disruption, resource consumption