distributed control algorithms for service differentiation in wireless packet networks michael...

14
Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres [email protected] [email protected] [email protected]

Upload: geoffrey-edwards

Post on 16-Dec-2015

213 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks

Michael Barry, Andrew T Campbell, Andras [email protected]

[email protected]@ericsson.com

Page 2: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

Distributed vs Centralized MAC

• Centralized– requires a base station– a base station schedules both up/downlink transmissions (e.g., polling)– better control over radio resources– service guarantees (fairness, delay, loss)– e.g.: cellular systems, GPRS

• Distributed– mobile hosts control their transmissions themselves– simple– robust (e.g., radio interference)– works in ad-hoc networks, or when channel is shared– e.g.: 802.11 DCF mode

Page 3: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

Distributed Service Differentiation• Is it possible to achieve service differentiation using exclusively

distributed algorithms?

• All components of the proposed service architecture are distributed

1. modified 802.11 DCF MAC to support “soft” prioritization

2. passive channel quality monitoring at the mobile

3. estimation of achievable service quality

4. distributed admission control

5. SLA Management

Page 4: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

IEEE 802.3 DCF

• Distributed MAC algorithm– When a packet is to be sent, choose random defer time

Tslot *Rand[0,CW]– Transmission attempt is made when defer time expires.

• Decrement when channel is idle• Success reset CW to CWmin for next frame• Failure increase CW to CW x 2, up to CWmax, and restart

• Highly random, unbounded delay• Best-effort• No fairness guarantees

Frame

Defer Timewait

wait

Station A

Station B

Station C

wait

Page 5: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

“Soft” Differentiated 802.11 MAC• The delay of a packet depends on the chosen random defer time

and the remaining defer times of other hosts

• Adjust CW so that high priority packets are more likely to grab the channel– increase upper limit of defer time (CWmin) for low priority traffic (LP)

– decrease CWmin for high priority traffic (HP)

Channel busy

HP

LP

CWmin

Defer transmission

Page 6: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

Service Differentiation - Analysis

• Impact of Different Contention Window Bounds

Page 7: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

Service Differentiation - Simulation

• Service differentiation is statistical and relative– High priority packets experience lower delay on average

• Channel capacity is shared between classes

Average Delay Throughput

Increasing Traffic Increasing Traffic

Del

ay[

s]

0.1

0.08

0.06

0.04

0.02

1e+06

8e+05

6e+05

4e+05

2e+05

1.2e+06

Th

rou

gh

pu

t[b

ps]

Page 8: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

Channel Monitoring• Soft Differentiation is not enough

– most real-time services require absolute and not relative guarantees

• Admission control is needed to maintain continuous QoS– should be distributed and robust

• Passive Approach

– is non intrusive, observes activity on the channel

– extended to estimate QoS of a new session based observation

• Virtual Algorithms– Run in parallel to real apps and MAC

– Test State of channel to estimate available level of service

Page 9: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

Passive Virtual MAC• Virtual MAC emulates a real 802.11 MAC

– receives virtual packets from a virtual application

– chooses random defer time

– observes channel and decrements timer

– if timer expires, “transmits”

– estimates collisions by observing transmission from other hosts

– adjusts Cwmin (backoff), retransmits, etc...

Radio channel state

packetdecr. timer decr. timer

collision!

delay of virtual packet

success!

Page 10: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

Evaluation of V-MAC

• V-MAC delays are precise

number of voice flows

aver

age

dela

y[s]

measured

V-MAC

0.02

0.01

2010

Page 11: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

Distributed Admission Control

• A V-MAC runs using high priority virtual packets in mobile hosts before staring session– since it is passive it can be on, long before the session starts

• If V-MAC delay estimates exceed a conservative value, reject– the safety margin is required because of unanticipated interference, bursts in

best-effort traffic, session dynamics in neighboring areas

• Algorithms run in both End Hosts– Provides comprehensive analysis of channel state

– Lessens Impact of Hidden Nodes

Page 12: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

Distributed Admission Control• 10 Access Points, 100 mobiles, shared channel• 50 mobiles generate Web traffic (TCP, Pareto file sizes)• 50 mobiles generate voice (32kbps on/off)• Admission Control

– Admit new voice calls if delay is under 10 ms.

– Web traffic always admitted

AP and host positions Average Delay of admitted voice calls

0.01

0.005

0.015

0.02

dela

y[s]

Page 13: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

QoS management in Access Networks

• Supporting Fast Mobility in the access network• Routing - Micromobility

• QoS: - Distributed Algorithms

• Security: - Ongoing, AAA

• SLAs: - Context Transfer

• Move QoS State as fast as routing information during handoff.• SLA Tokens

Page 14: Distributed Control Algorithms for Service Differentiation in Wireless Packet Networks Michael Barry, Andrew T Campbell, Andras Veres michael.barry@broadcom.ie

Conclusions and Future Work• Distributed algorithms can be used for service differentiation

• Soft guarantees are enough for many RT applications

• High priority service degrades “gracefully” with increasing low priority traffic load (we can forget admission control for best-effort)

• Passive V-MAC can efficiently estimate delays introduced by the MAC

• The distributed algorithm is robust in high interference, shared environment

• Tokens can be used to transfer state information in the network.