tcp over 802.11e doug leith & peter clifford hamilton institute, ireland

26
Hamilton Institute TCP over 802.11e Doug Leith & Peter Clifford Hamilton Institute, Ireland.

Upload: ciqala

Post on 19-Jan-2016

32 views

Category:

Documents


1 download

DESCRIPTION

TCP over 802.11e Doug Leith & Peter Clifford Hamilton Institute, Ireland. Overvie w Known that 802.11b TCP uploads exhibit unfairness and sustained lockout. Known that 802.11b unfairness exists between uploads and downloads. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

TCP over 802.11e

Doug Leith & Peter CliffordHamilton Institute,

Ireland.

Page 2: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Overview

• Known that 802.11b TCP uploads exhibit unfairness and sustained lockout.

• Known that 802.11b unfairness exists between uploads and downloads.

• Setting 802.11e parameters correctly resolves these problems.

• Analytic model of the MAC behaviour of the solution.

• Model TCP dynamics in an arbitrary WLAN.

• Round Trip Time unfairness for uploads disappears when WLAN is bottleneck.

• Convergence rate calculations for competing TCP flows possible.

Setup

• 802.11b/e, single cell, infrastructure mode, DCF, TCP SACK, saturated TCP.

• Applicable to any situation where WLAN is the bottleneck.

Page 3: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

TCP Upload Unfairness in 802.11b

Upload ACKs

Upload data

AP

•Usual 802.11b unfairness between upload data and upload ACKs:• MAC ensures fair channel access: Upload data wins n/(n+1) of transmission opportunities, Upload ACKs win only 1/(n+1) Very few ACKs get through.• AP queue (which is relatively underprovisioned) overflows, drops occur.• [Pilosof, Ramjee, Shavitt, Sinha, INFOCOM 2003]

•TCP specific unfairness between competing flows in the same direction:• ACK clocking disrupted, burstiness increased• Congestion window growth hampered• Possible exponential increase in retry timer flow gets locked out.• [Detti, Graziosi, Minichiello, Salsano, Sangregorio]

Page 4: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

TCP Upload Unfairness in 802.11b

Page 5: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Restoring Upload Fairness using 802.11e

Prevent loss of much more ACKs than data:• Use separate queues for ACKs and data, • Set 802.11e AIFS parameter to highly prioritise the queue containing ACKs.

• No ACK overload as they are created & limited by data.• High priority will prevent any ACK queues from filling and avoid concerns about correct queue provisioning.

Page 6: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Restoring Upload Fairness using 802.11e

Page 7: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Advantages of the solution

ACK prioritisation essentially decouples MAC layer contention from transport layer congestion control.

Throughput only slightly reduced.

Expected normal TCP behaviour restored.

CWmin parameter still available for differential prioritisation of TCP flows.

Page 8: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Although TCP traffic is not saturated, the rate of ACKs is matched to the rate of data packets. Can use this fact to extend Bianchi type model fairly easily: Assume that a data packet is immediately followed by its TCP ACK, and that the TCP ACK is never involved in a collision.

• Agrees well with simulation• ACKs never colliding is not a legitimate assumption (e.g. ACK arrives after data packet has already counted down a lot)

Better to extend Battiti & Li model by using a hold state which data must count down through if there is a TCP ACK present. Replace TCP ACK saturation assumption with rate matching assumption.

Both models easily adjusted to delayed ACK case.

Analytic MAC Model

Page 9: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Analytic MAC Model

Page 10: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Analytic MAC Model

Page 11: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

TCP Unfairness between uploads and downloads in 802.11b

Page 12: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Restoring Fairness between uploads and downloads using 802.11e

At AP:• Use separate queues for TCP data, • Determine current number nd of destination stations,• Use 802.11e TxOP burst parameter to send first packet to each of the destination stations whenever the AP gains medium access.

Must prioritise the download ACKs using AIFS to avoid same problems as before.

• Accommodates both bursty short lived traffic and long lived traffic.• If time sensitive traffic present can use CWmin parameter also to give higher priority smaller bursts. • Can apply Battiti & Li model to calculate correct value of CWmin for the

download ACKs.

Page 13: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Restoring Fairness between uploads and downloads using 802.11e

Page 14: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

TCP Congestion Window Model Assumptions

• Station interface queues are large enough so that they do not empty when a TCP flow backs off its congestion window following packet loss:

• Wireless stations always saturated.• Activity on medium is decoupled from the congestion window size.

• The wireless link provides stochastic service:• Random packet losses.• Service and loss rates independent of congestion window size.

Page 15: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

TCP Congestion Window Evolution Model

[A. Berman, R. Shorten, D. Leith, “Positive Matrices Associated with SynchronisedCommunications Networks”, Linear Algebra and its Applications 393, 2004]

Page 16: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

TCP Congestion Window Evolution Model

Page 17: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

TCP congestion window evolution model

• Positive real eigenvalues,• Unique largest eigenvalue = 1,• Network dynamics globally exponentially stable,• Equilibrium point equal to eigenvector of eigenvalue 1,• Equilibrium value of congestion window of flow i is

• Convergence rate of mean congestion windows bounded above by largestmean backoff factor.

Page 18: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

WLAN bottleneck: Throughput

When the WLAN is the bottleneck, the throughput achieved by the competing TCP upload flows is independent of the TCP AIMD parameters:

Page 19: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Varying bottleneck: Upload RTT Fairness

When the bottleneck is the WLAN, normal upload TCP Round Trip Time unfairness does not occur.

Page 20: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Varying bottleneck: Changing upload fairness

When the bottleneck varies in location, a sharp transition in behaviour occurs.

Page 21: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Convergence Rate: Downloads

• Convergence rate is the time that the network takes to reach steady state following a disturbance.

• Download TCP flows’ convergence rate is determined by the AIMD backoff factor.

Usual TCP backoff factor .5: Model predicts4 congestion epochs toconvergence.

Page 22: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Convergence Rate: Downloads

Unusual TCP backoff factor .8: Model predicts14 congestion epochs toconvergence.

Page 23: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Convergence Rate: Uploads, WLAN bottleneck

Instant convergenceindependent of backoff factor. E.g.backoff factor .5:

Page 24: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Convergence Rate: Uploads, WLAN bottleneck

Instant convergenceindependent of backoff factor. E.g.backoff factor .95:

Page 25: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Conclusions

• Using the power of 802.11e and the correct choice of parameters we can restore TCP fairness and expected TCP behaviour to both uploads and downloads.

• Analytic MAC model presented.

• In any similar asymmetric TCP situation, ACKs should be prioritised.

• Modelled TCP dynamics in an arbitrary WLAN.

• Round Trip Time unfairness for uploads disappears when WLAN is the bottleneck.

• Convergence rate calculations for competing TCP flows possible.

• Instant convergence for uploads when WLAN is the bottleneck.

Publications

• Modelling TCP Dynamics in Wireless Networks [Wirelesscom 2005].• TCP Fairness in 802.11e WLANs [Wirelesscom 2005].• Using the 802.11e EDCF to Achieve TCP Upload Fairness Over WLAN Links [WiOpt 05].

Page 26: TCP over 802.11e Doug Leith &  Peter Clifford Hamilton Institute,  Ireland

Hamilton Institute

Ongoing Work

Mixed voice and data.

Results verified experimentally. [D. Malone, A. Ng, D. Leith, SIGCOMM E-WIND ‘05]

Prioritisation using CWmin: