Joseph Camp, Edward Knightly ACM MobiCom 2008

Modulation Rate Adaptation in Urban and Vehicular Environments: Cross-Layer

Implementation and Experimental Evaluation

Joseph Camp, Edward KnightlyACM MobiCom 2008

BackgroundWhy need modulation rate adaptation?

Time-varying link quality – Mobility of sender, receiver, or obstacles – Multiple paths existing

Ideal modulation rate for channel condition

Ideal rate

For Protocol DesignerUse available information (loss, SNR, …) to

track ideal modulation rate

Rate Selection Protocol

SNR, loss, … Rate choice

ProblemALL existing rate adaptation algorithms fail

to track the ideal rateUrban propagation environment ( ex: multi-

path )Even with non-mobile sender and receiver

Result becomes loss and under-utilization

Select rate

ProposeUnderstand the origins of the failure to track

link variationCompare the result in in-door (lab) and out-

door (urban or downtown )

Propose (cont)Unified Implementation Platform

Implement multiple algorithms on a common platform

Extract General Rate Adaptation PrinciplesEvaluate rate selection accuracy packet-by-packetCompare against ideal rate found via exhaustive

searchUse repeatable controlled channelsAccurately measured outdoor channels

Design core mechanisms to track real-world link variation

WARP - Wireless Open-Access Research Platform Limits of Off-the-shelf platforms

Programmability and observability

WARP is clean-slate MAC and PHYNeeded to implement CSMA/CA (802.11-like MAC)

Cross-layer rate adaptation frameworkCore mechanisms for rate selection protocolsChannel measurementsEvaluation of selected rate versus ideal rate

Core Mechanisms for Rate Adaptation Loss-triggered Rate Adaptation

Consecutive-Packet DecisionHistorical-Decision

Collision/Fading Differentiation

SNR-triggered Rate AdaptationEqual Air-time Assurance

Rate Adaptation Accuracy MatricesIdeal rate found via exhaustive search of

channel conditionConsider case where at least one modulation

rate succeeds

Channel Condition

Idea

l Rat

e

Rate Adaptation Accuracy Matrices (cont)Rate Selection Accuracy Categories

Over-selection (loss)Accurate (achieving optimal rate)Under-selection (under-utilization)

Channel Condition

Idea

l Rat

e

Over-selection

Accurate

Under-selection

In-door (lab) experimentsImpact of Coherence Time

Increase fading of the channel to evaluate if rate adaptation can track

In-door (lab) experiments (cont)Similar performance with long coherence of

channelSNR: high overhead penaltyEqual Air-time Assurance: overcomes overhead

penaltyDissimilar performance at short coherence of

channel

Rate Choice InaccuraciesConsecutive low due to under-selectionSNR: extremely low throughput due to over-

selection

SNR-based Coherence Time SensitivityFast to slow channel fading

Accurate at long coherenceOver-select at <1ms

Over-selection caused by coherence time sensitivity of SNR-rate relationship

1 ms

accurate

Coherence Time Training for SNRConsider different SNR thresholds according

to coherence time

Consideration of SNR and Coherence TimeConsider different SNR thresholds according

to coherence timeIdeal rate = f(SNR, CT)

Retrain SNR-based decision (for the same protocol)

Joint consideration of SNR and coherence time provides large gains

Residential Urban and Downtown ScenariosResidential Urban (TFA)

Single-family residential, dense foliageCoherence Time: 100 ms on averageDriven to 15 ms with mobility of

scatterers (in static topology)Downtown Houston

Both sides of street lined with tall buildings (strong multipath)

Coherence Time: 80 ms on averageDriven to 300 us with mobility of scatterers (in

static topology)

Rate adaptation accuracy in outdoorResidential urbanInaccurate in outdoor settings for loss-

triggeredConsecutive-decision due to the mobility of

scattersHistorical-decision due to different modulation

rate selection

Rate adaptation accuracy in outdoor (cont)

Downtown - strong multipathForce loss-based to under-selectSNR: over and under-select with low coherence

time

Static Sender to Mobile Receiver -UrbanSNR protocols are able to plateau for >4 sec

Per-packet decisionLoss-based protocols are only able to spike to

suboptimal rate choicesLoss sensitivity prevents protocol from

tracking

static

mobile

SummaryLoss-based core mechanisms under-select with

Fast-fading, interference, competing links (even with collision/fading differentiation), and mobile environments

SNR-based mechanisms over-select with fast-fading but haveLarge gains from considering SNR and coherence time

jointlyRobustness to interference, competing links, and mobility

Despite use of 4-way handshake, SNR-based protocols outperform loss-based protocols in practical environments