Wireless & Mobile Networking CS 752/852 - Spring 2011 Tamer Nadeem Dept. of Computer Science Lec #7: Medium Access Control – V Data Rate Control Slides adapted from Romit Roy Choudhury (Duke)

Page 2 Spring 2011 CS 752/852 - Wireless and Mobile Networking What is Data Rate ? Number of bits that you transmit per unit time under a fixed energy budget Too many bits/s: Each bit has little energy -> Hi BER Too few bits/s: Less BER but lower throughput

Page 3 Spring 2011 CS 752/852 - Wireless and Mobile Networking Some Basics Floor Noise Data Rate Received Power Channel Bandwidth Bit error (p) for BPSK and QPSK : SNR Friss’ Equation: Varying with time and space How do we choose the rate of modulation

Page 4 Spring 2011 CS 752/852 - Wireless and Mobile Networking b – Transmission rates

Page 5 Spring 2011 CS 752/852 - Wireless and Mobile Networking Static Rates SINR time # Estimate a value of SINR # Then choose a corresponding rate that would transmit packets correctly most of the times # Failure in some cases of fading  live with it

Page 6 Spring 2011 CS 752/852 - Wireless and Mobile Networking Adaptive Rate-Control SINR time # Observe the current value of SINR # Believe that current value is indicator of near-future value # Choose corresponding rate of modulation # Observe next value # Control rate if channel conditions have changed

Page 7 Spring 2011 CS 752/852 - Wireless and Mobile Networking Is there a tradeoff ? CE A D B Rate = 10

Page 8 Spring 2011 CS 752/852 - Wireless and Mobile Networking Is there a tradeoff ? CE A D B Rate = 20 Rate = 10 What about length of routes due to smaller range ?

Page 9 Spring 2011 CS 752/852 - Wireless and Mobile Networking Any other tradeoff ? Will carrier sense range vary with rate

Page 10 Spring 2011 CS 752/852 - Wireless and Mobile Networking Total interference CE A D B Rate = 20 Rate = 10 Carrier sensing estimates energy in the channel. Does not vary with transmission rate Carrier sensing estimates energy in the channel. Does not vary with transmission rate

Page 11 Spring 2011 CS 752/852 - Wireless and Mobile Networking Bigger Picture Rate control has variety of implications Any single MAC protocol solves part of the puzzle Important to understand e2e implications Does routing protocols get affected? Does TCP get affected? … Good to make a start at the MAC layer RBAR OAR Opportunistic Rate Control …

© Gavin Holland A Rate-Adaptive MAC Protocol for Multi-Hop Wireless Networks Gavin Holland HRL Labs Nitin Vaidya Paramvir Bahl UIUC Microsoft Research MOBICOM’01 Rome, Italy

Page 13 Spring 2011 CS 752/852 - Wireless and Mobile Networking Background Current WLAN hardware supports multiple data rates b – 1 to 11 Mbps a – 6 to 54 Mbps Data rate determined by the modulation scheme

Page 14 Spring 2011 CS 752/852 - Wireless and Mobile Networking Modulation schemes have different error characteristics Problem BER SNR (dB) 1 Mbps 8 Mbps But, SINR itself varies With Space and Time But, SINR itself varies With Space and Time

Page 15 Spring 2011 CS 752/852 - Wireless and Mobile Networking Impact Large-scale variation with distance (Path loss) SNR (dB) Distance (m) Mean Throughput (Kbps) Path Loss 1 Mbps 8 Mbps

Page 16 Spring 2011 CS 752/852 - Wireless and Mobile Networking Impact Small-scale variation with time (Fading) SNR (dB) Time (ms) Rayleigh Fading 2.4 GHz 2 m/s LOS

Page 17 Spring 2011 CS 752/852 - Wireless and Mobile Networking Question Which modulation scheme to choose? SNR (dB) Time (ms) Distance (m) 2.4 GHz 2 m/s LOS

Page 18 Spring 2011 CS 752/852 - Wireless and Mobile Networking Answer  Rate Adaptation Dynamically choose the best modulation scheme for the channel conditions Mean Throughput (Kbps) Distance (m) Desired Result

Page 19 Spring 2011 CS 752/852 - Wireless and Mobile Networking Design Issues How frequently must rate adaptation occur? Signal can vary rapidly depending on: carrier frequency node speed interference etc. For conventional hardware at pedestrian speeds, rate adaptation is feasible on a per-packet basis Coherence time of channel higher than transmission time

Page 20 Spring 2011 CS 752/852 - Wireless and Mobile Networking Cellular networks Adaptation at the physical layer Impractical for in WLANs For WLANs, rate adaptation best handled at MAC Adaptation  At Which Layer ? D C BA CTS: 8 RTS: Sender Receiver RTS/CTS requires that the rate be known in advance Why?

Page 21 Spring 2011 CS 752/852 - Wireless and Mobile Networking Who should select the data rate? Collision is at the receiver Channel conditions are only known at the receiver SS, interference, noise, BER, etc. The receiver is best positioned to select data rate A B

Page 22 Spring 2011 CS 752/852 - Wireless and Mobile Networking Previous Work PRNet Periodic broadcasts of link quality tables Pursley and Wilkins RTS/CTS feedback for power adaptation ACK/NACK feedback for rate adaptation Lucent WaveLAN “Autorate Fallback” (ARF) Uses lost ACKs as link quality indicator

Page 23 Spring 2011 CS 752/852 - Wireless and Mobile Networking Lucent WaveLAN “Autorate Fallback” (ARF) Sender decreases rate after N consecutive ACKS are lost Sender increases rate after Y consecutive ACKS are received or T secs have elapsed since last attempt B A DATA 2 Mbps Effective Range 1 Mbps Effective Range

Page 24 Spring 2011 CS 752/852 - Wireless and Mobile Networking Performance of ARF Time (s) Rate (Mbps) SNR (dB) Time (s) – Slow to adapt to channel conditions – Choice of N, Y, T may not be best for all situations Attempted to Increase Rate During Fade Dropped Packets Failed to Increase Rate After Fade

Page 25 Spring 2011 CS 752/852 - Wireless and Mobile Networking RBAR Approach Move the rate adaptation mechanism to the receiver Better channel quality information = better rate selection Utilize the RTS/CTS exchange to: Provide the receiver with a signal to sample (RTS) Carry feedback (data rate) to the sender (CTS)

Page 26 Spring 2011 CS 752/852 - Wireless and Mobile Networking RTS carries sender’s estimate of best rate CTS carries receiver’s selection of the best rate Nodes that hear RTS/CTS calculate reservation If rates differ, special subheader in DATA packet updates nodes that overheard RTS Receiver-Based Autorate (RBAR) Protocol C BA CTS (1) RTS (2) 2 Mbps 1 Mbps D DATA (1) 2 Mbps 1 Mbps

Page 27 Spring 2011 CS 752/852 - Wireless and Mobile Networking Performance of RBAR Time (s) SNR (dB) Time (s) Rate (Mbps) Time (s) RBAR ARF

Page 28 Spring 2011 CS 752/852 - Wireless and Mobile Networking Question to the class There are two types of fading Short term fading Long term fading Under which fading is RBAR better than ARF ? Under which fading is RBAR comparable to ARF ? Think of some case when RBAR may be worse than ARF

Page 29 Spring 2011 CS 752/852 - Wireless and Mobile Networking Implementation into

Page 30 Spring 2011 CS 752/852 - Wireless and Mobile Networking Implementation into PLCP Header

Page 31 Spring 2011 CS 752/852 - Wireless and Mobile Networking Implementation into Encode data rate and packet length in duration field of frames Rate can be changed by receiver Length can be used to select rate Reservations are calculated using encoded rate and length New DATA frame type with Reservation Subheader (RSH) Reservation fields protected by additional frame check sequence RSH is sent at same rate as RTS/CTS New frame is only needed when receiver suggests rate change Frame Control Duration DA SABSSID Sequence Control Body FCS Reservation Subheader (RSH) WHY

Page 32 Spring 2011 CS 752/852 - Wireless and Mobile Networking Ns-2 with mobile ad hoc networking extensions Rayleigh fading Scenarios: single-hop, multi-hop Protocols: RBAR and ARF RBAR Channel quality prediction: SNR sample of RTS Rate selection: Threshold-based Sender estimated rate: Static (1 Mbps) Performance Analysis BER SNR (dB) 1E-5 2 Mbps Threshold 8 Mbps Threshold

Page 33 Spring 2011 CS 752/852 - Wireless and Mobile Networking Performance Results Single-Hop Network

Page 34 Spring 2011 CS 752/852 - Wireless and Mobile Networking Single-Hop Scenario AB Mean Throughput (Kbps) Distance (m)

Page 35 Spring 2011 CS 752/852 - Wireless and Mobile Networking CBR source Packet Size = 1460 Varying Node Speed UDP Performance RBAR performs best Declining improvement with increase in speed Adaptation schemes over fixed RBAR over ARF Some higher fixed rates perform worse than lower fixed rates Mean Throughput (Kbps) Mean Node Speed (m/s) RBAR ARF WHY?

Page 36 Spring 2011 CS 752/852 - Wireless and Mobile Networking Varying Node Speed TCP Performance RBAR again performs best Overall lower throughput and sharper decline than with UDP Caused by TCP’s sensitivity to packet loss More higher fixed rates perform worse than lower fixed rates FTP source Packet size = 1460 Mean Throughput (Kbps) Mean Node Speed (m/s) RBAR ARF

Page 37 Spring 2011 CS 752/852 - Wireless and Mobile Networking No Mobility UDP Performance RSH overhead seen at high data rates Can be reduced using some initial rate estimation algorithm Limitations of simple threshold-based rate selection seen Generally, still better than ARF Distance (m) CBR source Packet size = 1460 RBAR ARF Mean Throughput (Kbps) WHY?

Page 38 Spring 2011 CS 752/852 - Wireless and Mobile Networking No Mobility UDP Performance RBAR-P – RBAR using a simple initial rate estimation algorithm Previous rate used as estimated rate in RTS Better high-rate performance Other initial rate estimation and rate selection algorithms are a topic of future work Distance (m) CBR source Packet size = 1460 RBAR-P Mean Throughput (Kbps) Why useful ?

Page 39 Spring 2011 CS 752/852 - Wireless and Mobile Networking RBAR Summary Modulation schemes have different error characteristics Significant performance improvement may be achieved by MAC-level adaptive modulation Receiver-based schemes may perform best Proposed Receiver-Based Auto-Rate (RBAR) protocol Implementation into Future work RBAR without use of RTS/CTS RBAR based on the size of packets Routing protocols for networks with variable rate links

Page 40 Spring 2011 CS 752/852 - Wireless and Mobile Networking Questions?

Three scenarios: 1)Both nodes use 11 Mb/s 2)Both nodes use 1 Mb/s 3)n 1 uses 11 Mb/s, n 2 uses 1 Mb/s Fast user’s throughput is severely degraded whereas Slow user achieves throughput even larger than it expects Multi-rate Anomaly Multi-rate Anomaly [Heusse03]

OAR: An Opportunistic Auto-Rate Media Access Protocol for Ad Hoc Networks B. Sadeghi, V. Kanodia, A. Sabharwal, E. Knightly Rice University Slides adapted from Shawn Smith

Page 43 Spring 2011 CS 752/852 - Wireless and Mobile Networking Motivation Consider the situation below ARF? RBAR? A BC

Page 44 Spring 2011 CS 752/852 - Wireless and Mobile Networking Motivation What if A and B are both at 56Mbps, and C is often at 2Mbps? Slowest node gets the most absolute time on channel? A BC A B C Timeshare Throughput Fairness vs Temporal Fairness

Page 45 Spring 2011 CS 752/852 - Wireless and Mobile Networking Opportunistic Scheduling Goal Exploit short-time-scale channel quality variations to increase throughput. Issue Maintaining temporal fairness (time share) of each node. Challenge Channel info available only upon transmission

Page 46 Spring 2011 CS 752/852 - Wireless and Mobile Networking Opportunistic Auto-Rate (OAR) In multihop networks, there is intrinsic diversity Exploiting this diversity can offer benefits Transmit more when channel quality great Else, free the channel quickly RBAR does not exploit this diversity It optimizes per-link throughput

Page 47 Spring 2011 CS 752/852 - Wireless and Mobile Networking OAR Idea Basic Idea If bad channel, transmit minimum number of packets If good channel, transmit as much as possible A B CD A C Data

Page 48 Spring 2011 CS 752/852 - Wireless and Mobile Networking Why is OAR any better ? alternates between transmitters A and C Why is that bad A B CD A C Data Is this diagram correct ?

Page 49 Spring 2011 CS 752/852 - Wireless and Mobile Networking Why is OAR any better ? Bad channel reduces SINR  increases transmit time Fewer packets can be delivered A B CD A C Data

Page 50 Spring 2011 CS 752/852 - Wireless and Mobile Networking OAR Protocol Steps Transmitter estimates current channel Can use estimation algorithms Can use RBAR, etc. If channel better than base rate (2 Mbps) Transmit proportionally more packets E.g., if channel can support 11 Mbps, transmit (11/2 ~ 5) pkts OAR upholds temporal fairness Each node gets same duration to transmit Sacrifices throughput fairness  the network gains !!

Page 51 Spring 2011 CS 752/852 - Wireless and Mobile Networking MAC Access Delay Simulation Back to back packets in OAR decrease the average access delay Increase variance in time to access channel Why?

Page 52 Spring 2011 CS 752/852 - Wireless and Mobile Networking OAR Protocol Rates in IEEE b: 2, 5.5, and 11 Mbps Number of packets transmitted by OAR ~

Page 53 Spring 2011 CS 752/852 - Wireless and Mobile Networking Simulations Three Simulation experiments 1.Fully connected networks: all nodes in radio range of each other Number of Nodes, channel condition, mobility, node location 2.Asymmetric topology 3.Random topologies Implemented OAR and RBAR in ns-2 with extension of Ricean fading model [Punnoose et al ‘00]

Page 54 Spring 2011 CS 752/852 - Wireless and Mobile Networking Fully Connected Setup Every node can communicate with everyone Each node’s traffic is at a constant rate and continuously backlogged Channel quality is varied dynamically

Page 55 Spring 2011 CS 752/852 - Wireless and Mobile Networking Fully Connected Throughput Results OAR has 42% to 56% gain over RBAR Increase in gain as number of flows increases Note that both RBAR and OAR are significantly better than standard (230% and 398% respectively)

Page 56 Spring 2011 CS 752/852 - Wireless and Mobile Networking Asymmetric Topology Results OAR maintains time shares of IEEE Significant gain over RBAR

Page 57 Spring 2011 CS 752/852 - Wireless and Mobile Networking OAR thoughts OAR does not offer benefits when OAR may not be suitable for applications like With TCP how can OAR get affected ?

Page 58 Spring 2011 CS 752/852 - Wireless and Mobile Networking OAR thoughts OAR does not offer benefits when Neighboring nodes do not experience diverse channel conditions Coherence time is shorter than N packets With TCP can OAR get affected ? Back-to-back packets can increase TCP performance However, bottleneck bandwidth can get congested quick Also, variance of RTT can increase

Multi-user Data Rate Adjustment Algorithm for Enterprise Networks Nazif Tas, Tamer Nadeem, Ashok Agrawala INFOCOM 2011

Previous Work  Data Rate Adjustment Mechanisms  Agreement between sender and the transmitter on rate to use.  Sender-based: statistical.  Easy, resource efficient.  ARF, AARF, CARA, AMRR…  Receiver-based: receiver chooses data rate and notifies the sender.  More accurate, unnecessary resource usage.  RBAR  Multi-rate Anomaly Solutions  Infrastructure-based solutions  AP selection, traffic scheduling, etc  [Wang, Ji, Im04, …]  MAC Layer modifications  Packet bursting, minimum CW adjustment, etc  [Mai09, Yah-hong07, …]  Higher Layer adjustments  TCP window adjustment, traffic slow down  [Yoo08, Kashibuchi09, …] Requires holistic view of the system Requires modifications in the standards Can we lessen the effects of Multi-Rate anomaly efficiently in a distributed & seamless manner with no modifications in the standards? Requires non-seamless cross layer support

IEEE DCF CSMA/CA: carrier sensing Binary Exponential Backoff Binary Exponential Backoff (BEB) IEEE BEB

Fairness [Tan04] : throughput seen by data rate d users in a network with n s number of slow users and n f number of fast users. Fast Slow Multi-rate Anomaly Baseline Fairness Criteria

Bianchi Model with Retry Limit For data rate d i, k: retry limit Can we use arbitrary limits to mitigate Multi-rate Anomaly? Our extension supports arbitrary retry limits per user groups

Multi-user Data Rate Adjustment Algorithm (MORAL) Automatically adjust the retry limits of each user such that Improve fairness: Take the fair throughput share without hurting others. Distributed computation: No centralized decision making. Standard Coherence: No packet modification, restructuring or protocol alteration. “The first thing to note is that [happiness] is a relative quality. We experience it differently according to our circumstances. What makes one person glad may be a source of suffering to another” Two step operation: 1.Collect information about the overall network. 2.Adjust the retry limit accordingly MAC Observe Decide Act LA

MORAL Details: Heuristic Principles After each transmission cycle, we have two pieces of information: c current : number of transmissions this cycle for the current node Successful transmission (c current =1) Failed transmission(c current =0) c calculated : Fair number of transmissions for the current node

Experiments I Default MORAL 11, 1 Mb/s – 20 Users Each 1 Mb/s 11 Mb/s 60% 61.3%

Experiments II 11, 5.5, 2, 1 Mb/s– 10 Users Each DefaultMORAL 31.8% 43.0%

Experiments III 93.8% 30 1 Mb/s, Mb/s 1 Mb/s 11 Mb/s 1 Mb/s 11 Mb/s Default MORAL 20.8% 48.0%

Experiments IV 11, 1 Mb/s – 20 Users Each 23.9% 40.2%

Summary MORAL is an effective MAC layer link adaptation mechanism which lessens the effects of multi-rate anomaly and promises - Better fairness - Increased throughput MORAL is - Fully compatible with the current standards - Totally distributed - Highly adaptable - Easily deployable in any IEEE compliant devices

Page 71 Spring 2011 CS 752/852 - Wireless and Mobile Networking Other Ideas (Briefly)

Page 72 Spring 2011 CS 752/852 - Wireless and Mobile Networking Exploiting Diversity in Rate Adaptation Yet another idea exploits multiple user diversity Among many intermediate nodes, who has best channel Use that node as forwarding node Forwarding node can change with time Due to channel fluctuations at different time and space WLAN Channel Conditions SNR TIME USERS AP

Page 73 Spring 2011 CS 752/852 - Wireless and Mobile Networking The Protocol Overview SIFS DIFS GRTS ACK 0~2 CTS 1 CTS k CTS 2 … sender user 1 user 2 user k DATA 0DATA 1 DATA 2 SF MAD using Packet Concatenation (PAC) Since at least one intermediate node is likely to have good channel condition, transmitter can transmit at a high data rate or concatenate Multiple packets Choosing subset of neighbor-group is important Coherence time of channel must be greater than packet chain Group needs to really have independent channel gain Correlated channel gains can lead to performance hit.

Page 74 Spring 2011 CS 752/852 - Wireless and Mobile Networking What lies ahead ? Routing based on rate-control Choosing routes that contain high-rate links ETX metric proposed from MIT accomodates link character Opportunistic routing from MIT again – takes neighbor diversity into account (best paper Sigcomm 2005) Fertile area for a project … Dual of rate-control is power control One might be better than the other Decision often depends on the scenario  open problem Directional antennas for DD link for data/ack Rate control can be introduced  Not been studied yet … many many more

Page 75 Spring 2011 CS 752/852 - Wireless and Mobile Networking Questions ?