Modeling Per-flow Throughput and Capturing Starvation in CSMA - PowerPoint PPT Presentation

Modeling Per-flow Throughput and Capturing Starvation in CSMA Multi-hop Wireless Networks Michele Garetto Theodoros Salonidis Edward W. Knightly Rice Networks Group http://www.ece.rice.edu/networks

Modeling Per-flow Throughput and Capturing Starvation in CSMA Multi-hop Wireless Networks Michele Garetto Theodoros Salonidis Edward W. Knightly Rice Networks Group http://www.ece.rice.edu/networks

Example : 50 nodes 1000 900 800 700 600 Y (meters) 500 400 300 200 100 0 0 100 200 300 400 500 600 700 800 900 1000 X (meters)

Example : 50 nodes 1000 50 tx-rx pairs 900 (link flows) 800 700 802.11 DCF 600 Y (meters) (CSMA/CA) 500 Saturated traffic 400 300 Perfect channel 200 100 0 0 100 200 300 400 500 600 700 800 900 1000 X (meters)

Example : 50 nodes Sensing range 1000 Single cell 900 800 700 600 Y (meters) 500 400 300 200 100 0 0 100 200 300 400 500 600 700 800 900 1000 X (meters)

Example : 50 nodes 50 Single cell 40 Throughput (pkt/s) All flows receive equal throughput 30 20 10 0 0 5 10 15 20 25 30 35 40 45 50 Node ID

Example : 50 nodes 1000 1000 Sensing Range = 400m 900 900 800 800 700 700 600 600 Y (meters) 500 500 400 400 300 300 200 200 100 100 0 0 0 0 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000 X (meters)

Example : 50 nodes 250 A few rich 200 flows Throughput (pkt/s) 150 100 Many starving flows ! 50 0 0 5 10 15 20 25 30 35 40 45 50 Rank

Example : 50 nodes 250 ideal channel fading + capture 200 Throughput (pkt/s) 150 100 50 0 0 5 10 15 20 25 30 35 40 45 50 55 Rank

Our contributions � Develop an analytical model to compute per-flow throughput in arbitrary network topologies employing 802.11 DCF � Explain the origin of starvation in CSMA- based multi-hop wireless networks � Propose metrics to quantify starvation due to the MAC protocol operation

Related work on CSMA models � Models for single-cell networks (WLANs) � Leverage symmetric channel state � Accurately capture carrier sense, Binary Exponential Backoff, RTS/CTS (e.g. [Bianchi00]) � Models for multi-hop networks � Assumption � Throughput proportional to number of interferers [Boorstyn87, Carvalho04, Kar05] � Cannot capture the CSMA “disproportionalities” or predict zero throughput

Our approach � Decoupling technique � Describe behavior of each node based on its private view of the channel state. � Throughput expression � Express throughput of each node as a function of its � Sensed fraction of busy time (b, Tb) � Collision probability p � Basic iteration � Compute p, (b,T b ) variables of each node subject to the current variables of other nodes. � Iterative solution � Perform basic iteration until convergence

Analytical model � The “channel view” of a node: channel busy due to Node’s transmission Node’s transmission activity of other nodes idle slot collides is successful … … t � Modeled as a renewal-reward process P [event Ts occurs] Throughput (pkt/s) = Average duration of an event (s)

Analytical model � Define: = probability that the node sends a packet = conditional collision probability = conditional busy channel probability � Event probabilities: Success Busy channel Idle Collision … t …

Analytical model � f bianchi (.) � � Deterministic decreasing function of p � Captures Binary Exponential Backoff � Throughput formula: � Unknown variables (different for each node) � Collision probability � Busy channel probability � Average busy time

How can collision probability p be disproportionately large ? The “information asymmetry” scenario 37 pkts/sec a A Flow A->a starves 446 pkts/sec ( = 0.85) B b ( = 0)

Why is collision probability p disproportionately large ? The “information asymmetry” scenario Flow A->a starves 37 (pkts/sec) a due to high packet loss A 446 (pkts/sec) ( = 0.85) Starvation cause: B b ( = 0) A contends randomly B knows when to contend View of A RTS ? View of B … … t

How can busy channel product bT b be disproportionally large ? The “flow-in-the-middle” scenario 30 449 pkts/sec 449 a c b No packet losses Flow A->a starves C B A ( = 0) ( = 0) ( = 0)

Why is busy channel product bT b disproportionally large ? The “flow-in-the-middle” scenario 30 449 449 Flow A->a starves a c b No packet losses Starvation cause: A senses busy medium for a very long time C B A TxOp for A ( = 0) ( = 0) ( = 0) � Channel view of A: B B B B C C C C

Computation of busy channel parameters (b,T b ) for flow Aa Challenge � Not all neighbors of A are mutually � within range and their activities are inter- dependent. Clique computation � A Find minimum number of maximal � cliques M covering all neighbors

Computation of busy channel parameters (b,T b ) for flow Aa M=3 Challenge � Not all neighbors of A are mutually � within range and their activities are inter- 2 1 4 dependent. Clique computation � A Find minimum number of maximal � cliques M covering all neighbors 6 7 Virtual nodes (VN) graph � 5 VN = set of non-empty clique � intersections VN Graph: Connect two VNs if they � 3 share at least one clique

Computation of busy channel parameters (b,T b ) for link Aa 1 Challenge 4 2 � Not all neighbors of A are mutually � within range and their activities are inter- dependent. 7 Clique computation � Find minimum number of maximal � 5 6 cliques M covering all neighbors Virtual nodes (VN) graph � VN = set of non-empty clique � intersections VN Graph: Connect two VNs if they 3 � share at least one clique Computation of busy period � Find the aggregate busy time around � node i based on VN activities

Computation of collision probability p for link Aa � 4 classes of packet loss due to link Bb 1) Coordinated losses: P co 2) Information Asymmetry: P ia a A a A B b B b 4) Far Hidden terminals: P fh 3) Near Hidden terminals: P nh a A a A b B b B p = 1 – (1-P co )(1-P ia )(1-P nh )(1-P fh )

Network solution � Basic iteration � Compute p, (b,Tb) of each node subject to the variables of other nodes. � Network solution � Multivariate system of coupled non-linear equations � Perform basic iteration until convergence � Model features � Incorporates all starvation effects due to CSMA MAC � Can analyze arbitrary topology � Predicts individual flow throughput � Supports non-saturated flows

Model vs Sim – 50-nodes example 300 sim model 250 Throughput (pkt/s) 200 150 100 50 0 0 5 10 15 20 25 30 35 40 45 50 Rank

Measuring Starvation � Objectives � Capture how individual flows are treated by different solutions � Distinguish between imbalance due to topology (number of contenders) and starvation due to the MAC protocol � Reference system: Slotted Aloha � Starvation structurally eliminated

Conclusions � Multi-hop wireless networks employing 802.11 (or other variants of CSMA) are subject to severe starvation (under high load) � This is a fundamental problem CSMA due to lack of coordination between out-of-range transmitters � We developed an analytical model to predict per- flow throughput in arbitrary topologies and characterize starvation

Thank you For more information: www.ece.rice.edu/~thsalon