The System Issues of Rate Adaptation May 19 The System Issues of Rate Adaptation Hujun Yin hujun_yin@vivato.net Hujun Yin,Vivato

Motivation Traditionally, rate adaptation is done at link level May 19 Motivation Traditionally, rate adaptation is done at link level Ignoring the impact of other links at MAC Ignoring the impact of link adaptation on other links Understand the impact of rate adaptation on system performance Considering the synergy of multiple links Identify the issues of link based rate adaptation Provide insight into system based rate adaptation Hujun Yin,Vivato

Assumptions Network is operating in infrastructure mode May 19 Assumptions Network is operating in infrastructure mode One AP and multiple clients Communication only between AP and clients No multiuser reception capability If collision happens, at most one packet can be successfully received No non-802.11 interference Packet error is only introduced by low SNR or packet collision Hujun Yin,Vivato

Link Based Rate Adaptation May 19 Link Based Rate Adaptation SNR based rate adaptation Chooses data rate from a rate-SNR lookup table PER based rate adaptation Monotonic relation between data rate and PER Adjust data rate so that PER keeps constant Decreases data rate when PER raises Increases data rate when PER drops Rate adaptation is done from link perspective without considering the impact on system performance Hujun Yin,Vivato

System Performance Criteria May 19 System Performance Criteria Maximize total throughput under certain constraints Only consider PER constraint here S : total throughput; N: number of nodes; Ri: data rate of node i; pi: PER of node i; : utilization time of node i; : channel idle time; : channel busy time; p0: PER constraint Hujun Yin,Vivato

System PER Model The packet error is contributed by two sources: May 19 System PER Model The packet error is contributed by two sources: pe: wireless channel & noise (SNR related) pc: MAC collisions pe is a function of data rate and SNR pc is a function of packet arrival rate and channel access mechanism Hujun Yin,Vivato

Rate Adaptation Criteria Revised May 19 Rate Adaptation Criteria Revised Optimal data rates are not only a function of SNR, but also a function of network traffic Data rate adaptation is link dependent Traditional link adaptation only considers SNR is suboptimal Data rate shall be optimized with regard to interaction between links (collision) Hujun Yin,Vivato

Examples: TDMA MAC Each node is assigned to unique time slots May 19 Examples: TDMA MAC Each node is assigned to unique time slots No MAC collisions: pc=0 Links are isolated Maximize overall throughput equivalents to maximize individual link throughput Link based rate adaptation is optimal Hujun Yin,Vivato

Examples: ALOHA Random Access May 19 Examples: ALOHA Random Access Total throughput Probability of collision If N>>1 nodes are identical with aggregated traffic rate l and data rate R Pc decreases with data rate R pe increases with data rate R PER=pe+pc is no longer a monotonic function of rate R Hujun Yin,Vivato

Instability of Link PER Based Rate Adaptation May 19 Instability of Link PER Based Rate Adaptation Consider two rates: R1=R, pe1=10% R2=R/10, pe2=0 Network operates at 20% PER Pc=10%, l=R/10, S=0.081R pe1->15% Switch to R2: pe2=0, pc=63%, PER=63%, S=0.037R Stay at R1:pe1=15%, pc=10%, PER=24%, S=0.081R Rate adaptation to improve PHY PER may result in much worse system performance Hujun Yin,Vivato

Examples: 802.11 DCF MAC CSMA/CA Exponential Backoff Packet 2 DIFS May 19 Examples: 802.11 DCF MAC CSMA/CA Exponential Backoff Packet 2 DIFS Packet 1 Ack Tx Starts Sensing SIFS DIFS: Distributed Interframe Space SIFS: Short Interframe Space DIFS + Backoff Hujun Yin,Vivato

802.11 Packet Collision Probability May 19 802.11 Packet Collision Probability 802.11a initial backoff window only has 16 slots The probability of collision is close to 10% with two saturated nodes The probability of collision is close to 30% with 5 saturated nodes* The PER caused by packet collision can not be ignored in CSMA based MAC like 802.11a Rate adaptation in 802.11 is an involved system problem** High packet collision probability Variable packet size *G. Bianchi, ”Performance analysis of the IEEE 802.11 distributed coordination function”, IEEE JSAC, March 2000. ** H. Yin, “Multirate 802.11a Networks: System Performance Evaluation”, CISS, March 2002. Hujun Yin,Vivato

Conclusion Link adaptation depends on: Channel condition (SNR) May 19 Conclusion Link adaptation depends on: Channel condition (SNR) Activity of other links Rate adaptation considering PHY only is sub-optimal in random access MAC Rate adaptation in PHY simulation may provide biased results to random access MAC Limiting PHY PER at low level may be artificial if the nature of MAC contention is not considered Data rate adaptation has to be considered at system level The black-box approach can not provide accurate PHY-MAC interface abstraction for all possible MAC proposals Hujun Yin,Vivato

Ways to Improve Rate Adaptation Abstraction May 19 Ways to Improve Rate Adaptation Abstraction Partition rate adaptation Coarse rate adaptation at MAC Deals the network aspect of rate adaptation Fine rate adaptation at PHY Deals the channel aspect of rate adaptation Link adaptation localization Other links only impacted if the rate of one link varies significantly Coarse adaptation at MAC sets constraints on the fine rate adaptation at PHY Hujun Yin,Vivato