QoS Provisioning for IEEE 802.11 MAC Protocols Ye Ge The Ohio State University Jennifer C. Hou University of Illinois at Urbana-Champaign Sunghyun Choi Seoul National University Presented by Andrew Tzakis

Outline Motivation Overview of IEEE 802.11 MAC Protocols PMAC --- An Analytical Model for Multi-Class p-persistent version of 802.11 DCF to Achieve Flow Rate Differentiation Implementation Issues for PMAC Analysis of Arbitration IFS (AIFS) in 802.11e Conclusion

Motivation Wireless Local Area Networks (WLANs) based on IEEE 802.11 standard are getting very popular. Existing Applications Future Applications E-Mails File Transfer Web Access and Browsing …… Audio/Video Streaming HDTV Smart Home Networking Best Effort Require QoS Support ! (delay, rate etc.)

Existing Solutions Legacy IEEE 802.11 MAC

Existing Solutions - DCF Distributed Coordination Function (DCF) Based on Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) ACK frames are used to acknowledge the successful reception of data frames Each station maintains a Contention Window (CW) and a Backoff Timer

Existing Solutions - DCF In the backoff stage, the station decreases the Backoff Timer by one time slot for every consecutive idle slot after the medium has been idle for a DIFS. The backoff timer decreasing is suspended whenever the channel is sensed busy and resumed after the channel has been idle for a DIFS again. The station is allowed to transmit whenever the backoff timer reaches zero The Backoff Timer is initialized according to the Binary Exponential Backoff (BEB) algorithm Backoff Timer = Random() × aSlotTime Random() is a random integer drawn from an uniform distribution on the interval [0, CW] CW is double every unsuccessful transmission until reaching its maximal value allowed, that is, CWnew = min(CWold * 2 – 1, CWmax ) CWinitial = CWmin= 31, CWmax= 1023 Limited retry times Basic Access

Existing Solutions - PCF Point Coordinated Function (PCF) Contention Free Period (CFP) and Contention Period (CP) alternatively appear periodically. Within CFP, the Access Point gains control of the channel using PCF Inter-frame Space (PIFS), which is shorter than DIFS. AP may poll each station to transmit in CFP without incurring frame collision.

Problems with the current solutions Why is DCF not enough? It treats all data traffic in a FCFS, best-effort manner All stations contend for the wireless medium with the same priority No differentiation between data flows with QoS requirements Why is PCF not enough? Periodical appearance of CFP and CP limiting its flexibility (because it is difficult to find a repetition period fits all flow requirements) Lacking mechanism to specify traffic requirement

IEEE 802.11e Draft EDCA is very similar to DCF which already has many analytical models to characterize its data transmission. EDCA will be the focus of this paper.

IEEE 802.11e Draft - EDCA Enhanced Distributed Channel Access Multiple Access Priorities Multiple Queues Per-queue channel access Internal collision resolution AIFS[i] = a SIFSTime + AIFSN[i] × aSlotTime

IEEE 802.11e Draft – EDCA Default EDCA Parameter Set AC CWmin CWmax AIFSN AC_BK aCWmin aCWmax 7 AC_BE 3 AC_VI (aCWmin-1)/2 2 AC_VO (aCWmin-3)/4

Proposed Solution - PMAC Provide (proportional) service differentiation and achieve pre-specified targeted throughput ratios, while still maximizing the total system capacity. Through tuning IEEE 802.11e EDCA parameters. Less is more. When trying to achieve QoS guarantees, it may not be desirable to tune multiple parameters.

An Analytical Model for Multi-Class Service Differentiation PMAC An Analytical Model for Multi-Class Service Differentiation p-persistent model: After idle for DIFS time, each station transmits at the start of each idle slot with fixed probability p. (backoff interval is sampled from a geometric distribution) 1-p p DIFS

PMAC virtual transmission time (Tv) successful transmission virtual transmission time (Tv) collision DIFS Channel throughput ρ can be expressed as where m is the data packet payload size

PMAC Assumptions: There are P classes of stations, each of which contains Ni stations. A class-i station transmits its frame in a slot with probability pi in the p-persistent version of IEEE 802.11. All the stations always have packets ready for transmission (i.e., the asymptotic condition holds). The size, mi, of a packet sent by a class-i station is uniformly distributed between (x0, x1).

PMAC - virtual transmission length Average virtual transmission time length E(TV) can be expressed as virtual transmission time (Tv) collision successful Ncol 1 Ncol 2 … packet collision times in a virtual transmission time the duration of a collision period Where: the length of an idle period the length of the successful transmission To find E(Tv),only need to calculate E(Ncol), E(Tc) and E(I)

PMAC - E(Ncol) Derivation of E(Ncol) Pcollision is the probability that more than one will transmit given that at least one is transmitting E(Ncol) is the number of transmission attempts times the probability of colliding raised to the number of collisions, time the probability for one success.

PMAC - E(Tc) Derivation of E(Tc) where All combin. Time for a given combin. Probability of a combin. where (Sum for every class and every number of nodes in each class){ Time it takes to collide given a specific combination of colliding stations * the probability that that specific combination of colliding stations would occur.

PMAC - E(I) Derivation of E(I) Time for one slot * Sum of all i, i * ( ( Probability that at least one station is transmits )( Probability the number of transmitting stations is equal to zero )^(i))

PMAC – Ppkt(i) Define [( Number of stations of class i ( (Probability one from class i (pi) transmits * Probability that all others from class i do not transmit (1-pi)^(ni-1) ) ) Probability that all other classes do not transmit] -------------------------------------------------------------- Sum of the probabilities that one from each class will transmit while all other stations do not transmit.

PMAC – Throughput ratios The channel throughput attained by class-i stations can be expressed as and the throughput ratio between class-i and class-j traffic can be derived as and the throughput ratio between a class-i station and a class-j station can be expressed as

PMAC – Relative ratio to class 1 Suppose all the flow throughput ratios are given in terms of the relative ratio to a class 1 flow (i.e. ) and the data frame size of all traffic classes are of the same distribution ( E(Mi) = E(Mj) ). Constraint! To achieve the desired throughput ratio, pi has to be set to the value determined by the following equation Protocol capacity can be optimized by finding optimal value of p1 to minimize E(Tv) subject to the constraint of the above relation between pi and p1.

Simulation Results (ns-2 simulation) throughput They choose different values of p1 and used the analytical model to get p2 using a ratio of 2.0 and then the throughput is measured. each of which has a CBR trafc source that generates packets at a rate high enough to emulate the asymptotic condition All stations send CBR packets of sizes 500 bytes to the base station. We do not consider TCP trafc, because we are primarily interested in the performance of IEEE 802.11 EDCA optimal value

Simulation Results per flow throughput ratio

Implementation Issues for PMAC Finding an Approximate Solution Mapping p to Contention Window Size Dealing with Network Dynamics

An Approximate Solution Based on the observation that under normal, non-contention conditions, even if only class I nodes are active, the probability that at least one station starts to transmit is far less than one. By assuming we can make the following approximate

An Approx. Solution - simulation Packet size = 500 bytes

Mapping p to cw From “Performance evaluation and enhancement of the CSMA/CA MAC protocol for 802.11 wireless LANs, the probability that a station transmits in a randomly-chosen slot is: By setting CWmin and CWmax to CW* where CW* is chosen such that the probability that a station transmits in a slot is equal to the optimal transmission probability derived in our analytical model. We can extend the analytical model results for the multiple class p-persistent MAC protocol to the contention window based MAC protocol

Dealing with Network Dynamics To estimate the number of active stations in each class, we online count the number of active stations of class i from the channel access history overheard in the past Hi successful transmissions. we set the value of Hi to the largest integer k such that the probability that the specific station finishes at least on successful transmission in the Hi virtual transmission times is larger than α , that is

Dealing with Network Dynamics

Dealing with Network Dynamics

Dealing with Network Dynamics two classes (10 nodes in each class) and greedy traffic the number of active nodes

Dealing with Network Dynamics two classes (10 nodes in each class) and greedy traffic p is adaptively determined p1 p2 the transmission probabilities

Dealing with Network Dynamics two classes (10 nodes in each class) and greedy traffic aggregated class 1 class 2 the throughputs

Dealing with Network Dynamics two classes and on-off traffic the transmission probabilities

Dealing with Network Dynamics two classes and on-off traffic under-utilized The failure to keeping the desired throughput differentiation during any time interval in the case of on-off traffic is, in part, due to the fact that the analytical result is derived under the uid model assumption, which does not hold in the case of bursty traffic. fully utilized the per flow throughputs of two classes

Analysis of AIFS Ratio of average per flow throughput is a function of: Transmission probability (CW) AIFS values Study through Simulation Set on class to a fixed AIFSN of 2, and vary the other from 2 – 8. Set the ratio of class 2 to class 1 as 2.0 Use analytical model to calculate p1 and p2

Analysis of AIFS Analysis of AIFS PMAC m=5, n=15,

Analysis of AIFS Conclusion: Too many dimensions of design freedom AIFS1=AIFS2= . . . =AIFSP=DIFS may be a better choice

Conclusion Derived an analytical model Through different transmission probabilities, service differentiation is achieved. Low design dimensions are better. Simulations show that targeted ratio’s are acquired and throughput is high.