A SCHEME FOR IMPROVED DELAY AND FRAME DROP PROBABILITIES IN e NETWORKS

Transcription

1 A SCHEME FOR IMPROVED DELAY AND FRAME DROP PROBABILITIES IN 82.11e NETWORKS Daniel Fokum and Cory Beard School of Computing and Engineering, University of Missouri Kansas City 534 Robert H. Flarsheim Hall 511 Rockhill Road, Kansas City, MO 6411 USA {dtf6wc, ABSTRACT This paper describes a scheme, AEDCF-, that allows stations implementing IEEE 82.11e to have varying maximum contention window sizes and persistence factors. The values of these parameters are adjusted using knowledge of the station s collision probability. Our research indicates that stations implementing this scheme experience much lower frame delays and frame dropping probabilities under heavy loads. In addition, this scheme has the benefit of not requiring many modifications within the station. KEY WORDS WLAN, 82.11e, EDCF 1. Introduction Wireless networks based on the IEEE standard are increasingly popular due to their ease of implementation and their interoperability with the Ethernet standard. In spite of their ease of implementation, networks have the following flaws: One of the medium access functions, the Point Coordination Function (PCF), is inefficient. The medium access functions are unable to provide different levels of service to different traffic classes. The medium access control functions use the wireless medium inefficiently. In light of these flaws, the IEEE s working group commissioned a task group, called Task Group E (TgE) to develop a new standard called 82.11e that will provide quality of service over 82.11a and 82.11b networks. The 82.11e standard will provide quality of service primarily by allowing for prioritization of different traffic streams within an network. Quality of Service (QoS) is a term used to define the different characteristics of a traffic flow. Some of these characteristics include frame delay, jitter (the amount of variation in frame delay), frame dropping probability, priority etc. In this document we will evaluate the effects of the different 82.11g parameters on the quality of service that is enjoyed by a traffic class. The rest of this document is laid out as follows. In Section 2 we present an overview of EDCF and its operation. In Section 3 we provide a description of the proposed AEDCF- scheme. In Section 4 we describe our simulation methodology and results. Concluding remarks are provided in Section Overview of EDCF and Its Operation A. EDCF Overview The standard divides time into two types of periods a contention-free period, and a contention period. Together both periods of time are called a superframe. During the contention-free period, stations are polled by the access point (acting as a hybrid coordinator). During the contention period all stations, including the hybrid coordinator, contend for access to the medium. As of December 9, 25, the IEEE had not yet published the final 82.11e standard. As a result the information about the standard was gleaned from several papers on 82.11e, as well as a document from the TgE website. The Enhanced Distributed Coordination Function (EDCF) is used during the contention period of a superframe. EDCF provides prioritization by allowing for different traffic classes called access categories (AC). Each AC has a separate queue within each station along with different contention window parameters. In addition, each access category has to sense the medium for a different interframe space time prior to beginning a transmission. This time value is defined as an Arbitration Interframe Space (AIFS). Each AIFS is equal to a DIFS time + k*aslottime, where k can potentially be zero and DIFS stands for DCF Interframe Space. It should be observed that the lower the value of k, the higher the AC s priority, because the AC would have to sense the medium for a shorter time interval. In addition to the new access functions described above, 82.11e replaces the dot11maxtransmitmsdulifetime parameter with a new

2 variable called dot11msdulifetime. This variable is defined per access category, and it refers to the amount of time that a frame from a given access category can remain queued up. Finally, [1] states that frame losses in 82.11e occur if the number of transmission attempts for a given frame exceeds either the short/long retry limit, or if the queuing delay for the frame exceeds the dot11msdulifetime limit for this traffic class. From [2] it is seen that the 82.11e MAC also has a new variable called TXOPlimit that defines the amount of time that an access category has access to the medium. Once an access category wins access to the medium, it may continue to transmit frames separated by SIFS without sensing the medium for AIFS[AC], as long as the station does not use the medium for a period that exceeds TXOPlimit[AC]. In addition it is seen that the eight priorities defined by the 82.1D standard are mapped to four access categories in 82.11e. Another enhancement of the 82.11e MAC is the fact that stations are now able to communicate directly with each other without necessarily involving the access point. This feature is known as the Direct Link Protocol of 82.11e. From [3] it is seen that if multiple queues within a station count down to zero simultaneously, a virtual collision occurs. If a virtual collision occurs, the transmission opportunity is handed to the AC with the highest priority, while the other AC s execute the backoff algorithm. When a virtual collision occurs, the transmission attempt counter for each frame in the lower priority queues is not incremented to indicate that a collision occurred. The new 82.11e standard also implements one new feature that should help it use the medium more effectively. This feature is called the group acknowledgement mechanism [4]. When this mechanism is used, a station may send several frames without waiting for individual acknowledgement of the frames. After the station has sent a burst of frames it will send out a group acknowledgement request (GroupAckReq). The receiver will respond by sending a GroupAck frame that will list information on all the frames that were correctly received. As with the basic ACK mechanism, both the GroupAckReq and GroupAck frames are separated by an SIFS interval. If the GroupAck frame shows that any frames were not received correctly, the sender will try to resend those frames as long as it can do so (Recall that frames are dropped if the frame retry limit for those frames is exceeded, or if the frames have been in the queue for more than dot11msdulifetime[i].) Figure 1 shows the timing relationships that exist between the interfame spaces under 82.11e. AIFS[j] Immediate access when AIFS[i] Medium is free >= DIFS/AIFS[i] DIFS/AIFS DIFS/AIFS Busy Medium Defer Access PIFS SIFS Contention Window Backoff-Window Slot time Next Frame Select Slot and Decrement Backoff as long as medium is idle Figure 1. Some IFS relationships B. EDCF Operation In the basic access mode a QoS station (QSTA) will have at least four access category queues that will contend for the medium independently. Each AC queue will independently sense that the medium is idle for the AIFS[i] time that is defined for that queue. Next, the AC queue will randomly pick a backoff interval between and the current size of the contention window, CW, and count down to zero. After one slot time (the slot time is a time value that depends on the physical layer being used) the medium is sensed again. If the medium is found to be idle the backoff counter, w, is decremented. If the medium is found to be busy, the backoff counter is frozen until the medium is found to be idle for AIFS[i] time. At this time the backoff counter is resumed. Once the backoff counter reaches zero, the frame is transmitted. If multiple AC queues count down to zero simultaneously, the transmission opportunity is granted to the AC with the highest priority, while the other colliding queues double their contention windows and attempt to sense the medium again. When the receiving station receives the transmitted frame, it verifies that the frame was received correctly, and that no other stations were transmitting at the same time the frame was received. If both of these conditions are true, the destination station sends an acknowledgement (ACK) frame to the sender after a period of time called a Short Interframe Space (SIFS). If the sending station does not receive an ACK frame within an ACKTimeout time, the sending station assumes that a collision took place on the medium. Following a collision, the sending station increases its contention window as follows: CW new = ((CW old +1)*2) 1, as long as CW new is less than or equal to the maximum contention window, CW max, defined for this PHY. In addition to the contention window doubling procedure, the station also increments its retry counter by one to indicate the number of transmission attempts for this frame. 3. Description of Scheme A. Description Our scheme is an extension of the Adaptive EDCF (AEDCF) scheme initially proposed by Romdhani et al. [5]. Unlike standard EDCF that always resets the contention window to CW min following the transmission of a frame, AEDCF resets the contention window based on a station s collision probability. Reference [5] defines the collision probability, f avg, as the ratio of the number of collisions experienced within a period, T update, to the number of frames sent within that same period. In order to minimize the effect of short-term changes in the collision probability, Romdhani et al. propose the use of an Exponentially Weighted Moving Average to smoothen the collision probability. In preliminary investigations, one of the authors observed that under high loads the highest priority traffic

3 class had many retransmission attempts per frame. Based on this observation, it was concluded that if each AC was allowed to pick a backoff counter from a larger interval the collision rate might be reduced. With this motivation, we set out to propose AEDCF- (Adaptive EDCF with Varying CW max and Persistence Factor). Another motivation for our research was to allow stations to select some of their MAC parameters independently. It was our opinion that since EDCF is a distributed function, stations should be able to compute some of their MAC parameters without relying on the hybrid coordinator to broadcast those values. Stations implementing the AEDCF- algorithm operated as shown below. Each station was to check its average collision probability, f avg (j), prior to executing the exponential backoff algorithm. If f avg (j) was greater than some threshold value, γ, the station was allowed to increase CW max [i] temporarily. If f avg (j) was greater than another threshold value, β. the persistence factor the factor, PF, used to increase the contention window was set to 4, otherwise it was set to 2. Our algorithm operated as shown below: f avg (j) = (1-α) * f curr (j) + α * f avg (j-1) MF[i] = min ((1 + (i*2))* f avg,.8) CW new [i] = max (CW min [i], CW old [i] *MF[i]) If (f avg >β) then PF[i] = 4 Else PF[i]= 2 If (a collision occurred) If (f avg > γ) then If ((CW[i]+1) * PF [i]) <= (2 * (CW max [i] + 1)) then CW new [i] = (CW[i]+1) * PF [i] Else If ((CW[i]+1) * PF [i]) <= CW max [i]) then CW new [i] = (CW[i]+1) * PF [i] B. Impact of Changes Adaptive EDCF with varying CW and PF can be implemented at each station with minimal impact. In addition to the four additional registers mentioned in [5] to store the values for α, PF[i], f avg (j-1) and T update, our scheme will require two additional registers to store the values of β and γ. Computing the new persistence factor will take one comparison operation. In the event of a collision, determining whether or not to double CW max will take two comparisons. Choosing to double CW max will take one addition and one multiplication operation. The other multiplication and addition operations are a normal part of the standard EDCF procedure; therefore, they do not add to the complexity of the function. 4. Simulation Results A. Simulation Parameters and Methodology The algorithm described above was simulated using CSIM programs. In the simulations the number of competing stations was held at 2, 25 or 3 stations, while varying the mean frame interarrival time from 1 msec to 1 msec in descending steps of 2 msec. In each of our simulation exercises, at least 25% of the competing stations were marked as having high priority traffic i.e. having traffic from class, at least 5% of the stations were marked as having medium priority traffic, and the rest of the stations were marked as having low priority traffic i.e. traffic from class 2. All the stations were assumed to be within range of each other; therefore, we did not have to account for the hidden terminal problem. Each simulation was run for 3, frame arrivals, following a warm-up period in which 12,5 frames were processed. Our frame sizes were uniformly distributed between 1 and 495 octets, with no fragmentation allowed. In addition the frames were entered into queues of infinite lengths at each station. The simulations used the 82.11g MAC, and it was assumed that all control and data frames were sent over ideal channels at the maximum link rate of 54 Mbps. The simulation parameters shown in table 1 were used in all our simulations, many of which are taken directly from the 82.11g standard. All the parameters marked with an asterisk (*) were drawn from the standard [6]; those with two asterisks came from [6,7] and the HCF group s recommendation [3]. Except for the contention window sizes, the rest of the variables came from [1]. The contention window sizes were chosen such that the highest priority access category had priority over DCF. In addition the contention window sizes were chosen such that each access category had at least two opportunities to back-off prior to attaining its maximum contention window size. Table 1. MAC parameters used in simulations MAC parameter Value CW min [] 15 CW max [] 63 CW min [1] 31 CW max [1] 127 CW min [2] 63 CW max [2] 123 aslottime* 2 µs asifstime* 1 µs adifstime* 5 µs aaifstime[]** 5 µs aaifstime[1]** 7 µs aaifstime[2]** 9 µs ampdumaxlength* 495 octets dot11longretrylimit* 7 dot11msdulifetime[] 6 ms dot11msdulifetime[1] 1 ms dot11msdulifetime[2] 2 ms MAC Header Length* 34 octets PHY Header Length* 24 octets ACK Length* 14 octets

4 In addition, in our simulations we set β = 2*γ, where γ = This value of γ was approximately half of the maximum average weighted collision probability seen in our simulations. B. Discussion Our simulations indicated that the new scheme outperformed both basic EDCF and adaptive EDCF at high loads, as shown in figures 2 and 3. The results also indicate that the performance of AEDCF- improves as the number of stations increases. Observe from figure 2 that the system response time for EDCF levels off at about.32 s. This is a result of EDCF s tendency to drop frames more frequently. In our simulations the system response time was computed only for those frames that were successfully transmitted. As load increases EDCF drops more and more frames, and so there are fewer frames available for the system response time computation. Recall that AEDCF- doubles the CW max, then doubles the persistence factor next. Both of these changes result in the contention window being chosen from a larger interval. As a result, stations have a much lower probability of colliding under AEDCF- than under EDCF. At higher loads, therefore, this gain in performance becomes more apparent. In addition to the new scheme s excellent response time performance under heavy loads, AEDCF- also drops frames at a lower rate than EDCF, as illustrated in figure 4. From this figure we see that at the highest load (.1 ms between frame arrivals) 73% of the class frames are dropped, whereas 99.9% of the class frames are dropped for basic EDCF. Our research [8] indicates that this value was 91.5% for AEDCF. Based on this, we can conclude that the highest priority traffic gains most from the new scheme. It is worth noting, that the other traffic classes also gain from lower frame drop probabilities at the very high loads. AEDCF- does have one flaw it introduces higher jitter values for the traffic under heavy loads. In our simulations, we measured jitter by collecting the variance of the system response time. Our data indicates that AEDCF- starts out having comparable or lower jitter values than both EDCF and AEDCF. Once each of these medium access functions attains its maximum throughput, AEDCF- begins to display poor jitter performance. This degradation in performance is a result of the lower tendency of AEDCF- to drop frames. Since more frames get transmitted, these frames then adversely affect the variance of the system response time. In spite of this flaw, we are of the opinion that AEDCF- is superior to both EDCF and AEDCF since it has much better response times, and lower frame dropping probabilities. Response time/s System Response Time vs Interarrival time for basic EDCF and AEDCF with varying PF and CW max STA EDCF 25 STA EDCF 3 STA EDCF 2 STA AEDCF- 3 STA AEDCF- Figure 2. System Response Time for EDCF vs. AEDCF- Response time/s System Response Time vs Interarrival time for AEDCF and AEDCF with varying PF and CW max STA AEDCF 25 STA AEDCF 3 STA AEDCF 2 STA AEDCF- 3 STA AEDCF- Figure 3. System Response Time for AEDCF vs. AEDCF- Frame drop probability Frame dropping probability vs Interarrival time for EDCF and AEDCF-, 3 stations AEDCF- Class AEDCF- Class 1 AEDCF- Class 2 EDCF Class EDCF Class 1 EDCF Class 2 Figure 4. Frame Dropping Probabilities for EDCF vs. AEDCF-

5 System Response Time variance vs Interarrival time for basic EDCF and AEDCF with varying PF and CW max References Response time variance/s STA EDCF 25 STA EDCF 3 STA EDCF 2 STA AEDCF- 3 STA AEDCF- Figure 5. System Response Time Variance for EDCF vs. AEDCF- 5. Conclusion The new 82.11e standard will allow for prioritization in networks; however, this prioritization comes at the cost of higher collision probabilities. In this paper we presented an algorithm called AEDCF-. This algorithm can help a station deal with temporary instability in the WLAN caused by saturation, plus it has the benefit of being fully distributed. As a result QSTAs can compute values for CW min [i] and AIFS[i] without having to rely on a HC to broadcast these values. In light of the advantages offered by AEDCF-, it is our recommendation that any attempts to offer quality of service in 82.11e LANs focus on CW min, and CW max. Our research [8] indicates that the frame retry limits do not influence frame delay significantly. Our work in this paper may be extended to investigate the effects of starting the AEDCF algorithm with different values for CW min and CW max. In addition, AEDCF s performance needs to be investigated when the group ACK mechanism is used. The AEDCF algorithm may be extended to adapt the packet burst length based on the number of acknowledgements that are received [5]. Finally, some research needs to be done in making the 82.11e MAC fairer. By incorporating the use of different sizes for CW min, CW max and AIFS the 82.11e MAC has become less fair than DCF. One approach to alleviating the unfairness of the 82.11e MAC might be to allow stations to use a given TXOP to transmit data from the queue that won access to the medium as well as another queue other than the queue that originally won access to the medium. The second frame that is transmitted may be chosen by choosing a frame from a queue that has the largest queue length to AIFS length ratio. [1] A. Grilo, M. Macedo, & M. Nunes. A Scheduling Algorithm for QoS Support in IEEE82.11e Networks, IEEE Wireless Communications, 1(3), 23, [2] S. Mangold, S. Choi, G.R. Hiertz, O. Klein, & B. Walke. Analysis of IEEE 82.11e for QoS Support in Wireless LANs, IEEE Wireless Communications, 1(6), 23, 4-5. [3] M. Wentink, S. Choi, & M. Hoeben. HCF Ad Hoc Group Recommendation Normative Text to EDCF Access Category, Mar. 22; s/documentholder/2-241.zip [4] Y. Xiao, IEEE 82.11e: QoS Provisioning at the MAC Layer, IEEE Wireless Communications Magazine 11(3), 24, [5] L. Romdhani, Q. Ni, & T. Turletti Adaptive EDCF: enhanced service differentiation for IEEE wireless ad-hoc networks, Proc. Wireless Communications and Networking Conference, 23, (WCNC 23), [6] IEEE Std , Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, IEEE, [7] IEEE Std g, Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, Amendment 4: Further Higher Data Rate Extension in the 2.4 GHz Band, IEEE, 23. [8] D. Fokum, The Effects of MAC Layer Parameters on Quality of Service Provided by 82.11e, master s thesis, Dept. Computer Science and Electrical Eng., University of Missouri Kansas City, 25.