- Email: [email protected]
A simple recursive scheme for adjusting the contention window size in IEEE 802.11e wireless ad hoc networks
Computer Communications 29 (2006) 3789–3803 www.elsevier.com/locate/comcom
A simple recursive scheme for adjusting the contention window size in IEEE 802.11e wireless ad hoc networks Hassan Artail, Haidar Safa *, Joe Naoum-Sawaya, Bissan Ghaddar, Sami Khawam American University of Beirut, P.O. Box: 11-0236, Riad El-Solh, Beirut 1107 2020, Lebanon Received 20 February 2006; received in revised form 16 June 2006; accepted 16 June 2006 Available online 14 July 2006
Abstract The IEEE 802.11e standard has been introduced recently for providing Quality of Service (QoS) capabilities in the emerging wireless local area networks. This standard introduces a contention window based Enhanced Distributed Channel Access (EDCA) technique that provides a prioritized traffic to guarantee the minimum bandwidth needed for time critical applications. However, the EDCA technique resets the Contention Window (CW) of the mobile station statically after each successful transmission. This static behavior does not adapt to the network state since it reduces the network usage and results in bad performance and poor link utilization whenever the demand for link utilization increases. This paper proposes a new adaptive differentiation technique for IEEE 802.11e Wireless Local Area Networks that takes into account the network state before resetting the contention window. In the new technique, the congestion level of the network is sensed by using previous CW values. Three other enhancement techniques that focus on network adaptation are also discussed. Their main limitations are the high complexity of the implemented algorithms and their slow adaptation to the network state when the channel experiences bursty traffic. The proposed technique is compared to the original differentiation techniques of IEEE 802.11a and IEEE 802.11e standards, as well as to the enhancement schemes. Results show that the proposed adaptive technique outperforms IEEE 802.11e and is comparable to the other enhancement schemes while maintaining relatively low complexity requirements. 2006 Elsevier B.V. All rights reserved. Keywords: Enhanced Distributed Channel Access; IEEE 802.11e; Quality of Service; Wireless LAN; MANET
1. Introduction The emerging IEEE 802.11 family of wireless technologies has shown tremendous growth and acceptance as a solution for Wireless Local Area Networks (WLANs). Simultaneously with the growth of WLANs, there is the trend for multimedia applications over IP with Quality of Service (QoS) support. The use of such applications is now a reality in corporate networks and promises to expand to the global Internet. The original IEEE 802.11 networks are best effort networks and do not support QoS for time critical applications [5]. A new IEEE 802.11 Standard (IEEE 802.11e) has been introduced to replace the best effort services by more sophisticated services that
*
Corresponding author. Tel.: +961 1 350000X4270. E-mail address: [email protected] (H. Safa).
0140-3664/$ - see front matter 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.comcom.2006.06.016
guarantee QoS attributes such as bandwidth, delay, and jitter [6]. This standard focuses on replacing the conventional Distributed Coordination Function (DCF) and the optional Point Coordination Function (PCF) of the Medium Access Control (MAC) layer by a Hybrid Coordination Function (HCF). The HCF defines two medium access mechanisms: a contention-based channel access referred to as Enhanced Distributed Channel Access (EDCA), and controlled channel access referred to as HCF Controlled Channel Access (HCCA). In this work we focus on the EDCA only. The EDCA is a Contention Window (CW) based channel access function and is based on priority allocation for different categories of traffic to guarantee the minimum bandwidth needed for high priority applications, thus increasing the network performance. Similarly to the traditional DCF technique, the EDCA technique of IEEE 802.11e employs a static reset to the CW. This static behavior is a drawback since it does not offer any
3790
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
room for adaptation to the network state [6]. It also reduces the network usage and is reflected in bad performance and poor QoS whenever the demand for medium utilization increases [7]. Several techniques have been proposed to enhance the performance of IEEE 802.11e by adapting CW to the network state [1,2,4,7,9,14,15]. Even though these techniques, which are presented in Section 2, outperform the EDCF technique used in IEEE 802.11e, their main limitations are the high complexity of their algorithms and their slow adaptation to the network state as explained later in Sections 2 and 3. Low complexity algorithms are favorable due to their simpler implementation and low demand on the system in terms of resource consumption. This paper introduces a new low complexity Adaptive EDCA algorithm that adapts the CW to channel conditions and adjusts it depending on the network utilization and performance. The proposed technique outperforms IEEE 802.11e and is comparable to the other enhancement schemes while maintaining relatively low complexity requirements and providing a faster adaptation to the network state. The rest of the paper is organized as follows. A brief overview of the DCF and EDCA of the IEEE 802.11 and 802.11e standards is first presented in the remainder of this section. In Section 2, we discuss the adaptive techniques that appeared recently in the literature, and then introduce the proposed scheme in Section 3. Section 4 introduces and evaluates the different techniques in terms of complexity and performance, and then finish in Section 5 with conclusions regarding our findings. 1.1. Overview of the MAC layer A review of the DCF and EDCA techniques of the IEEE 802.11 and 802.11e MAC layer is provided in this section. 1.1.1. IEEE 802.11 MAC layer The DCF technique of the IEEE 802.11 MAC layer employs a contention-based channel access function and uses carrier sensing multiple access with collision avoidance (CSMA/CA) mechanism [5]. The mechanism tries to avoid the event of two or more stations attempting to transmit a packet simultaneously, in which case an unsuccessful transmission occurs. Under DCF a station that intends to transmit monitors the channel until it is idle, and then waits for a time period of a Distributed Inter-Frame Space (DIFS) length. After sensing an idle channel for a DIFS period, the mobile station randomly selects a backoff timer (time slot) within a backoff window. The backoff timer is decreased only when the medium is idle; it is frozen when another station is transmitting (see Fig. 1). Each time the medium becomes idle, the station waits for a DIFS and continuously decrements the backoff timer. As soon as the backoff timer expires, the station is authorized to access the medium and transmit. The backoff timer is derived from a uniform distribution over the interval [0, CW 1], where CW is a value between [CWmin, CWmax].
Medium is idle
Medium is busy
Random backoff timer, value ∈ [0, CW], CW ∈[CWmin, CWmax] Transmission Time DIFS Medium is busy
Slot Time
Medium is idle Station enters backoff
Transmission by another station
Station resumes backoff Transmission Time
DIFS
DIFS
Fig. 1. IEEE 802.11 DCF Channel Access.
At the very first transmission attempt, CW value is equal to the initial backoff window size CWmin. For every unsuccessful transmission the value of CW is doubled until CWmax is reached. After transmitting a frame, the station expects to receive an acknowledgment (ACK) from the destination station following a Short Inter-Frame Space (SIFS) time. If the acknowledgment is not received, the sender assumes that the transmitted frame was collided, so it schedules a retransmission and enters the backoff process again. After every successful transmission the CW is reset to CWmin. DCF employs a discrete time backoff scale. The time that immediately follows the DIFS is slotted and the station is permitted to transmit only at the beginning of each slot time. The length of the slot time is set equal to the time needed at any station to detect the transmission of a packet from any other station. In all, DCF can only support best-effort services, not any QoS guarantees [9,12]. Typically, time-bounded services such as voice over IP, or audio/video conferencing require specified bandwidth, delay and jitter, but can tolerate some losses. However, in DCF mode, all the stations compete for the resources and channel with the same priorities. There is no differentiation mechanism to guarantee bandwidth, packet delay and jitter for high-priority traffic or multimedia flows [7]. All the MAC parameters including SIFS, DIFS, slot time, CWmin, and CWmax are given in Table 1 for the Direct Sequence Spread Spectrum (DSSS) physical channel (PHY). However, the MAC evaluation in the following sections is valid irrespective of the underlying PHY layer. 1.1.2. IEEE 802.11e MAC layer The new IEEE 802.11e standard introduces a Hybrid Coordination Function (HCF) [6], which defines two medium access mechanisms: a contention-based channel access referred to as Enhanced Distributed Channel Access (EDCA), and controlled channel access referred to as HCF Controlled Channel Access (HCCA). In this work
H. Artail et al. / Computer Communications 29 (2006) 3789–3803 Table 1 MAC Parameters for IEEE 802.11 Parameters
Value
SIFS DIFS CWmin CWmax Slot time
10 ls 50 ls 32 1024 20 ls
we focus on the EDCA only, in which QoS is realized with the introduction of Access Categories (ACs) and multiple independent backoff entities. Frames are delivered by parallel backoff entities within one 802.11e station. EDCA provides differentiated distributed channel access for eight traffic priorities that are mapped into four queues (ACs) as shown in Table 2. Thus, four backoff entities exist in every 802.11e station. Each AC queue works as an independent DCF station and uses its own contention parameters such as Arbitration Inter-Frame Space (AIFS), CWmin, and CWmax. Similar to a DCF station, each AC starts a backoff timer after detecting an idle channel for a time interval equal to an AFIS length. The backoff value is chosen to be a random number between [1, CW + 1], where initially CW is set to CWmin and increases whenever collision occurs up to CWmax. CW increases in accordance with the following equation [6]: CW new ½AC ¼ 2 CW current ½AC þ 1
ð1Þ
In case of successful transmission, the CW value of the AC queue is reset to CWmin. Additionally, in EDCA high priority traffic has a smaller AIFS, CWmax, and CWmin than low priority traffic as shown in Table 3. Thus, higher priority traffic will enter the contention period and access the wireless medium earlier than lower priority traffic. For example if i has lower priority than j, it follows that:
Table 2 User traffic priorities mapped to access categories User priority
Designation
Access category
1 2 0 3 4 5 6 7
BK (Background) BK (Background) BE (Best-effort) EE (Video/Excellent-effort) CL (Video/Controlled Load) VI (Video) VO (Voice) NC (Network Control)
AC_BK AC_BK AC_BE AC_BE AC_VI AC_VI AC_VO AC_VO
Table 3 Access categories Access category
AC_VO
AC_VI
AC_BE
AC_BK
AIFS CWmin CWmax TXOPlimit
2 7 15 0.003008
2 15 31 0.00601
3 31 1023 0
7 31 1023 0
CW min ½i > CW min ½j CW max ½i > CW max ½j
3791
ð2Þ
AIFS½i > AIFS½j In IEEE 802.11e, whenever a station seizes the channel, it will be able to transmit one or more frames for a Transmission Opportunity time interval (TXOP). A TXOP is an interval of time during which a given station has the right to transmit frames. A TXOP is defined by its starting time and duration. The duration of TXOP is limited by a parameter referred to as TXOPlimit (for more details about EDCA, refer to [6]). The main problem of this EDCA ad-hoc mode is that the values of CWmin, CWmax and backoff function of each queue are static and do not take into account wireless channel conditions. 2. Related work In [14], a new scheme is proposed for the purpose of improving the performance of IEEE 802.11e under different load rates and increasing the service differentiation in EDCF-based networks. The scheme uses a dynamic procedure to change the contention window value after a collision or a successful transmission. While in EDCF CWmin[i] and CWmax[i] values are statically set for each access category, this particular scheme, which we refer to as collision rate-based adaptive EDCF (CR-AEDCF) resets its CWs to adaptive values that are different from the CWmins, taking into account their current sizes and the average collision rate. This is driven by the assumption that when a collision occurs, another one is expected in the near future, and so to avoid bursty collisions, this scheme updates the contention window with values that are close to the previous ones but slowly decrease toward CWmins after successful transmissions. Furthermore, the scheme suggests changing the mechanism of doubling the CWs when a collision occurs and uses different factors that correspond to the different classes to increase the CWs. More specifically, each class updates its CW taking into account the estimated collision rate, which is calculated as a smoothed version of the ratio of the number of collisions to the total number of packets sent during a fixed time. Next, a multiplicative factor is computed using the collision rate such that it maintains the priority relationship between different classes and is upper-bounded by 0.8. This factor is then multiplied by the previous value of CW to become the new value while making sure that it does not fall below CWmin. In all, it is expected that on average, CR-AEDCF improves the performance of all types of traffics and outperforms the EDCF technique used in IEEE 802.11e by adaptively varying the contention window. A simple Static Slow Decrease (SSD) scheme was evaluated in [1]. The main purpose of this scheme is to slowly decrease CW values by employing a static factor (with a value of 0.5). It is obvious that such an approach is not
3792
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
suitable for all network conditions. This scheme is described in Eq. (3) below.
3. Proposed scheme
CW new ¼ 0:5 ðCW current CW min Þ þ CW min
The static reset method employed in IEEE 802.11 and 802.11e has been proved to be ineffective in maximizing channel utilization whenever the demand for channel access (e.g. traffic load) increases [8,11,14]. The probability of collisions in a busy wireless channel is high and will likely cause CW to approach CWmax almost every time before the station succeeds in transmitting its data. After a successful transmission, and the next time the station wants to transmit a frame during similar channel conditions, CW will start at CWmin and will subsequently increase (double) at each unsuccessful step until the station succeeds in transmitting its packet. As a result, a station may have to try several times before it succeeds and hence, significant time is wasted due to resetting CW to CWmin and consequently, the channel becomes underused. In low channel demand, the value of CW tends to acquire low values (close to CWmin and distant from CWmax) due to the low probability of collisions. In these conditions, high values of CW will lead to unnecessary waiting while the channel is idle. Consequently, it is favorable to adopt an adaptation technique that estimates the channel contention level and reflects the estimation in the value of CW. With this, we share the same objective as that of the CR-AEDCF but with the additional goal that the computational complexity of the algorithm is low. In high collision probabilities, the value of CW is driven toward higher values that are accustomed to the channel state, therefore getting values distant from CWmin and close to CWmax. To start with, our proposed scheme, which is a extension to our work in [10], needs to acquire channel state information in order to adjust the value of CW. Rather than measuring the collision rate directly, as in the CRAEDCF scheme, we use an indirect method that will be proven to be equally effective, through the recursive computation of the window size for CW, hence the name Simple Recursive Adaptive Enhanced DCF (SRAEDCF). This is justified by the fact that the value of CW directly reflects the state of the channel, as was detailed above. At the risk of repetition, we mention the following:
ð3Þ
In [8], an approach was proposed to replace the EDCA technique and is based on adapting the AIFS in response to network conditions, hence its name: Adaptive MAC Parameters (AMPA). The rationale behind adapting AIFS is to reduce the waiting time for high priority applications and increasing it for low priorities. When there is congestion in the network, high priority AIFS is decreased while best effort AIFS is increased. Conversely, if the network is in normal conditions, high priority AIFS is increased while best effort AIFS is decreased. It follows that AMPA favors high priority streams when the network is overloaded then and tries to serve all traffic when the network is in normal conditions. As in CR-AEDCF, network congestion level is first estimated using the ratio of the number of collisions to the number of sent packets. This value is then compared to threshold values above which the network is defined as congested leading to a decrease in AIFS. Conversely, below the threshold, the network is assumed in normal conditions and AIFS is increased for high priorities. Threshold values should be adapted to traffic priorities hence having higher thresholds for lower priorities. Hence, such a differentiation favors higher priority traffic to gain access to the channel. Threshold values (THX) are shown in Table 4. In general, AMPA respects particular QoS metrics, by providing bounded delays, and better throughput. This scheme achieves a stable performance regardless of the traffic load by using dynamic adaptation of AIFS depending on the network status. Recently, a link adaptation strategy that provides differentiation not only at the MAC layer, but also at the PHY layer was proposed in [2]. This strategy exploits the positive ACK procedure to evaluate the quality of the link. If the transmitter does not receive ACK frame it concludes that the last transmission was failed. For each active link, a transmitter maintains two counters, a success counter and a failure counter. If a frame is successfully transmitted, then the success counter is incremented by one and the failure counter is reset to 0. If the transmission fails then the failure counter is incremented by one and the success counter is reset to 0. These two counters are used to determine whether the quality of the link is good or not. The transmission rate is adjusted with respect to the link quality. Another recent approach that combines the slow decrease and the dynamic adaptation of CW was proposed in [4]. Table 4 Threshold differentiation Access category
AIFS
THX
AC_BK AC_BE AC_VI AC_VO
7 3 2 2
0.20 0.15 0.10 0.05
1. The value of CW varies in the range [CWmin, CWmax]. 2. When the channel is congested, CW will acquire values close to CWmax and distant from CWmin. 3. When the channel is free,CW will acquire values close to CWmin and distant from CWmax. Hence, it is feasible to estimate the channel congestion level by taking into consideration the current distance of CW from the limits CWmin and CWmax. That is, by dividing the distance (CWcurr CWmin) by the maximum distance (CWmax CWmin), we get an indication for the channel congestion level. It follows that the estimated congestion level can be written as:
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
CW curr CW min CW max CW min
ð4Þ
In the above, CF is a certainty factor (CF) that is included in the ratio in order to relate to the time interval between subsequent channel estimations, i.e., successful transmissions. In particular, the certainty of the estimate should be high whenever the time difference between subsequent estimates is small and should be low whenever the time difference is large. Actually, the value of CF should accomplish two functions: 1. Provide a high weight (between 0 and 1) when the difference in time between consecutive successful transmissions is in the order of milliseconds and a low weight when the difference is in the order of seconds (or minutes). 2. Maintain the differentiation between the four classes of access categories. For instance, video and audio streaming applications should correspond to higher weights relative to applications with no real-time requirements. To satisfy the first objective, a time decaying function such as the one in Eq. (5) may be used. This function, which is suitable for the AC_VI access category, provides an upper limit of 0.7 when the time difference is small and a lower limit of 0.4 when the difference is large. CF ¼ 0:3 expð0:001t2 Þ þ 0:4
ð5Þ
The 0.4 and 0.7 limits were chosen rather than the 0 and 1 as limits for the certainty factor. The main reason for this choice is that factors other than busy channel (e.g., buffer overflow) may introduce packet dropping. Thus limiting the certainty factor to a smaller length will allow the algorithm to account for these factors. However, rather than keeping track of the times of when successful transmission took place and performing extra computations at the MAC layer level, we suggest relying on application differentiation for weight variation. This is mainly due to the fact that the time between subsequent transmissions is directly dependent on application type. For instance, video and audio streaming applications (priorities 0 and 1) require fast frame transmission while other applications such as http and email services (priorities 2 and 3) transmit smaller number of frames in longer periods. This dependence on application time permits the assignment of higher weights for streaming applications and lower weights to other applications. Several simulations were run to assign weights on access categories and led to the values that are shown in Table 5. Following channel estimation, the next step is to figure out the value of CW that best adapts to this estimation. The proposed approach is based on estimating the distance from CWmin at which the probability of a successful transmission is maximized. Since CW has a minimum value of CWmin and a maximum value of CWmax, the maximum distance that CW can acquire is CWmax CWmin. To model a
Table 5 Values of CF Priority
CF
AC_VO AC_VI AC_BE AC_BK
0.8 0.7 0.4 0.3
slow decrease after each successful transmission, a wireless node should attempt transmission at a CW that is less than CWcurr, at which the node was able to transmit in the previous attempt. This slow decrease model limits the maximum distance to CWcurr CWmin. Since the ratio previously introduced in Eq. (4) acquires a value in the interval [0, CF], which increases with the congestion level, a good model for the estimated distance is: d ¼ ratio ðCW curr CW min Þ
ð6Þ
By making CWnew = CWmin + d and substituting (4) in (6), we get: CW new ¼ CW min þ ratio ðCW curr CW min Þ ¼ CW min þ CF
ðCW curr CW min Þ CW max CW min
2
ð7Þ
The ratio in (4) is a normalized value ranging between 0 and 1 · CF that reflects the weighted degree of channel contention, taking into account the four priority classes. This ratio would render a value close to 0 whenever the channel is free and a value close to 1 · CF whenever the channel is congested. Hence, multiplying this ratio by the factor (CWcurr CWmin) and adding the result to CWmin would result in a value bounded by [CWmin, CF · CWcurr]. This value of CWnew is thus a good representation of the backoff timer value needed for transmission for the given traffic priority, while taking into account the current network conditions. Given that our proposed scheme represents a simpler approach to CR-AEDCF, we compare the adaptability of CW to the variations in the channel conditions for both schemes. For this, we calculate and plot CWcurr versus the changes in channel congestion levels. In Fig. 2, we simulate Channel Condition
CR-AEDCF
SR-AEDCF
16 14 12 CW Value
ratio ¼ CF
3793
10 8 6 4 2 0 1
6
11 16 21
26 31 36
41 46 51 56 61 Time
66 71
76 81 86
91
Fig. 2. Value of CW as a function of channel conditions for the CRAEDCF scheme and our proposed scheme.
3794
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
a scenario in which a station is transmitting video streams during various channel conditions as depicted in the bottom curve, which indicates the occurrence of a collision or the lack of it. In this type of transmission, the class priority is 0 and hence, CWmin = 7 and CWmax = 15. Referring to the figure, between the time values 1 and 8, the channel is free and the wireless node is transmitting at CWmin. As the congestion level is increased, the CW value is adapted to the congestion level as seen in the linear increase in both schemes. Between the times values 15 and 20, the channel experiences bursty variations, thus prompting responses in the value of CW: The CR-AEDCF scheme decreases CW to a value close to CWmin while our proposed scheme offers a more conservative adaptation, as reflected in the comparably higher values of CW. This conservative slow variation offers better performance in bursty channel conditions since the channel is free momentarily and goes back to the busy state. Furthermore and as noticed when the channel gets more congested at time = 21, our scheme offers a faster adaptation, which brings the value of CW to the maximum value faster than CR-AEDCF. This conservative behavior, which suits bursty conditions, does not negatively impact performance when the channel is available for a period of time. When the channel is mostly idle (from time 37 to time 45), both schemes decrease CW steeply toward CWmin, although our proposed scheme is minimally lagging behind. A similar behavior can be noticed between the time values 55 and 63. 3.1. Complexity evaluation Complexity is a major concern for real time implementations. Low complexity algorithms are favorable due to their simpler implementation and low demand on the system in terms of resource consumption. In this section, we evaluate the complexity of our proposed approach and compare it to that of the other approaches discussed earlier. 3.1.1. Proposed scheme (SR-AEDCF) The ratio evaluation shown in Eq. (4) requires 2 subtractions, 1 division, and 1 multiplication. The evaluation of the new value of CW requires 1 multiplication and 1 addition. In total, this scheme requires 1 addition, 2 subtractions, 2 multiplications and 1 division. Finally in terms of storage, this scheme only requires one register for storing the value of CWcurr. 3.1.2. CR-AEDCF scheme The exact computational requirements can be retrieved from [14]: to compute the collision rate, 2 additions and one division are required. To do the smoothing, 1 subtraction, 1 addition, and 2 multiplications are additionally needed. One addition, 2 multiplications, and one comparison are needed to compute the multiplication factor. Finally, the new CW value is calculated by using 1 multiplication
and 1 comparison. As for storage, 6 registers are needed for storing the number of collisions, sent packets, and other variables. In total, 4 additions, 1 subtraction, 5 multiplications, 2 comparisons, and 6 storage registers are needed. 3.1.3. SSD scheme This scheme sets CW to half of its previous value. As a result, it evaluates CW at each state by employing 1 addition, 1 subtraction, and 1 multiplication. 3.1.4. AMPA scheme This scheme requires 2 additions and 1 division to compute the collision ratio. Further, 1 addition and two registers are needed to compute the new value of AIFS (involves two comparisons). One division and one register are also required to compute the value of the threshold (THX). In all, this scheme requires 3 additions, 2 divisions, 2 comparisons, and 3 memory registers for data storage. The summary of the above evaluation is shown in Fig. 3. Of particular importance is the comparison to the CRAEDCF scheme, where it is obvious the much less complexity that our proposed scheme offers. 4. Performance evaluation SR-AEDCF and CR-AEDCF provide distinct mechanisms to quantify network conditions. As deeply discussed in the previous sections, the advantage in the SR-AEDCF relies only on the current CW value to accurately evaluate channel usage and decide on the next CW. As opposed to SR-AEDCF, CR-AEDCF relies on statistics by keeping record of the number of all packet drops and the number of all transmitted packets. The disadvantage of the latter approach is the tardiness of the algorithm to converge to the current channel conditions. For instance, packets dropped way in the past are relevant for the current decision. This contradicts with the nature of wireless systems that is known to tumble fast due to many factors. As shown in Fig. 2, SR-AEDCF proved high competence in fast adaptation to network conditions. Some modifications might enhance the performance of CR-AEDCF, these might include the use of a moving average rather than a
Fig. 3. Computational complexity comparison.
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
simple average for estimating channel usage. However as shown in Section 3.1, SR-AEDCF outperforms CR-AEDCF in terms of complexity, hence introducing modifications such as moving average to CR-AEDCF will certainly complicate it more. In the following section we investigate the performance of the different schemes and in particular SR-AEDCF and CR-AEDCF through simulations. The performed simulations reveal the performance of the discussed schemes in terms of the following three parameters: bit rate, end-to-end packet delay, and packet drop rate. The bit rate measures the effective bit rate achieved by each node (e.g., the number of bits transmitted per second. End to end packet delay is the time a packet takes to reach the destination (e.g., source o destination delay). Packet drop rate is the number of packets dropped before reaching their destination. These parameters provide direct figures of the supported QoS. The simulation topology of this scenario is simple. It consists of 8 mobile nodes: 4 source nodes and 4 destination nodes. The four source nodes are labeled 0 through 3 and each node is transmitting with a different priority. Node 0 has a higher priority than Node 1, which has a higher priority than Node 2, which has a higher priority than Node 3. Each source is a constant bit rate source over UDP (User Datagram Protocol) with a transmission rate equal to 600 Kbps. The size of a transmitted packet is 512 bytes and it was assumed that all the nodes are in transmission range. The bit rate performance that is reported below was measured at the destination nodes and thus, reflects what the source nodes were able to transmit. Finally, we mention that the transmission channel was modeled using the Two-Ray model and show the simulation topology in Fig. 4. To simulate the proposed scheme, we used version 2.28 of the network simulator NS-2 [13], which implements many data link, network, and transport protocols and can be used to simulate large networks. An important feature of ns2 is the wireless model, which simplifies the simulation process of the model.
Node 4 Node 5
Node 0 Node 1
Node 6 Node 7
Node 2 Node 3 Fig. 4. Simulation Topology.
3795
4.1. Bit rate performance We start with the comparison of the bit rate performance for the five schemes: IEEE 802.11e standard, Collision Rate-based Adaptive EDCF (CR-AEDCF), Static Slow Decrease (SSD), Adaptive MAC Parameters (AMPA), and our own Simple Recursive Adaptive EDCF (SR-AEDCF). Fig. 5 presents the comparison of the four access categories (see Table 2) in four corresponding graphs. Further and to make the curves traceable, the results were smoothed using a moving window average with a window size equal to 10 samples. As can be implied from Fig. 5, Node 0 starts transmitting at time T = 0 s (Fig. 5 (upper-left)), Node 1 starts T = 10 s (Fig. 5 (upper-right)), Node 2 at T = 20 s (Fig. 5 (lower-left)), and Node 3 at T = 30 s (Fig. 5 (lower-right)). During the initial 10 s, Node 0 is the only one transmitting and consequently is able to occupy the entire available bandwidth. At T = 10 s, Node 1 starts sharing the channel resources with Node 0, which explains the relative drop in the throughput of the latter. Additionally, the bit rate plots show heavy oscillations as the number of transmitting nodes increases, although much of the oscillations were dampened as a result of the applied smoothing. When comparing the schemes, especially our proposed scheme to CR-AEDCF, the highest priority node was able to perform equally well, on average, under both schemes. The other three schemes, including IEEE 802.11e, exhibit much poorer performance as the number of nodes increases. Another notable observation is the approximate steady performance of both schemes, which indicates the effectiveness of the priority differentiation mechanism under the CR-AEDCF and SR-AEDCF schemes. Starting with the AC_VI access category (Node 1), SRAEDCF outperforms CR-AEDCF throughout the simulation. The difference in performance is more prevalent with Nodes 2 and 3, which is significantly important since it does not deprive lower priority applications from being able to access the wireless network. Referring back to Fig. 2, it is obvious that the fast convergence of CWcurr toward CWmin works against lower priority traffic for the benefit of higher priority traffic. As can be easily seen in the Fig. 5 (a), the throughput of the CR-AEDCF scheme is more stable than that of SR-AEDCF and is able to maintain a bit rate output at about 550 Kbps, even after the other three nodes entered the competition for channel utilization. The cost of this however is that traffic, including Excellent-effort Video transmission (see Table 2), may be blocked for prolonged periods of time from accessing the channel (Fig. 5 (d)). On the other hand, the moderate oscillations exhibited by the bit rate output of the SR-AEDCF scheme allows other types of traffic to get an opportunity to use the wireless channel. In Fig. 6, we provide a summary view of the results in Fig. 5 by plotting the averages with the corresponding standard deviations. Here, the relative performance of the five
3796
H. Artail et al. / Computer Communications 29 (2006) 3789–3803 802.11e CR-AEDCF
SSD SR-AEDCF
AMPA
802.11e CR-AEDCF 0.22
0.65
SSD SR-AEDCF
AMPA
Access category: AC_VI
0.19
Time (sec.) 802.11e CR-AEDCF
89
81
73
65
57
49
40
Time (sec.)
SSD SR-AEDCF
AMPA
802.11e CR-AEDCF 0.016
0.08
SSD SR-AEDCF
AMPA
Access category: AC_BK
Access category: AC_BE 0.012
0.06
Bit Rate (Mbps)
0.04
0.02
0.008
0.004
Time (sec.)
89
81
73
65
57
49
40
32
24
0
16
0.000
89
81
73
65
57
49
40
32
24
16
8
0
0.00
8
Bit Rate (Mbps)
32
0
81
72
63
54
45
36
0.10
27
0.25
18
0.13
9
0.35
24
0.16
8
0.45
16
Bit Rate (Mbps)
0.55
0
Bit Rate (Mbps)
Access category: AC_VO
Time (sec.)
Fig. 5. Bit rate performance comparison for the four access categories.
1.000 802.11e
SSD
CR-AEDCF
AMPA
SR-AEDCF
AC_VI
0.100
AC_BE
0.010 AC_BK
Node 3
Node 2
Node 1
Node 3/0
Node 2
Node 1
Node 3/0
Node 2
Node 1
Node 3/0
Node 2
Node 2
Node 1
Node 0
0.000
Node 1
0.001
Node 3/0
Average bit rate (Mbps)
AC_VO
Fig. 6. Averages and standard deviations of the bit rate performance of the four access categories (four nodes) for the five simulated schemes.
schemes is perhaps clearer, as reflected by the values of the averages and the lengths of the standard deviation bars. The 802.11e standard understandably exhibits the worst performance in terms of averages and standard deviations, as may be explained by the fact that the other four schemes were designed to improve upon this standard. With regard to SSD and AMPA, surprisingly the simpler SSD performs better. When comparing the best two performers; our SR-AEDCF scheme and CR-AEDCF, it is obvious that SR-AEDCF provides comparable highest-priority performance (Node 0), somewhat better performance at the next highest priority (Node 1), and a much better performance
by Node 2 (access category AC_BE). Finally, it could be noted that the overlapping standard deviation ranges (confidence levels) may be attributed to how the schemes react to fluctuating network conditions while the negative values shown in the case of CR-AEDCF are strictly due to curve fitting and do not represent real data. 4.2. Packet loss performance Fig. 7 presents the packet loss performance for all schemes in all four access categories. Such packets were lost during transmissions and thus represent the number of
H. Artail et al. / Computer Communications 29 (2006) 3789–3803 802.11e CR-AEDCF
SSD SR-AEDCF
AMPA
802.11e SSD CR-AEDCF SR-AEDCF 130 Access category: AC_VI
Access category: AC_VO
Number of lost packets
60
40
20
110
100
Time (sec.) 802.11e CR-AEDCF
89
81
73
65
57
49
40
32
24
0
89
81
73
65
57
49
40
32
24
8
16
0
8
90
0
Time (sec.)
SSD SR-AEDCF
AMPA
802.11e CR-AEDCF
300
SSD SR-AEDCF
AMPA
80 Access category: AC_BE
Access category: AC_BK
Number of lost packets
240 180 120 60
60
40
20
89
81
73
65
57
49
40
32
24
16
89
81
73
65
57
49
40
32
24
16
8
0
Time (sec.)
8
0
0
0
Number of lost packets
AMPA
120
16
Number of lost packets
80
3797
Time (sec.)
Fig. 7. Packet loss performance comparison for the four access categories.
collisions as experienced by each source node. Since packet loss signifies the contention level in the channel, it follows that higher packet loss implies less coordination among the competing source nodes. Consequently, packet loss is an indicative measure of the effectiveness of the algorithm that is responsible for adapting the transmission of packets to the channel conditions. Hence, studying the packet loss performance of a given node gives different insights into the effectiveness of the adaptation algorithm that is employed by the node. Those insights can not be derived from the bit rate performance, which is strictly a throughput measure. The plots in Fig. 7 illustrate the superior performance of the CR-AEDCF and SR-AEDCF schemes as compared to the others. An exception to this is the lowest priority category where SSD shows a zero packet-loss behavior whereas SR-AEDCF shows a small number of lost packets toward the end of the simulation. When comparing the performance of SR-AEDCF to that of CR-AEDCF in Fig. 7, we notice that in the highest priority category, a slightly higher number of packets were lost, while in the AC_VI and AC_BE categories, the same number of packets were lost on average. This, coupled with the results that were shown in Fig. 6, present major significance in that the much less-complex SR-AEDCF is almost as effective in adapting the DCF contention window size to channel conditions as the CR-AEDCF scheme. With regard to the three other schemes, including 802.11e, a much worse packet loss behavior is exhibited across all
access categories and throughout the entire simulation duration. We present a summary view of the data in Fig. 7 by plotting the total number of packets lost to that of the packets delivered for the four access categories in Fig. 8. 4.3. Packet end-to-end delay The third metric that was measured is source-to-destination packet delay for all nodes, while repeating the simulations for all five schemes. Packet delay indicates latency in the wireless network and reveals information that may not totally be inferred from the packet loss data. It is important to note that the delay includes all the activities that take place between the initial attempt by a station to transmit the concerned packet and the reception of this packet by the destination station. Hence, any retransmissions that may have taken place are included in this measurement. As seen in Fig. 9, the delays experienced by Nodes 1, 2, and 3 experience an initial high delay due to the utilization of the transmission medium by higher-level traffic. The magnitude of this initial delay understandably increases with additional nodes trying to transmit, as indicated in the figures. Node 1 for example had to compete with Node 0, while Node 2 had to compete with two other nodes transmitting higher priority traffic. By examining the Fig. 9, the superior performance of the SR-AEDCF scheme is observable in all access categories but it is closer to the IEEE 802.11e standard more than
3798
H. Artail et al. / Computer Communications 29 (2006) 3789–3803 3.0 Access category:AC_VO
Ratio of lost to delivered packets
Ratio of lost to delivered packets
3 2.5 2 1.5 1 0.5
2.0 1.5 1.0 0.5 0.0
0 802.11e
3.0
SSD
AMPA
CRSRAEDCF AEDCF
802.11e
SSD
AMPA
CRSRAEDCF AEDCF
3.0
Access category:AC_BE
Ratio of lost to delivered packets
Ratio of lost to delivered packets
Access category:AC_VI 2.5
2.5 2.0 1.5 1.0 0.5
Access category:AC_BK 2.5 2.0 1.5 1.0 0.5 0.0
0.0 802.11e
SSD
AMPA
802.11e
CRSRAEDCF AEDCF
SSD
AMPA
CRSRAEDCF AEDCF
Fig. 8. Ratio of number of lost packets to that of delivered packets.
802.11e CR-AEDCF 0.014
SSD WS-AEDCF
AMPA
802.11e CR-AEDCF
SSD WS-AEDCF
AMPA
0.30
Access category: AC_VO
Access category: AC_VI 0.25
Packet delay
Packet delay
0.012
0.010
0.20 0.15 0.10
0.008 0.05
Time (sec.) 802.11e CR-AEDCF
89
81
73
65
57
49
Time (sec.)
SSD WS-AEDCF
AMPA
802.11e CR-AEDCF
SSD WS-AEDCF
AMPA
30.0
6.0
Access category: AC_BK
Access category: AC_BE 5.0
Packet delay
24.0
4.0 3.0 2.0
18.0 12.0 6.0
1.0
Time (sec.)
Fig. 9. Packet end-to-end delay for the four access categories.
89
81
73
65
57
49
40
32
24
16
8
89
81
73
65
57
49
40
32
16
8
24
Time (sec.)
0
0.0
0.0
0
Packet delay
40
32
24
8
0
89
81
73
65
57
49
40
32
24
8
16
0
16
0.00
0.006
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
to any other scheme. Notably, the delay of transmission in the CR-AEDCF scheme stagnates in the two lowest priority categories after the initial rise and may be considered unacceptably high for many applications. It should be noted that the actual delay values are not indicative of the true delays that would occur in a real system as they are greatly affected by the simulation environment and employed tools. However, one can certainly trust the relative performance across schemes and across priorities. In the upper left graph of Fig. 9, the average delay of highest priority traffic is around 8 milliseconds (ms), while the steady sustainable delay for categories AC_VI, AC_BE, and AC_BK are 26, 100, and 750 ms, respectively. Finally, it is worth noting that packet delay can be correlated with the bit rate performance since faster transmission (lower delays) implies higher throughput and the other way around is true as well. 4.4. Transmitting at equal priorities In the above simulations, we have demonstrated the superiority of SR-AEDCF in scenarios where nodes of different priority classes are competing for transmission. In this scenario we investigate the behavior of SR-AEDCF when nodes with equal priority are competing. Considering the fact that the aim of introducing IEEE 802.11e is to provide bandwidth guarantees for voice and video streaming traffic we limit our evaluation to AC_VO class (highest priority), however the presented analysis will allow generalization to the remaining priority classes. As described before, IEEE 802.11e classifies traffic in eight priorities depending on application type where traffic with higher priority is serviced ahead of that with less priority. This approach may be illustrated as a queuing system with four queues (four access categories) where a single server provides service in the following approach. A client (e.g., traffic) is always selected for service from the queue with the highest priority. The server selects traffic from a queue with less priority when no more traffic is present in the higher priority queues. Moreover, demonstrating the effect of traffic prioritization filters down into analyzing the performance of the aforementioned queuing model and in particular the behavior of the server. Thus it is evident that traffic with high priority is serviced faster than that of low priority. It is also evident that the average waiting time and average size of the queue (measures directly related to bit rate, packet delay, and drop rate) for traffic in high priority queues are lower than those in lower priority queues. Using the queuing system analogy, we assume four sources, each generating traffic in a different queue where the four queues have the same priority. Having an equal priority, each queue is randomly selected by the server and traffic is serviced without any preference or differentiation. In this case, it is evident that the average wait time and size of the queue are the same for the four queues. Generalizing these queuing results to network performance, we expect that no guarantees will be provided to
3799
any of the nodes and the bandwidth will be equally divided between them. In fact, having different queues with equal priorities is identical to having a single queue with no priority. We may thus argue that in the case of equal priority nodes, the performance of IEEE 802.11e is close to that of IEEE 802.11a with no bandwidth guarantees. The aim of SR-AEDCF is adapting CW and thus, it has no effect on the core EDCA access mechanism. Hence, we expect that its performance in similar priority traffic conditions will be similar to that of IEEE 802.11e and consequently close to IEEE 802.11a. To confirm this behavior, four NS-2 simulation scenarios corresponding to the four access priorities were conducted, where the same topology was simulated for SR-AEDCF, original IEEE 802.11e, and IEEE 802.11a. In each scenario, four source nodes with identical priorities transmitted packets over UDP with a transmission rate equal to 600 Kbps. The purpose was to investigate the performance in terms of effective bit rate for all priorities. Fig. 10 shows that SR-AEDCF, like IEEE 802.11e and IEEE 802.11a, does not provide any guarantees in terms of bandwidth. Moreover, the bit rates are equally spread among the four nodes, a behavior that is due to the similarity in how the access mechanism works for equal priority conditions. As with EDCA, we believe that integrating SR-AEDCF with an access mechanism that provides guarantees in conditions where equal priority traffic is present will improve the performance. 4.5. Simulation over FTP/TCP traffic In this section we evaluate the performance of SRAEDCF for FTP/TCP traffic. In this simulation, we assume a network topology of 40 nodes. We start with a minimum traffic load that is generated by one node only then we gradually increase the traffic load by increasing the number of transmitting nodes till we reach a maximum network traffic load that is generated by all the 40 nodes. We investigate the effect of the network load on the TCP goodput and the number of segment retransmissions for the four adaptation schemes (SSD – AMPA – CR-AEDCF – SR-AEDCF). The TCP goodput is evaluated in terms of percentage gain relative to the original 802.11e goodput. Fig. 11 (a) shows that the goodput gain is negligible for the four discussed adaptation schemes for a network load generated by less than 8 wireless nodes. The figure shows that the gains for CR-AEDCF and SR-AEDCF are almost equal over the different network loads, while they both outperform AMPA and SSD. We observe that AMPA and SSD achieve equal gains for network loads generated by less than 18 nodes. SSD performance achieves 14% gain for network load generated by 20 nodes up to 40 nodes but AMPA outperforms SSD for this network load. We notice that CR-AEDCF and SR-AEDCF reach a 36% gain for a traffic load generated by all the 40 nodes. Fig. 11 (b) evaluates the TCP segment retransmissions in the four schemes. It shows that the four schemes have equal performance for a network traffic loads generated by less than 10
3800
H. Artail et al. / Computer Communications 29 (2006) 3789–3803 0.7
0.7 Access Priority: AC_VO
Access Priority: AC_VI 0.6
802.11a
0.5
802.11e
Bit rate (Mbps)
Bit rate (Mbps)
0.6
SR-AEDCF
0.4 0.3 0.2 0.1
802.11a
0.5
802.11e SR-AEDCF
0.4 0.3 0.2 0.1
0
0 0
10
20
30 40 50 Time (sec.)
60
70
80
0
0.7
10
30 40 50 Time (sec.)
60
70
80
0.7 Access Priority: AC_BE
0.6
Access Priority: AC_BK 0.6
802.11a
0.5
802.11e
Bit rate (Mbps)
Bit rate (Mbps)
20
SR-AEDCF
0.4 0.3 0.2 0.1
802.11a
0.5
802.11e SR-AEDCF
0.4 0.3 0.2 0.1
0
0 0
10
20
30 40 50 Time (sec.)
60
70
80
0
10
20
30 40 50 Time (sec.)
60
70
80
Fig. 10. Bit Rate for IEEE 802.11a, IEEE 802.11e, and SR-AEDCF for four nodes transmitting with equal priority at 600 Kbps.
900
40 SSD AMPA SR-AEDCF CR-AEDCF
700 600
SSD AMPA
35
SR-AEDCF CR-AEDCF
30 25
500
% Gain
Retransmissions per sec.
800
400 300
20 15
200
10
100
5 0
0 0
8
16 24 Number of stations
32
40
0
8
16 24 Number of stations
32
40
Fig. 11. Percent Gain of TCP goodput of the four schemes relative to 802.11e (left) and evaluation the TCP segment retransmissions (right).
nodes. On the other hand we notice that for traffic load generated by a number of nodes greater than 10, the SSD retransmission plot increases sharply while the CRAEDCF and SR-AEDCF maintain a relatively constant increase. For a load of 40 nodes SSD retransmissions are around 800 while CR-AEDCF and SR-AEDCF experience 400 retransmissions. AMPA’s performance is better than SSD but worse than CR-AEDCF and SR-AEDCF reaching 600 retransmissions with a traffic load of 40 nodes. This simulation revealed the remarkable performance of SR-AEDCF in TCP/FTP traffic conditions. Even though the performance of SR-AEDCF is similar to CR-AEDCF, SR-AEDCF presents a simpler and less complex adaptation framework.
4.6. Video transmission To confirm our simulation results, a scenario that involved the use of the EVALVID tool [3] to transmit an MPEG-4 video sequence is simulated. EVALVID is a package that can be integrated with NS-2. It allows the transmission of MPEG video sequences. The QoS of video streaming may be investigated by comparing the quality of video received at the destination nodes. The topology that was used consists of 20 nodes that are within transmission range to each other and in a severely congested network. The MPEG-4 sequence consists of 400 frames, each frame is fragmented into packets of 1000 bytes, and the rate is 30 frames per second. The sequence was sent from Node 1 to
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
Node 2 and the performance of each scheme is evaluated by inspecting the received video sequence. The other 18 nodes are generating constant bit rate traffic at 500 Kbits/s over UDP and are transmitting packets of size 1000 bytes each. Furthermore, these nodes are transmitting traffic with priority 4 so as to generate a heavily congested channel and test each scheme’s support of QoS. Samples of the received video sequences are shown in Fig. 12. As can be noticed from the figure, we also included the standard IEEE 802.11a in the simulations as a reference. This standard does not offer priority differentiation, which explains its very poor video quality, as experienced by the receiving node. Even though the main purpose of IEEE 802.11e is to provide support for QoS, we observe a reduced quality in the received video frames. By visually
3801
inspecting the received frames, we notice a similar quality under the AMPA (which works on adjusting the AIFS of the DCF) and the CR-AEDCF and SR-AEDCF schemes that work on adjusting the contention window size. On the other hand and as was expected, the performance of the simple slow decrease scheme (SSD) outperforms 802.11e but lags behind the three dynamic schemes in terms of QoS support. As shown in the bottom row of Fig. 12, this scheme experiences disorders at the start of transmission but the streaming quality becomes more enhanced in further frames but are delayed with respect to the frames achieved by other schemes. It is worth noting that we faced many problems (crashing at runtime) while programming the EVALVID tool to implement the CR-AEDCF scheme due to its
802.11a Standard
IEEE 802.11e scheme
Proposed SR_AEDCF scheme
CR-AEDCF scheme
AMPA scheme
SSD scheme
Fig. 12. Video streaming evaluation over the different schemes.
3802
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
rate, end-to-end packet delay, and packet drop rate. In addition, we evaluated the proposed approach, under heavy loads, using EVALVID tool to generate and transmit MPEG-4 video sequences. Results have shown the effectiveness of the proposed technique in providing a high performance scheme with a relatively low complexity.
41.0
PSNR (decibels)
37.0 33.0 29.0 25.0
802.11e, start frame: 94 AMPA, start frame: 138
21.0
SSD, start frame: 70 CR-AEDCF, start frame: 68
References
SR-AEDCF, start frame: 68
17.0 0
80
160 240 Frame Number
320
400
Fig. 13. PSNR for all the received from for the five schemes.
computational complexity and the requirement to track many variables. This justifies our motivation for designing a less complex scheme with the same or better effectiveness of an adaptation scheme that adjusts the MAC layer parameters to utilize the wireless channel in the most effective manner. To obtain a more concrete measure of quality, we generated the peak signal to noise ratio (PSNR) of the delivered video frames for all five discussed schemes with respect to the original video. We accomplished this by using a trial version of the YUVTools software [16], which is a set of utilities for playing, converting and analyzing YUV video data (including generating PSNR values). The starting frames in all five schemes were received damaged and delayed but to varying degrees. To generate reasonably accurate PSNR value, we clipped those frames using the tool above and aligned the first reasonably good frame in each scheme with the corresponding one in the original movie. Fig. 13 shows the PSNR values for each scheme, where it is shown that AMPA rendered close results to that of CR-AEDCF and SR-AEDCF although it had the highest number of damaged frames at the start, as indicated on the figure itself. It is also shown that SRAEDCF (heavy line) has a very comparable performance to that of CR-AEDCF, including the starting frames. 5. Conclusion In this paper, we have evaluated the performance of the IEEE 802.11e standard. We have also proposed a new adaptive differentiation technique for resetting the value of the contention window after each successful transmission. The proposed adaptive technique takes into account the current level of link utilization when resetting such value. We have performed several simulation scenarios, using NS-2, to evaluate the proposed technique compared to IEEE 802.11e, CR-AEFCF, AMPA, and SSD. During the evaluation, we have focused on three parameters: bit
[1] I. Aad, C. Castelluccia, Differentiation mechanisms for IEEE 802.11, in: Proceedings of IEEE Infocom 2001, Anchorage, Alaska, USA, April 2001, pp. 209–218. [2] M. Bandinelli, F. Chifi, R. Fantacci, D. Tarchi, G. Vannuccini, A link adaptation strategy for QoS support in IEEE 802.11e-based WLANs, in: IEEE Wireless Communications and Networking Conference, New Orleans, USA, March 2005, pp. 120–125. [3] Evalvid Simulation Tool, Available from. [4] M. Frikha, F. Ben Said; L. Maalej, F. Tabbana, Enhancing IEEE 802.11e standard in congested environments, in: International Conference on Internet and Web Applications and Services/Advanced International Conference on Telecommunications, 2006. AICT-ICIW ’06, Guadeloupe, France, February 2006, pp. 78–78. [5] IEEE Std 802.11. Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specification, 1999 edition. [6] IEEE Std 802.11e-2005 IEEE Standard for Information technology – Telecommunications and information exchange between systems – LAN/MAN Specific requirements Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications Amendment 8: Medium Access Control (MAC) Quality of Service Enhancements. [7] D. Gu, J. Zhang, QoS enhancement in IEEE 802.11 wireless area networks, IEEE Commun. Mag. 41 (6) (2003) 120–124. [8] A. Ksentini, M. Naimi, A. Nafss, M. Gueroui, Adaptive service differentiation for QoS provisioning in IEEE 802.11 wireless ad hoc networks, in: International Conference on Mobile Computing and Networking, Proceedings of the First ACM International Workshop on Performance Evaluation of Wireless Ad hoc, Sensor, and Ubiquitous Networks, Venezia, Italy, 2004, pp. 39–45. [9] S. Mangold, S. Choi, P. May, G. Hiertz, O. Klein, B. Walke, Analysis of IEEE 802.11e for QoS support in wireless LAN, IEEE Wireless Commun. (2003) 40–50. [10] J. Naoum-Sawaya, B. Ghaddar, S. Khawam, H. Safa, H. Artail, Z. Dawy, Adaptive approach for QoS support in IEEE 802.11e wireless LAN, in: IEEE International Conference on Wireless And Mobile Computing, Networking And Communications, 2005, (WiMob’2005), pp. 167–173, Montreal, Canada, August 22–24, 2005. [11] Q. Ni, Performance analysis and enhancements for IEEE 802.11e wireless networks, IEEE Netw. 19 (4) (2005) 21–27. [12] Q. Ni, L. Romdhani, T. Turletti, A survey of QoS enhancements for IEEE 802.11 wireless LAN, Wireless Commun. Mobile Comput. 4 (2004) 547–566. [13] NS-2 simulator (April 2002). .
H. Artail et al. / Computer Communications 29 (2006) 3789–3803 Hassan Artail worked as a system development supervisor at the Scientific Labs of DaimlerChrysler, Michigan before joining AUB in 2001. At DaimlerChrysler, he worked for 11 years in the field of software and system development for vehicle testing applications, covering the areas of instrument control, computer networking, distributed computing, data acquisition, and data processing. He obtained a B.S. and M.S. in Electrical Engineering from the University of Detroit in 1985 and 1986 respectively and a Ph.D. in Electrical and Computer Engineering in from Wayne State University in 1999. His research is in the areas of Internet and Mobile Computing, Distributed Computing and Systems, and computer networking and security.
Haidar Safa received a B.Sc. in Computer Science in 1991 from Lebanese university, Lebanon, M.Sc. in Computer Science in 1996 from University of Quebec at Montreal (UQAM), and a Ph.D. in Electrical and Computer Engineering in 2001 from Ecole Polytechnique de Montreal. He joined ADC Telecommunications and SS8 Networks in 2001 where he worked on designing and developing networking and system software. In 2003, he joined AUB where he is currently teaching and doing research in Networking. Dr. Safa’s research interests include mobility management in wireless networks, Quality of Service, plus routing and network security.
3803
Joe Naoum-Sawaya finished the Bachelor degree in Computer and Communication Engineering from the American University of Beirut in 2005. He is currently pursuing his graduate studies at the University of Waterloo, in the department of Management Sciences. His research interests involve the application of large-scale linear and integer optimization to problems inspired from communication networks.
Bissan Ghaddar finished the Bachelor degree in Computer and Communication Engineering from the American University of Beirut in 2005. She is currently pursuing her graduate studies at the University of Waterloo, in the department of Management Sciences. Her research interests are applying Semi-definite Optimization to large scale problems with application to communication networks.
Sami Khawam finished the Bachelor degree in Electrical Engineering from the American University of Beirut in 2005. His research interests are in computer networking and application software development.
A simple recursive scheme for adjusting the contention window size in IEEE 802.11e wireless ad hoc networks Hassan Artail, Haidar Safa *, Joe Naoum-Sawaya, Bissan Ghaddar, Sami Khawam American University of Beirut, P.O. Box: 11-0236, Riad El-Solh, Beirut 1107 2020, Lebanon Received 20 February 2006; received in revised form 16 June 2006; accepted 16 June 2006 Available online 14 July 2006
Abstract The IEEE 802.11e standard has been introduced recently for providing Quality of Service (QoS) capabilities in the emerging wireless local area networks. This standard introduces a contention window based Enhanced Distributed Channel Access (EDCA) technique that provides a prioritized traffic to guarantee the minimum bandwidth needed for time critical applications. However, the EDCA technique resets the Contention Window (CW) of the mobile station statically after each successful transmission. This static behavior does not adapt to the network state since it reduces the network usage and results in bad performance and poor link utilization whenever the demand for link utilization increases. This paper proposes a new adaptive differentiation technique for IEEE 802.11e Wireless Local Area Networks that takes into account the network state before resetting the contention window. In the new technique, the congestion level of the network is sensed by using previous CW values. Three other enhancement techniques that focus on network adaptation are also discussed. Their main limitations are the high complexity of the implemented algorithms and their slow adaptation to the network state when the channel experiences bursty traffic. The proposed technique is compared to the original differentiation techniques of IEEE 802.11a and IEEE 802.11e standards, as well as to the enhancement schemes. Results show that the proposed adaptive technique outperforms IEEE 802.11e and is comparable to the other enhancement schemes while maintaining relatively low complexity requirements. 2006 Elsevier B.V. All rights reserved. Keywords: Enhanced Distributed Channel Access; IEEE 802.11e; Quality of Service; Wireless LAN; MANET
1. Introduction The emerging IEEE 802.11 family of wireless technologies has shown tremendous growth and acceptance as a solution for Wireless Local Area Networks (WLANs). Simultaneously with the growth of WLANs, there is the trend for multimedia applications over IP with Quality of Service (QoS) support. The use of such applications is now a reality in corporate networks and promises to expand to the global Internet. The original IEEE 802.11 networks are best effort networks and do not support QoS for time critical applications [5]. A new IEEE 802.11 Standard (IEEE 802.11e) has been introduced to replace the best effort services by more sophisticated services that
*
Corresponding author. Tel.: +961 1 350000X4270. E-mail address: [email protected] (H. Safa).
0140-3664/$ - see front matter 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.comcom.2006.06.016
guarantee QoS attributes such as bandwidth, delay, and jitter [6]. This standard focuses on replacing the conventional Distributed Coordination Function (DCF) and the optional Point Coordination Function (PCF) of the Medium Access Control (MAC) layer by a Hybrid Coordination Function (HCF). The HCF defines two medium access mechanisms: a contention-based channel access referred to as Enhanced Distributed Channel Access (EDCA), and controlled channel access referred to as HCF Controlled Channel Access (HCCA). In this work we focus on the EDCA only. The EDCA is a Contention Window (CW) based channel access function and is based on priority allocation for different categories of traffic to guarantee the minimum bandwidth needed for high priority applications, thus increasing the network performance. Similarly to the traditional DCF technique, the EDCA technique of IEEE 802.11e employs a static reset to the CW. This static behavior is a drawback since it does not offer any
3790
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
room for adaptation to the network state [6]. It also reduces the network usage and is reflected in bad performance and poor QoS whenever the demand for medium utilization increases [7]. Several techniques have been proposed to enhance the performance of IEEE 802.11e by adapting CW to the network state [1,2,4,7,9,14,15]. Even though these techniques, which are presented in Section 2, outperform the EDCF technique used in IEEE 802.11e, their main limitations are the high complexity of their algorithms and their slow adaptation to the network state as explained later in Sections 2 and 3. Low complexity algorithms are favorable due to their simpler implementation and low demand on the system in terms of resource consumption. This paper introduces a new low complexity Adaptive EDCA algorithm that adapts the CW to channel conditions and adjusts it depending on the network utilization and performance. The proposed technique outperforms IEEE 802.11e and is comparable to the other enhancement schemes while maintaining relatively low complexity requirements and providing a faster adaptation to the network state. The rest of the paper is organized as follows. A brief overview of the DCF and EDCA of the IEEE 802.11 and 802.11e standards is first presented in the remainder of this section. In Section 2, we discuss the adaptive techniques that appeared recently in the literature, and then introduce the proposed scheme in Section 3. Section 4 introduces and evaluates the different techniques in terms of complexity and performance, and then finish in Section 5 with conclusions regarding our findings. 1.1. Overview of the MAC layer A review of the DCF and EDCA techniques of the IEEE 802.11 and 802.11e MAC layer is provided in this section. 1.1.1. IEEE 802.11 MAC layer The DCF technique of the IEEE 802.11 MAC layer employs a contention-based channel access function and uses carrier sensing multiple access with collision avoidance (CSMA/CA) mechanism [5]. The mechanism tries to avoid the event of two or more stations attempting to transmit a packet simultaneously, in which case an unsuccessful transmission occurs. Under DCF a station that intends to transmit monitors the channel until it is idle, and then waits for a time period of a Distributed Inter-Frame Space (DIFS) length. After sensing an idle channel for a DIFS period, the mobile station randomly selects a backoff timer (time slot) within a backoff window. The backoff timer is decreased only when the medium is idle; it is frozen when another station is transmitting (see Fig. 1). Each time the medium becomes idle, the station waits for a DIFS and continuously decrements the backoff timer. As soon as the backoff timer expires, the station is authorized to access the medium and transmit. The backoff timer is derived from a uniform distribution over the interval [0, CW 1], where CW is a value between [CWmin, CWmax].
Medium is idle
Medium is busy
Random backoff timer, value ∈ [0, CW], CW ∈[CWmin, CWmax] Transmission Time DIFS Medium is busy
Slot Time
Medium is idle Station enters backoff
Transmission by another station
Station resumes backoff Transmission Time
DIFS
DIFS
Fig. 1. IEEE 802.11 DCF Channel Access.
At the very first transmission attempt, CW value is equal to the initial backoff window size CWmin. For every unsuccessful transmission the value of CW is doubled until CWmax is reached. After transmitting a frame, the station expects to receive an acknowledgment (ACK) from the destination station following a Short Inter-Frame Space (SIFS) time. If the acknowledgment is not received, the sender assumes that the transmitted frame was collided, so it schedules a retransmission and enters the backoff process again. After every successful transmission the CW is reset to CWmin. DCF employs a discrete time backoff scale. The time that immediately follows the DIFS is slotted and the station is permitted to transmit only at the beginning of each slot time. The length of the slot time is set equal to the time needed at any station to detect the transmission of a packet from any other station. In all, DCF can only support best-effort services, not any QoS guarantees [9,12]. Typically, time-bounded services such as voice over IP, or audio/video conferencing require specified bandwidth, delay and jitter, but can tolerate some losses. However, in DCF mode, all the stations compete for the resources and channel with the same priorities. There is no differentiation mechanism to guarantee bandwidth, packet delay and jitter for high-priority traffic or multimedia flows [7]. All the MAC parameters including SIFS, DIFS, slot time, CWmin, and CWmax are given in Table 1 for the Direct Sequence Spread Spectrum (DSSS) physical channel (PHY). However, the MAC evaluation in the following sections is valid irrespective of the underlying PHY layer. 1.1.2. IEEE 802.11e MAC layer The new IEEE 802.11e standard introduces a Hybrid Coordination Function (HCF) [6], which defines two medium access mechanisms: a contention-based channel access referred to as Enhanced Distributed Channel Access (EDCA), and controlled channel access referred to as HCF Controlled Channel Access (HCCA). In this work
H. Artail et al. / Computer Communications 29 (2006) 3789–3803 Table 1 MAC Parameters for IEEE 802.11 Parameters
Value
SIFS DIFS CWmin CWmax Slot time
10 ls 50 ls 32 1024 20 ls
we focus on the EDCA only, in which QoS is realized with the introduction of Access Categories (ACs) and multiple independent backoff entities. Frames are delivered by parallel backoff entities within one 802.11e station. EDCA provides differentiated distributed channel access for eight traffic priorities that are mapped into four queues (ACs) as shown in Table 2. Thus, four backoff entities exist in every 802.11e station. Each AC queue works as an independent DCF station and uses its own contention parameters such as Arbitration Inter-Frame Space (AIFS), CWmin, and CWmax. Similar to a DCF station, each AC starts a backoff timer after detecting an idle channel for a time interval equal to an AFIS length. The backoff value is chosen to be a random number between [1, CW + 1], where initially CW is set to CWmin and increases whenever collision occurs up to CWmax. CW increases in accordance with the following equation [6]: CW new ½AC ¼ 2 CW current ½AC þ 1
ð1Þ
In case of successful transmission, the CW value of the AC queue is reset to CWmin. Additionally, in EDCA high priority traffic has a smaller AIFS, CWmax, and CWmin than low priority traffic as shown in Table 3. Thus, higher priority traffic will enter the contention period and access the wireless medium earlier than lower priority traffic. For example if i has lower priority than j, it follows that:
Table 2 User traffic priorities mapped to access categories User priority
Designation
Access category
1 2 0 3 4 5 6 7
BK (Background) BK (Background) BE (Best-effort) EE (Video/Excellent-effort) CL (Video/Controlled Load) VI (Video) VO (Voice) NC (Network Control)
AC_BK AC_BK AC_BE AC_BE AC_VI AC_VI AC_VO AC_VO
Table 3 Access categories Access category
AC_VO
AC_VI
AC_BE
AC_BK
AIFS CWmin CWmax TXOPlimit
2 7 15 0.003008
2 15 31 0.00601
3 31 1023 0
7 31 1023 0
CW min ½i > CW min ½j CW max ½i > CW max ½j
3791
ð2Þ
AIFS½i > AIFS½j In IEEE 802.11e, whenever a station seizes the channel, it will be able to transmit one or more frames for a Transmission Opportunity time interval (TXOP). A TXOP is an interval of time during which a given station has the right to transmit frames. A TXOP is defined by its starting time and duration. The duration of TXOP is limited by a parameter referred to as TXOPlimit (for more details about EDCA, refer to [6]). The main problem of this EDCA ad-hoc mode is that the values of CWmin, CWmax and backoff function of each queue are static and do not take into account wireless channel conditions. 2. Related work In [14], a new scheme is proposed for the purpose of improving the performance of IEEE 802.11e under different load rates and increasing the service differentiation in EDCF-based networks. The scheme uses a dynamic procedure to change the contention window value after a collision or a successful transmission. While in EDCF CWmin[i] and CWmax[i] values are statically set for each access category, this particular scheme, which we refer to as collision rate-based adaptive EDCF (CR-AEDCF) resets its CWs to adaptive values that are different from the CWmins, taking into account their current sizes and the average collision rate. This is driven by the assumption that when a collision occurs, another one is expected in the near future, and so to avoid bursty collisions, this scheme updates the contention window with values that are close to the previous ones but slowly decrease toward CWmins after successful transmissions. Furthermore, the scheme suggests changing the mechanism of doubling the CWs when a collision occurs and uses different factors that correspond to the different classes to increase the CWs. More specifically, each class updates its CW taking into account the estimated collision rate, which is calculated as a smoothed version of the ratio of the number of collisions to the total number of packets sent during a fixed time. Next, a multiplicative factor is computed using the collision rate such that it maintains the priority relationship between different classes and is upper-bounded by 0.8. This factor is then multiplied by the previous value of CW to become the new value while making sure that it does not fall below CWmin. In all, it is expected that on average, CR-AEDCF improves the performance of all types of traffics and outperforms the EDCF technique used in IEEE 802.11e by adaptively varying the contention window. A simple Static Slow Decrease (SSD) scheme was evaluated in [1]. The main purpose of this scheme is to slowly decrease CW values by employing a static factor (with a value of 0.5). It is obvious that such an approach is not
3792
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
suitable for all network conditions. This scheme is described in Eq. (3) below.
3. Proposed scheme
CW new ¼ 0:5 ðCW current CW min Þ þ CW min
The static reset method employed in IEEE 802.11 and 802.11e has been proved to be ineffective in maximizing channel utilization whenever the demand for channel access (e.g. traffic load) increases [8,11,14]. The probability of collisions in a busy wireless channel is high and will likely cause CW to approach CWmax almost every time before the station succeeds in transmitting its data. After a successful transmission, and the next time the station wants to transmit a frame during similar channel conditions, CW will start at CWmin and will subsequently increase (double) at each unsuccessful step until the station succeeds in transmitting its packet. As a result, a station may have to try several times before it succeeds and hence, significant time is wasted due to resetting CW to CWmin and consequently, the channel becomes underused. In low channel demand, the value of CW tends to acquire low values (close to CWmin and distant from CWmax) due to the low probability of collisions. In these conditions, high values of CW will lead to unnecessary waiting while the channel is idle. Consequently, it is favorable to adopt an adaptation technique that estimates the channel contention level and reflects the estimation in the value of CW. With this, we share the same objective as that of the CR-AEDCF but with the additional goal that the computational complexity of the algorithm is low. In high collision probabilities, the value of CW is driven toward higher values that are accustomed to the channel state, therefore getting values distant from CWmin and close to CWmax. To start with, our proposed scheme, which is a extension to our work in [10], needs to acquire channel state information in order to adjust the value of CW. Rather than measuring the collision rate directly, as in the CRAEDCF scheme, we use an indirect method that will be proven to be equally effective, through the recursive computation of the window size for CW, hence the name Simple Recursive Adaptive Enhanced DCF (SRAEDCF). This is justified by the fact that the value of CW directly reflects the state of the channel, as was detailed above. At the risk of repetition, we mention the following:
ð3Þ
In [8], an approach was proposed to replace the EDCA technique and is based on adapting the AIFS in response to network conditions, hence its name: Adaptive MAC Parameters (AMPA). The rationale behind adapting AIFS is to reduce the waiting time for high priority applications and increasing it for low priorities. When there is congestion in the network, high priority AIFS is decreased while best effort AIFS is increased. Conversely, if the network is in normal conditions, high priority AIFS is increased while best effort AIFS is decreased. It follows that AMPA favors high priority streams when the network is overloaded then and tries to serve all traffic when the network is in normal conditions. As in CR-AEDCF, network congestion level is first estimated using the ratio of the number of collisions to the number of sent packets. This value is then compared to threshold values above which the network is defined as congested leading to a decrease in AIFS. Conversely, below the threshold, the network is assumed in normal conditions and AIFS is increased for high priorities. Threshold values should be adapted to traffic priorities hence having higher thresholds for lower priorities. Hence, such a differentiation favors higher priority traffic to gain access to the channel. Threshold values (THX) are shown in Table 4. In general, AMPA respects particular QoS metrics, by providing bounded delays, and better throughput. This scheme achieves a stable performance regardless of the traffic load by using dynamic adaptation of AIFS depending on the network status. Recently, a link adaptation strategy that provides differentiation not only at the MAC layer, but also at the PHY layer was proposed in [2]. This strategy exploits the positive ACK procedure to evaluate the quality of the link. If the transmitter does not receive ACK frame it concludes that the last transmission was failed. For each active link, a transmitter maintains two counters, a success counter and a failure counter. If a frame is successfully transmitted, then the success counter is incremented by one and the failure counter is reset to 0. If the transmission fails then the failure counter is incremented by one and the success counter is reset to 0. These two counters are used to determine whether the quality of the link is good or not. The transmission rate is adjusted with respect to the link quality. Another recent approach that combines the slow decrease and the dynamic adaptation of CW was proposed in [4]. Table 4 Threshold differentiation Access category
AIFS
THX
AC_BK AC_BE AC_VI AC_VO
7 3 2 2
0.20 0.15 0.10 0.05
1. The value of CW varies in the range [CWmin, CWmax]. 2. When the channel is congested, CW will acquire values close to CWmax and distant from CWmin. 3. When the channel is free,CW will acquire values close to CWmin and distant from CWmax. Hence, it is feasible to estimate the channel congestion level by taking into consideration the current distance of CW from the limits CWmin and CWmax. That is, by dividing the distance (CWcurr CWmin) by the maximum distance (CWmax CWmin), we get an indication for the channel congestion level. It follows that the estimated congestion level can be written as:
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
CW curr CW min CW max CW min
ð4Þ
In the above, CF is a certainty factor (CF) that is included in the ratio in order to relate to the time interval between subsequent channel estimations, i.e., successful transmissions. In particular, the certainty of the estimate should be high whenever the time difference between subsequent estimates is small and should be low whenever the time difference is large. Actually, the value of CF should accomplish two functions: 1. Provide a high weight (between 0 and 1) when the difference in time between consecutive successful transmissions is in the order of milliseconds and a low weight when the difference is in the order of seconds (or minutes). 2. Maintain the differentiation between the four classes of access categories. For instance, video and audio streaming applications should correspond to higher weights relative to applications with no real-time requirements. To satisfy the first objective, a time decaying function such as the one in Eq. (5) may be used. This function, which is suitable for the AC_VI access category, provides an upper limit of 0.7 when the time difference is small and a lower limit of 0.4 when the difference is large. CF ¼ 0:3 expð0:001t2 Þ þ 0:4
ð5Þ
The 0.4 and 0.7 limits were chosen rather than the 0 and 1 as limits for the certainty factor. The main reason for this choice is that factors other than busy channel (e.g., buffer overflow) may introduce packet dropping. Thus limiting the certainty factor to a smaller length will allow the algorithm to account for these factors. However, rather than keeping track of the times of when successful transmission took place and performing extra computations at the MAC layer level, we suggest relying on application differentiation for weight variation. This is mainly due to the fact that the time between subsequent transmissions is directly dependent on application type. For instance, video and audio streaming applications (priorities 0 and 1) require fast frame transmission while other applications such as http and email services (priorities 2 and 3) transmit smaller number of frames in longer periods. This dependence on application time permits the assignment of higher weights for streaming applications and lower weights to other applications. Several simulations were run to assign weights on access categories and led to the values that are shown in Table 5. Following channel estimation, the next step is to figure out the value of CW that best adapts to this estimation. The proposed approach is based on estimating the distance from CWmin at which the probability of a successful transmission is maximized. Since CW has a minimum value of CWmin and a maximum value of CWmax, the maximum distance that CW can acquire is CWmax CWmin. To model a
Table 5 Values of CF Priority
CF
AC_VO AC_VI AC_BE AC_BK
0.8 0.7 0.4 0.3
slow decrease after each successful transmission, a wireless node should attempt transmission at a CW that is less than CWcurr, at which the node was able to transmit in the previous attempt. This slow decrease model limits the maximum distance to CWcurr CWmin. Since the ratio previously introduced in Eq. (4) acquires a value in the interval [0, CF], which increases with the congestion level, a good model for the estimated distance is: d ¼ ratio ðCW curr CW min Þ
ð6Þ
By making CWnew = CWmin + d and substituting (4) in (6), we get: CW new ¼ CW min þ ratio ðCW curr CW min Þ ¼ CW min þ CF
ðCW curr CW min Þ CW max CW min
2
ð7Þ
The ratio in (4) is a normalized value ranging between 0 and 1 · CF that reflects the weighted degree of channel contention, taking into account the four priority classes. This ratio would render a value close to 0 whenever the channel is free and a value close to 1 · CF whenever the channel is congested. Hence, multiplying this ratio by the factor (CWcurr CWmin) and adding the result to CWmin would result in a value bounded by [CWmin, CF · CWcurr]. This value of CWnew is thus a good representation of the backoff timer value needed for transmission for the given traffic priority, while taking into account the current network conditions. Given that our proposed scheme represents a simpler approach to CR-AEDCF, we compare the adaptability of CW to the variations in the channel conditions for both schemes. For this, we calculate and plot CWcurr versus the changes in channel congestion levels. In Fig. 2, we simulate Channel Condition
CR-AEDCF
SR-AEDCF
16 14 12 CW Value
ratio ¼ CF
3793
10 8 6 4 2 0 1
6
11 16 21
26 31 36
41 46 51 56 61 Time
66 71
76 81 86
91
Fig. 2. Value of CW as a function of channel conditions for the CRAEDCF scheme and our proposed scheme.
3794
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
a scenario in which a station is transmitting video streams during various channel conditions as depicted in the bottom curve, which indicates the occurrence of a collision or the lack of it. In this type of transmission, the class priority is 0 and hence, CWmin = 7 and CWmax = 15. Referring to the figure, between the time values 1 and 8, the channel is free and the wireless node is transmitting at CWmin. As the congestion level is increased, the CW value is adapted to the congestion level as seen in the linear increase in both schemes. Between the times values 15 and 20, the channel experiences bursty variations, thus prompting responses in the value of CW: The CR-AEDCF scheme decreases CW to a value close to CWmin while our proposed scheme offers a more conservative adaptation, as reflected in the comparably higher values of CW. This conservative slow variation offers better performance in bursty channel conditions since the channel is free momentarily and goes back to the busy state. Furthermore and as noticed when the channel gets more congested at time = 21, our scheme offers a faster adaptation, which brings the value of CW to the maximum value faster than CR-AEDCF. This conservative behavior, which suits bursty conditions, does not negatively impact performance when the channel is available for a period of time. When the channel is mostly idle (from time 37 to time 45), both schemes decrease CW steeply toward CWmin, although our proposed scheme is minimally lagging behind. A similar behavior can be noticed between the time values 55 and 63. 3.1. Complexity evaluation Complexity is a major concern for real time implementations. Low complexity algorithms are favorable due to their simpler implementation and low demand on the system in terms of resource consumption. In this section, we evaluate the complexity of our proposed approach and compare it to that of the other approaches discussed earlier. 3.1.1. Proposed scheme (SR-AEDCF) The ratio evaluation shown in Eq. (4) requires 2 subtractions, 1 division, and 1 multiplication. The evaluation of the new value of CW requires 1 multiplication and 1 addition. In total, this scheme requires 1 addition, 2 subtractions, 2 multiplications and 1 division. Finally in terms of storage, this scheme only requires one register for storing the value of CWcurr. 3.1.2. CR-AEDCF scheme The exact computational requirements can be retrieved from [14]: to compute the collision rate, 2 additions and one division are required. To do the smoothing, 1 subtraction, 1 addition, and 2 multiplications are additionally needed. One addition, 2 multiplications, and one comparison are needed to compute the multiplication factor. Finally, the new CW value is calculated by using 1 multiplication
and 1 comparison. As for storage, 6 registers are needed for storing the number of collisions, sent packets, and other variables. In total, 4 additions, 1 subtraction, 5 multiplications, 2 comparisons, and 6 storage registers are needed. 3.1.3. SSD scheme This scheme sets CW to half of its previous value. As a result, it evaluates CW at each state by employing 1 addition, 1 subtraction, and 1 multiplication. 3.1.4. AMPA scheme This scheme requires 2 additions and 1 division to compute the collision ratio. Further, 1 addition and two registers are needed to compute the new value of AIFS (involves two comparisons). One division and one register are also required to compute the value of the threshold (THX). In all, this scheme requires 3 additions, 2 divisions, 2 comparisons, and 3 memory registers for data storage. The summary of the above evaluation is shown in Fig. 3. Of particular importance is the comparison to the CRAEDCF scheme, where it is obvious the much less complexity that our proposed scheme offers. 4. Performance evaluation SR-AEDCF and CR-AEDCF provide distinct mechanisms to quantify network conditions. As deeply discussed in the previous sections, the advantage in the SR-AEDCF relies only on the current CW value to accurately evaluate channel usage and decide on the next CW. As opposed to SR-AEDCF, CR-AEDCF relies on statistics by keeping record of the number of all packet drops and the number of all transmitted packets. The disadvantage of the latter approach is the tardiness of the algorithm to converge to the current channel conditions. For instance, packets dropped way in the past are relevant for the current decision. This contradicts with the nature of wireless systems that is known to tumble fast due to many factors. As shown in Fig. 2, SR-AEDCF proved high competence in fast adaptation to network conditions. Some modifications might enhance the performance of CR-AEDCF, these might include the use of a moving average rather than a
Fig. 3. Computational complexity comparison.
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
simple average for estimating channel usage. However as shown in Section 3.1, SR-AEDCF outperforms CR-AEDCF in terms of complexity, hence introducing modifications such as moving average to CR-AEDCF will certainly complicate it more. In the following section we investigate the performance of the different schemes and in particular SR-AEDCF and CR-AEDCF through simulations. The performed simulations reveal the performance of the discussed schemes in terms of the following three parameters: bit rate, end-to-end packet delay, and packet drop rate. The bit rate measures the effective bit rate achieved by each node (e.g., the number of bits transmitted per second. End to end packet delay is the time a packet takes to reach the destination (e.g., source o destination delay). Packet drop rate is the number of packets dropped before reaching their destination. These parameters provide direct figures of the supported QoS. The simulation topology of this scenario is simple. It consists of 8 mobile nodes: 4 source nodes and 4 destination nodes. The four source nodes are labeled 0 through 3 and each node is transmitting with a different priority. Node 0 has a higher priority than Node 1, which has a higher priority than Node 2, which has a higher priority than Node 3. Each source is a constant bit rate source over UDP (User Datagram Protocol) with a transmission rate equal to 600 Kbps. The size of a transmitted packet is 512 bytes and it was assumed that all the nodes are in transmission range. The bit rate performance that is reported below was measured at the destination nodes and thus, reflects what the source nodes were able to transmit. Finally, we mention that the transmission channel was modeled using the Two-Ray model and show the simulation topology in Fig. 4. To simulate the proposed scheme, we used version 2.28 of the network simulator NS-2 [13], which implements many data link, network, and transport protocols and can be used to simulate large networks. An important feature of ns2 is the wireless model, which simplifies the simulation process of the model.
Node 4 Node 5
Node 0 Node 1
Node 6 Node 7
Node 2 Node 3 Fig. 4. Simulation Topology.
3795
4.1. Bit rate performance We start with the comparison of the bit rate performance for the five schemes: IEEE 802.11e standard, Collision Rate-based Adaptive EDCF (CR-AEDCF), Static Slow Decrease (SSD), Adaptive MAC Parameters (AMPA), and our own Simple Recursive Adaptive EDCF (SR-AEDCF). Fig. 5 presents the comparison of the four access categories (see Table 2) in four corresponding graphs. Further and to make the curves traceable, the results were smoothed using a moving window average with a window size equal to 10 samples. As can be implied from Fig. 5, Node 0 starts transmitting at time T = 0 s (Fig. 5 (upper-left)), Node 1 starts T = 10 s (Fig. 5 (upper-right)), Node 2 at T = 20 s (Fig. 5 (lower-left)), and Node 3 at T = 30 s (Fig. 5 (lower-right)). During the initial 10 s, Node 0 is the only one transmitting and consequently is able to occupy the entire available bandwidth. At T = 10 s, Node 1 starts sharing the channel resources with Node 0, which explains the relative drop in the throughput of the latter. Additionally, the bit rate plots show heavy oscillations as the number of transmitting nodes increases, although much of the oscillations were dampened as a result of the applied smoothing. When comparing the schemes, especially our proposed scheme to CR-AEDCF, the highest priority node was able to perform equally well, on average, under both schemes. The other three schemes, including IEEE 802.11e, exhibit much poorer performance as the number of nodes increases. Another notable observation is the approximate steady performance of both schemes, which indicates the effectiveness of the priority differentiation mechanism under the CR-AEDCF and SR-AEDCF schemes. Starting with the AC_VI access category (Node 1), SRAEDCF outperforms CR-AEDCF throughout the simulation. The difference in performance is more prevalent with Nodes 2 and 3, which is significantly important since it does not deprive lower priority applications from being able to access the wireless network. Referring back to Fig. 2, it is obvious that the fast convergence of CWcurr toward CWmin works against lower priority traffic for the benefit of higher priority traffic. As can be easily seen in the Fig. 5 (a), the throughput of the CR-AEDCF scheme is more stable than that of SR-AEDCF and is able to maintain a bit rate output at about 550 Kbps, even after the other three nodes entered the competition for channel utilization. The cost of this however is that traffic, including Excellent-effort Video transmission (see Table 2), may be blocked for prolonged periods of time from accessing the channel (Fig. 5 (d)). On the other hand, the moderate oscillations exhibited by the bit rate output of the SR-AEDCF scheme allows other types of traffic to get an opportunity to use the wireless channel. In Fig. 6, we provide a summary view of the results in Fig. 5 by plotting the averages with the corresponding standard deviations. Here, the relative performance of the five
3796
H. Artail et al. / Computer Communications 29 (2006) 3789–3803 802.11e CR-AEDCF
SSD SR-AEDCF
AMPA
802.11e CR-AEDCF 0.22
0.65
SSD SR-AEDCF
AMPA
Access category: AC_VI
0.19
Time (sec.) 802.11e CR-AEDCF
89
81
73
65
57
49
40
Time (sec.)
SSD SR-AEDCF
AMPA
802.11e CR-AEDCF 0.016
0.08
SSD SR-AEDCF
AMPA
Access category: AC_BK
Access category: AC_BE 0.012
0.06
Bit Rate (Mbps)
0.04
0.02
0.008
0.004
Time (sec.)
89
81
73
65
57
49
40
32
24
0
16
0.000
89
81
73
65
57
49
40
32
24
16
8
0
0.00
8
Bit Rate (Mbps)
32
0
81
72
63
54
45
36
0.10
27
0.25
18
0.13
9
0.35
24
0.16
8
0.45
16
Bit Rate (Mbps)
0.55
0
Bit Rate (Mbps)
Access category: AC_VO
Time (sec.)
Fig. 5. Bit rate performance comparison for the four access categories.
1.000 802.11e
SSD
CR-AEDCF
AMPA
SR-AEDCF
AC_VI
0.100
AC_BE
0.010 AC_BK
Node 3
Node 2
Node 1
Node 3/0
Node 2
Node 1
Node 3/0
Node 2
Node 1
Node 3/0
Node 2
Node 2
Node 1
Node 0
0.000
Node 1
0.001
Node 3/0
Average bit rate (Mbps)
AC_VO
Fig. 6. Averages and standard deviations of the bit rate performance of the four access categories (four nodes) for the five simulated schemes.
schemes is perhaps clearer, as reflected by the values of the averages and the lengths of the standard deviation bars. The 802.11e standard understandably exhibits the worst performance in terms of averages and standard deviations, as may be explained by the fact that the other four schemes were designed to improve upon this standard. With regard to SSD and AMPA, surprisingly the simpler SSD performs better. When comparing the best two performers; our SR-AEDCF scheme and CR-AEDCF, it is obvious that SR-AEDCF provides comparable highest-priority performance (Node 0), somewhat better performance at the next highest priority (Node 1), and a much better performance
by Node 2 (access category AC_BE). Finally, it could be noted that the overlapping standard deviation ranges (confidence levels) may be attributed to how the schemes react to fluctuating network conditions while the negative values shown in the case of CR-AEDCF are strictly due to curve fitting and do not represent real data. 4.2. Packet loss performance Fig. 7 presents the packet loss performance for all schemes in all four access categories. Such packets were lost during transmissions and thus represent the number of
H. Artail et al. / Computer Communications 29 (2006) 3789–3803 802.11e CR-AEDCF
SSD SR-AEDCF
AMPA
802.11e SSD CR-AEDCF SR-AEDCF 130 Access category: AC_VI
Access category: AC_VO
Number of lost packets
60
40
20
110
100
Time (sec.) 802.11e CR-AEDCF
89
81
73
65
57
49
40
32
24
0
89
81
73
65
57
49
40
32
24
8
16
0
8
90
0
Time (sec.)
SSD SR-AEDCF
AMPA
802.11e CR-AEDCF
300
SSD SR-AEDCF
AMPA
80 Access category: AC_BE
Access category: AC_BK
Number of lost packets
240 180 120 60
60
40
20
89
81
73
65
57
49
40
32
24
16
89
81
73
65
57
49
40
32
24
16
8
0
Time (sec.)
8
0
0
0
Number of lost packets
AMPA
120
16
Number of lost packets
80
3797
Time (sec.)
Fig. 7. Packet loss performance comparison for the four access categories.
collisions as experienced by each source node. Since packet loss signifies the contention level in the channel, it follows that higher packet loss implies less coordination among the competing source nodes. Consequently, packet loss is an indicative measure of the effectiveness of the algorithm that is responsible for adapting the transmission of packets to the channel conditions. Hence, studying the packet loss performance of a given node gives different insights into the effectiveness of the adaptation algorithm that is employed by the node. Those insights can not be derived from the bit rate performance, which is strictly a throughput measure. The plots in Fig. 7 illustrate the superior performance of the CR-AEDCF and SR-AEDCF schemes as compared to the others. An exception to this is the lowest priority category where SSD shows a zero packet-loss behavior whereas SR-AEDCF shows a small number of lost packets toward the end of the simulation. When comparing the performance of SR-AEDCF to that of CR-AEDCF in Fig. 7, we notice that in the highest priority category, a slightly higher number of packets were lost, while in the AC_VI and AC_BE categories, the same number of packets were lost on average. This, coupled with the results that were shown in Fig. 6, present major significance in that the much less-complex SR-AEDCF is almost as effective in adapting the DCF contention window size to channel conditions as the CR-AEDCF scheme. With regard to the three other schemes, including 802.11e, a much worse packet loss behavior is exhibited across all
access categories and throughout the entire simulation duration. We present a summary view of the data in Fig. 7 by plotting the total number of packets lost to that of the packets delivered for the four access categories in Fig. 8. 4.3. Packet end-to-end delay The third metric that was measured is source-to-destination packet delay for all nodes, while repeating the simulations for all five schemes. Packet delay indicates latency in the wireless network and reveals information that may not totally be inferred from the packet loss data. It is important to note that the delay includes all the activities that take place between the initial attempt by a station to transmit the concerned packet and the reception of this packet by the destination station. Hence, any retransmissions that may have taken place are included in this measurement. As seen in Fig. 9, the delays experienced by Nodes 1, 2, and 3 experience an initial high delay due to the utilization of the transmission medium by higher-level traffic. The magnitude of this initial delay understandably increases with additional nodes trying to transmit, as indicated in the figures. Node 1 for example had to compete with Node 0, while Node 2 had to compete with two other nodes transmitting higher priority traffic. By examining the Fig. 9, the superior performance of the SR-AEDCF scheme is observable in all access categories but it is closer to the IEEE 802.11e standard more than
3798
H. Artail et al. / Computer Communications 29 (2006) 3789–3803 3.0 Access category:AC_VO
Ratio of lost to delivered packets
Ratio of lost to delivered packets
3 2.5 2 1.5 1 0.5
2.0 1.5 1.0 0.5 0.0
0 802.11e
3.0
SSD
AMPA
CRSRAEDCF AEDCF
802.11e
SSD
AMPA
CRSRAEDCF AEDCF
3.0
Access category:AC_BE
Ratio of lost to delivered packets
Ratio of lost to delivered packets
Access category:AC_VI 2.5
2.5 2.0 1.5 1.0 0.5
Access category:AC_BK 2.5 2.0 1.5 1.0 0.5 0.0
0.0 802.11e
SSD
AMPA
802.11e
CRSRAEDCF AEDCF
SSD
AMPA
CRSRAEDCF AEDCF
Fig. 8. Ratio of number of lost packets to that of delivered packets.
802.11e CR-AEDCF 0.014
SSD WS-AEDCF
AMPA
802.11e CR-AEDCF
SSD WS-AEDCF
AMPA
0.30
Access category: AC_VO
Access category: AC_VI 0.25
Packet delay
Packet delay
0.012
0.010
0.20 0.15 0.10
0.008 0.05
Time (sec.) 802.11e CR-AEDCF
89
81
73
65
57
49
Time (sec.)
SSD WS-AEDCF
AMPA
802.11e CR-AEDCF
SSD WS-AEDCF
AMPA
30.0
6.0
Access category: AC_BK
Access category: AC_BE 5.0
Packet delay
24.0
4.0 3.0 2.0
18.0 12.0 6.0
1.0
Time (sec.)
Fig. 9. Packet end-to-end delay for the four access categories.
89
81
73
65
57
49
40
32
24
16
8
89
81
73
65
57
49
40
32
16
8
24
Time (sec.)
0
0.0
0.0
0
Packet delay
40
32
24
8
0
89
81
73
65
57
49
40
32
24
8
16
0
16
0.00
0.006
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
to any other scheme. Notably, the delay of transmission in the CR-AEDCF scheme stagnates in the two lowest priority categories after the initial rise and may be considered unacceptably high for many applications. It should be noted that the actual delay values are not indicative of the true delays that would occur in a real system as they are greatly affected by the simulation environment and employed tools. However, one can certainly trust the relative performance across schemes and across priorities. In the upper left graph of Fig. 9, the average delay of highest priority traffic is around 8 milliseconds (ms), while the steady sustainable delay for categories AC_VI, AC_BE, and AC_BK are 26, 100, and 750 ms, respectively. Finally, it is worth noting that packet delay can be correlated with the bit rate performance since faster transmission (lower delays) implies higher throughput and the other way around is true as well. 4.4. Transmitting at equal priorities In the above simulations, we have demonstrated the superiority of SR-AEDCF in scenarios where nodes of different priority classes are competing for transmission. In this scenario we investigate the behavior of SR-AEDCF when nodes with equal priority are competing. Considering the fact that the aim of introducing IEEE 802.11e is to provide bandwidth guarantees for voice and video streaming traffic we limit our evaluation to AC_VO class (highest priority), however the presented analysis will allow generalization to the remaining priority classes. As described before, IEEE 802.11e classifies traffic in eight priorities depending on application type where traffic with higher priority is serviced ahead of that with less priority. This approach may be illustrated as a queuing system with four queues (four access categories) where a single server provides service in the following approach. A client (e.g., traffic) is always selected for service from the queue with the highest priority. The server selects traffic from a queue with less priority when no more traffic is present in the higher priority queues. Moreover, demonstrating the effect of traffic prioritization filters down into analyzing the performance of the aforementioned queuing model and in particular the behavior of the server. Thus it is evident that traffic with high priority is serviced faster than that of low priority. It is also evident that the average waiting time and average size of the queue (measures directly related to bit rate, packet delay, and drop rate) for traffic in high priority queues are lower than those in lower priority queues. Using the queuing system analogy, we assume four sources, each generating traffic in a different queue where the four queues have the same priority. Having an equal priority, each queue is randomly selected by the server and traffic is serviced without any preference or differentiation. In this case, it is evident that the average wait time and size of the queue are the same for the four queues. Generalizing these queuing results to network performance, we expect that no guarantees will be provided to
3799
any of the nodes and the bandwidth will be equally divided between them. In fact, having different queues with equal priorities is identical to having a single queue with no priority. We may thus argue that in the case of equal priority nodes, the performance of IEEE 802.11e is close to that of IEEE 802.11a with no bandwidth guarantees. The aim of SR-AEDCF is adapting CW and thus, it has no effect on the core EDCA access mechanism. Hence, we expect that its performance in similar priority traffic conditions will be similar to that of IEEE 802.11e and consequently close to IEEE 802.11a. To confirm this behavior, four NS-2 simulation scenarios corresponding to the four access priorities were conducted, where the same topology was simulated for SR-AEDCF, original IEEE 802.11e, and IEEE 802.11a. In each scenario, four source nodes with identical priorities transmitted packets over UDP with a transmission rate equal to 600 Kbps. The purpose was to investigate the performance in terms of effective bit rate for all priorities. Fig. 10 shows that SR-AEDCF, like IEEE 802.11e and IEEE 802.11a, does not provide any guarantees in terms of bandwidth. Moreover, the bit rates are equally spread among the four nodes, a behavior that is due to the similarity in how the access mechanism works for equal priority conditions. As with EDCA, we believe that integrating SR-AEDCF with an access mechanism that provides guarantees in conditions where equal priority traffic is present will improve the performance. 4.5. Simulation over FTP/TCP traffic In this section we evaluate the performance of SRAEDCF for FTP/TCP traffic. In this simulation, we assume a network topology of 40 nodes. We start with a minimum traffic load that is generated by one node only then we gradually increase the traffic load by increasing the number of transmitting nodes till we reach a maximum network traffic load that is generated by all the 40 nodes. We investigate the effect of the network load on the TCP goodput and the number of segment retransmissions for the four adaptation schemes (SSD – AMPA – CR-AEDCF – SR-AEDCF). The TCP goodput is evaluated in terms of percentage gain relative to the original 802.11e goodput. Fig. 11 (a) shows that the goodput gain is negligible for the four discussed adaptation schemes for a network load generated by less than 8 wireless nodes. The figure shows that the gains for CR-AEDCF and SR-AEDCF are almost equal over the different network loads, while they both outperform AMPA and SSD. We observe that AMPA and SSD achieve equal gains for network loads generated by less than 18 nodes. SSD performance achieves 14% gain for network load generated by 20 nodes up to 40 nodes but AMPA outperforms SSD for this network load. We notice that CR-AEDCF and SR-AEDCF reach a 36% gain for a traffic load generated by all the 40 nodes. Fig. 11 (b) evaluates the TCP segment retransmissions in the four schemes. It shows that the four schemes have equal performance for a network traffic loads generated by less than 10
3800
H. Artail et al. / Computer Communications 29 (2006) 3789–3803 0.7
0.7 Access Priority: AC_VO
Access Priority: AC_VI 0.6
802.11a
0.5
802.11e
Bit rate (Mbps)
Bit rate (Mbps)
0.6
SR-AEDCF
0.4 0.3 0.2 0.1
802.11a
0.5
802.11e SR-AEDCF
0.4 0.3 0.2 0.1
0
0 0
10
20
30 40 50 Time (sec.)
60
70
80
0
0.7
10
30 40 50 Time (sec.)
60
70
80
0.7 Access Priority: AC_BE
0.6
Access Priority: AC_BK 0.6
802.11a
0.5
802.11e
Bit rate (Mbps)
Bit rate (Mbps)
20
SR-AEDCF
0.4 0.3 0.2 0.1
802.11a
0.5
802.11e SR-AEDCF
0.4 0.3 0.2 0.1
0
0 0
10
20
30 40 50 Time (sec.)
60
70
80
0
10
20
30 40 50 Time (sec.)
60
70
80
Fig. 10. Bit Rate for IEEE 802.11a, IEEE 802.11e, and SR-AEDCF for four nodes transmitting with equal priority at 600 Kbps.
900
40 SSD AMPA SR-AEDCF CR-AEDCF
700 600
SSD AMPA
35
SR-AEDCF CR-AEDCF
30 25
500
% Gain
Retransmissions per sec.
800
400 300
20 15
200
10
100
5 0
0 0
8
16 24 Number of stations
32
40
0
8
16 24 Number of stations
32
40
Fig. 11. Percent Gain of TCP goodput of the four schemes relative to 802.11e (left) and evaluation the TCP segment retransmissions (right).
nodes. On the other hand we notice that for traffic load generated by a number of nodes greater than 10, the SSD retransmission plot increases sharply while the CRAEDCF and SR-AEDCF maintain a relatively constant increase. For a load of 40 nodes SSD retransmissions are around 800 while CR-AEDCF and SR-AEDCF experience 400 retransmissions. AMPA’s performance is better than SSD but worse than CR-AEDCF and SR-AEDCF reaching 600 retransmissions with a traffic load of 40 nodes. This simulation revealed the remarkable performance of SR-AEDCF in TCP/FTP traffic conditions. Even though the performance of SR-AEDCF is similar to CR-AEDCF, SR-AEDCF presents a simpler and less complex adaptation framework.
4.6. Video transmission To confirm our simulation results, a scenario that involved the use of the EVALVID tool [3] to transmit an MPEG-4 video sequence is simulated. EVALVID is a package that can be integrated with NS-2. It allows the transmission of MPEG video sequences. The QoS of video streaming may be investigated by comparing the quality of video received at the destination nodes. The topology that was used consists of 20 nodes that are within transmission range to each other and in a severely congested network. The MPEG-4 sequence consists of 400 frames, each frame is fragmented into packets of 1000 bytes, and the rate is 30 frames per second. The sequence was sent from Node 1 to
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
Node 2 and the performance of each scheme is evaluated by inspecting the received video sequence. The other 18 nodes are generating constant bit rate traffic at 500 Kbits/s over UDP and are transmitting packets of size 1000 bytes each. Furthermore, these nodes are transmitting traffic with priority 4 so as to generate a heavily congested channel and test each scheme’s support of QoS. Samples of the received video sequences are shown in Fig. 12. As can be noticed from the figure, we also included the standard IEEE 802.11a in the simulations as a reference. This standard does not offer priority differentiation, which explains its very poor video quality, as experienced by the receiving node. Even though the main purpose of IEEE 802.11e is to provide support for QoS, we observe a reduced quality in the received video frames. By visually
3801
inspecting the received frames, we notice a similar quality under the AMPA (which works on adjusting the AIFS of the DCF) and the CR-AEDCF and SR-AEDCF schemes that work on adjusting the contention window size. On the other hand and as was expected, the performance of the simple slow decrease scheme (SSD) outperforms 802.11e but lags behind the three dynamic schemes in terms of QoS support. As shown in the bottom row of Fig. 12, this scheme experiences disorders at the start of transmission but the streaming quality becomes more enhanced in further frames but are delayed with respect to the frames achieved by other schemes. It is worth noting that we faced many problems (crashing at runtime) while programming the EVALVID tool to implement the CR-AEDCF scheme due to its
802.11a Standard
IEEE 802.11e scheme
Proposed SR_AEDCF scheme
CR-AEDCF scheme
AMPA scheme
SSD scheme
Fig. 12. Video streaming evaluation over the different schemes.
3802
H. Artail et al. / Computer Communications 29 (2006) 3789–3803
rate, end-to-end packet delay, and packet drop rate. In addition, we evaluated the proposed approach, under heavy loads, using EVALVID tool to generate and transmit MPEG-4 video sequences. Results have shown the effectiveness of the proposed technique in providing a high performance scheme with a relatively low complexity.
41.0
PSNR (decibels)
37.0 33.0 29.0 25.0
802.11e, start frame: 94 AMPA, start frame: 138
21.0
SSD, start frame: 70 CR-AEDCF, start frame: 68
References
SR-AEDCF, start frame: 68
17.0 0
80
160 240 Frame Number
320
400
Fig. 13. PSNR for all the received from for the five schemes.
computational complexity and the requirement to track many variables. This justifies our motivation for designing a less complex scheme with the same or better effectiveness of an adaptation scheme that adjusts the MAC layer parameters to utilize the wireless channel in the most effective manner. To obtain a more concrete measure of quality, we generated the peak signal to noise ratio (PSNR) of the delivered video frames for all five discussed schemes with respect to the original video. We accomplished this by using a trial version of the YUVTools software [16], which is a set of utilities for playing, converting and analyzing YUV video data (including generating PSNR values). The starting frames in all five schemes were received damaged and delayed but to varying degrees. To generate reasonably accurate PSNR value, we clipped those frames using the tool above and aligned the first reasonably good frame in each scheme with the corresponding one in the original movie. Fig. 13 shows the PSNR values for each scheme, where it is shown that AMPA rendered close results to that of CR-AEDCF and SR-AEDCF although it had the highest number of damaged frames at the start, as indicated on the figure itself. It is also shown that SRAEDCF (heavy line) has a very comparable performance to that of CR-AEDCF, including the starting frames. 5. Conclusion In this paper, we have evaluated the performance of the IEEE 802.11e standard. We have also proposed a new adaptive differentiation technique for resetting the value of the contention window after each successful transmission. The proposed adaptive technique takes into account the current level of link utilization when resetting such value. We have performed several simulation scenarios, using NS-2, to evaluate the proposed technique compared to IEEE 802.11e, CR-AEFCF, AMPA, and SSD. During the evaluation, we have focused on three parameters: bit
[1] I. Aad, C. Castelluccia, Differentiation mechanisms for IEEE 802.11, in: Proceedings of IEEE Infocom 2001, Anchorage, Alaska, USA, April 2001, pp. 209–218. [2] M. Bandinelli, F. Chifi, R. Fantacci, D. Tarchi, G. Vannuccini, A link adaptation strategy for QoS support in IEEE 802.11e-based WLANs, in: IEEE Wireless Communications and Networking Conference, New Orleans, USA, March 2005, pp. 120–125. [3] Evalvid Simulation Tool, Available from
H. Artail et al. / Computer Communications 29 (2006) 3789–3803 Hassan Artail worked as a system development supervisor at the Scientific Labs of DaimlerChrysler, Michigan before joining AUB in 2001. At DaimlerChrysler, he worked for 11 years in the field of software and system development for vehicle testing applications, covering the areas of instrument control, computer networking, distributed computing, data acquisition, and data processing. He obtained a B.S. and M.S. in Electrical Engineering from the University of Detroit in 1985 and 1986 respectively and a Ph.D. in Electrical and Computer Engineering in from Wayne State University in 1999. His research is in the areas of Internet and Mobile Computing, Distributed Computing and Systems, and computer networking and security.
Haidar Safa received a B.Sc. in Computer Science in 1991 from Lebanese university, Lebanon, M.Sc. in Computer Science in 1996 from University of Quebec at Montreal (UQAM), and a Ph.D. in Electrical and Computer Engineering in 2001 from Ecole Polytechnique de Montreal. He joined ADC Telecommunications and SS8 Networks in 2001 where he worked on designing and developing networking and system software. In 2003, he joined AUB where he is currently teaching and doing research in Networking. Dr. Safa’s research interests include mobility management in wireless networks, Quality of Service, plus routing and network security.
3803
Joe Naoum-Sawaya finished the Bachelor degree in Computer and Communication Engineering from the American University of Beirut in 2005. He is currently pursuing his graduate studies at the University of Waterloo, in the department of Management Sciences. His research interests involve the application of large-scale linear and integer optimization to problems inspired from communication networks.
Bissan Ghaddar finished the Bachelor degree in Computer and Communication Engineering from the American University of Beirut in 2005. She is currently pursuing her graduate studies at the University of Waterloo, in the department of Management Sciences. Her research interests are applying Semi-definite Optimization to large scale problems with application to communication networks.
Sami Khawam finished the Bachelor degree in Electrical Engineering from the American University of Beirut in 2005. His research interests are in computer networking and application software development.