SCW: Sliding Contention Window For Efficient Service Differentiation in IEEE 802.11 Networks Abdelhamid Nafaa1, Adlen Ksentini2, and Ahmed Mehaoua1 1
University of Versailles – CNRS-PRiSM Lab. 45 avenue des Etats-Unis 78035 Versailles – France {anaf, mea}@prism.uvsq.fr
Abstract—Many works have recently addressed the IEEE 802.11 QoS issues by proposing different monitoring-based Contention Window (CW) differentiation techniques. In network saturation, however, it is difficult to guarantee firm services differentiation while achieving high network exploitation. Particularly, most of existing QoS-capable MAC protocols rely on backoff interval sampled from a dynamic range [0 CWi]. In order to ensure more deterministic service differentiation, we propose a new MAC protocol featuring a sliding contention window (SCW) for each network flow. The different flows are now able to select backoff intervals from different (separated) CW ranges. The SCW dynamically adjusts to changing network conditions, but remains within a per-class predefined range, in order to maintain a separation between different service classes. Simulation results show that compared to the EDCA scheme of 802.11e, SCW consistently excels, in terms of network utilization, strict service separation, and service-level fairness.
I. INTRODUCTION The audio/video streaming application imposes stringent requirements on communication QoS metrics, such as loss rate, delay and jitter. It is important to guarantee these requirements at IEEE 802.11 MAC-level [1], allowing continuity and interaction with higher layer QoS mechanisms (cross-layer QoS) [2]. To tackle the QoS issues at MAC level, the IEEE formed the 802.11 task group [3] to notably design a generic framework for supporting QoS mechanisms. Among the 802.11e TG proposals, EDCA (Enhanced Distributed Channel Access) introduces priority-based CSMA/CA (Collision Sense Medium Access / Collision Avoidance). Substantial amount of works [4],[5],[6],[7] were carried out focusing on enhancing EDCA and developing differentiated services mechanisms. They proposed different priority schemes through differentiating the inter-frame spaces (IFS), minimum/maximum contention windows, Transmission Opportunity (TXOP) durations, and even contention window increasing process. These approaches, however, still provide more probabilistic service assurances rather than deterministic one. Actually, the best-effort (BE) traffics can randomly select a small backoff counter, and then access the medium after wining the contention. This is particularly prevalent in heavy loaded network conditions. As a consequence, the performance of high priority services may undergo severe degradation caused by (1) diminution of the available bandwidth due to frequent medium occupation by the BE traffics, (2) introducing additional delays due to recurrent backoff counter freezing (i.e., high priority flows in deferring state). This phenomenon is obviously caused by the conventional backoff scheme that randomly selects backoff values with a uniform distribution in [0, CWi] regardless the traffic priority i. In addition to the aforementioned inter-class QoS separation issues, the intra-class QoS coordination is determinant to achieve fairness among competing flows. In fact, some fairness problems arise due to resetting CW’s of successful sender while other nodes continue to maintain larger contention windows. This confines the flows in deferring state to seize the channel, resulting in channel
2
LICP, University of Cergy-Pontoise 3 avenue Adolph Chauvin - 95302 Cergy-Pontoise – France
[email protected]
domination by the successful nodes. Although some works have proposed adaptive CW schemes, designed to coordinate MAC parameters between different stations [5][8], they still differentiate based on nodes rather than flows (i.e., achieving rather throughput fairness). If traffic were balanced between nodes, achieving fairness between flows within the same traffic class would require the MAC parameters on different nodes to remain harmonized. However, load is typically unbalanced in real WLAN deployment, with a large variation in traffic class (TC) volume from one node to the next. It is, therefore, crucial to provide intra-TC fairness in terms of perceived QoS. Our main concern, in this paper, is to provide a QoS-capable IEEE 802.11 MAC protocol that achieves high wireless link utilization, while providing strict service differentiation and fairness as well. Thus, we propose for each TC’s flow a Sliding Contention Window (SCW) that allows selecting backoff counter from a sliding/bounded CW range, confining oscillations of both throughput and delays. The different sliding CW’s vary dynamically within a per-class defined backoff range, and may overlap to achieve high bandwidth efficiency in every network configuration. Each flow’s SCW reacts based on the degree to which class-defined QoS metrics (loss rate) are satisfied, providing service-level fairness regardless the flow’s bitrate. The remainder of this paper is organized as follows. The next section will provide background material on the 802.11 MAC, and summarize related works on QoS enhancements. Section III describes the design of the Sliding Contention Window protocol. In Section IV, we compare the performance of SCW to EDCA. Finally, we have drawn several key conclusions from this work; these are stated in Section V.
II. BACKGROUND AND RELATED WORKS ON QOS PROVISIONING FOR IEEE 802.11 A. Legacy IEEE 802.11 MAC Protocols The IEEE 802.11 MAC defines two transmission modes for data packets: the mandatory Distributed Coordination Function (DCF) based on CSMA/CA and, the optional contention-free Point Coordination Function (PCF), where the Access Point controls all transmissions based on a polling mechanism. Basically, PCF was designed to support real time traffic. However, PCF involves several drawbacks, among which excessive data control overhead and scalability limitations. In DCF mode, a station must sense the medium before initiating the transmission of a packet. If the medium is sensed as being idle for a time interval greater than a Distributed Inter Frame Space (DIFS) then the station transmits the packet. Otherwise, the transmission is deferred and a backoff process is started. More specifically, the station computes a random delaying period in the range of 0 to the so-called Contention Window (CW). The backoff time interval is randomly computed in terms of time-slots, and then used to initialize the backoff timer. This timer is decreased only when the medium is
0-7803-8966-2/05/$20.00 (C) 2005 IEEE
idle; it is frozen when another station is detected as transmitting. Each time the medium becomes idle for a period longer than DIFS, the backoff timer is periodically decremented by one for every idle slot-time. As soon as the backoff timer expires, the station accesses the medium. A collision occurs when two or more stations start transmission in the same slot. An acknowledgement is used to notify the sending station that the transmitted frame has been successfully received. If no acknowledgement is received, the station assumes that the frame transmission failed and schedules a retransmission by reentering the backoff process. After a successful or unsuccessful frame transmission, the station executes a new backoff process and so on.
[
]
Tbackoff = Rand( 0,2 j ⋅ CWmin ) ∗ Tslot
DCF adopts an exponential backoff scheme, which refers to the process of increasing the range of backoff delay by doubling the contention window (CW) size with every transmission retry until a predefined maximum (CWmax) is reached; we assume a maximum of j retransmissions.
B. IEEE 802.11 MAC Protocol QoS Issues The need for better access mechanism supporting service differentiation has led Task Group E of IEEE 802.11 to amend the actual IEEE 802.11 standard. The QoS support is realized with the introduction of “traffic class” or Traffic Categories (TCs). Service differentiation under EDCA is achieved by providing different MAC parameters to different priority classes. Thus, each station maintains multiple backoff instances parameterized with TC specific parameters. And each TC[i] flow uses TC-specific parameters such as AIFS[i], CW[i]min, TXOP[i], and Persistence Factor PF[i]. MAC Service Data Units are now delivered through these multiple backoff instances instead of a single instance. As a result, high priority classes will get more transmission time than low priority classes. Our main concern, in this work, is to provide service differentiation at the distributed wireless MAC layer. Although EDCA allows setting different static MAC parameters for each service class, it still does not propose dynamic MAC parameters adaptation according to network condition fluctuations. It is much suitable to provide network-interactive adaptation schemes that perform well in every network configuration. Recently, some works have been done to cover the adaptive issues in IEEE 802.11. In AEDCF (Adaptive EDCF) [4], after each successful transmission, the authors propose to smoothly reset the CW values based on the actual average collision rate. The recently proposed AF-EDCF (Adaptive Fair EDCF) [5] aims at reducing the effect of idle time slots through using a TC[i]-based adaptive backoff threshold taking into account the channel load. This consists in increasing the contention window during deferring periods when the channel is busy, and using an adaptive fast backoff deceasing mechanism when the channel is idle. Generally, the adaptive backoff-based differentiation [4],[5],[6],[7] provides priority medium access for multimedia streams by reducing the probability of collision between frames belonging to different traffic classes. These approaches, however, are not sufficient to provide deterministic QoS guarantees, while achieving fairness and bandwidth efficiency. Managing the contending flows through appropriate CW scheme is a key component to effectively separate between service classes. To overcome this situation, two important issues must be addressed, inter-TC QoS differentiation and intraTC QoS coordination. First, almost all of existing approaches still provide more probabilistic service assurances rather than deterministic one.
Particularly, when the network is heavily loaded, though the best effort flows have a large CW, they still can frequently access the medium by randomly selecting short backoff intervals. Besides limiting the bandwidth available for backlogged multimedia flows, this phenomenon introduces further delays due to backoff freezing (i.e., flows in deferring states). Differentiation between TCs (i.e. inter-TC separation) becomes more “randomness” when the number of active flows increases. As a consequence, the wireless network doesn’t operate at optimal performance by serving the multimedia flows and allowing best effort traffic to use residual resources. Second, in case of collision, the colliding flows enter in retransmission cycle and increase theirs respective contention windows, while other active flows belonging to the same traffic class still maintain small CW range. This leads to unbalanced medium access opportunities among flows of the same TC.
III. SCW: SLIDING CONTENTION WINDOW A. CW Range -based Service Differentiation Most of existing works that addressed QoS issues in IEEE 802.11, use the conventional CW scheme that randomly select the backoff counter in the range [0, W] with uniform distribution. Here, W depends on the flow priority, retransmission stage, or even network load. It is difficult, in such a condition, to provide strict QoS guarantees for high priority flows, while maximizing the network exploitation. To cope with this issue, we control the backoff randomness through providing strict separation between CW ranges of each traffic class i, j, k, etc. Meanwhile, overlapping between CW ranges of the different traffic classes is permitted in order to achieve high medium exploitation in relaxed network conditions (see Figure 1). Within SCW scheme, we associate to each traffic class i a sliding contention window SCW[i] defined by a lower bound (CW[i]LB) and an upper bound (CW[i]UB). These CW bounds delimit the interval from which TC[i]’s flows select a random backoff value. The sliding CW associated to a given TC[i] varies dynamically within a per-TC predefined CW range (CW[i]min, CW[i]max). Figure 1 illustrates the Sliding Contention Window (SCW) operation for 3 traffic classes. All packets from flows belonging to TC[i] use the same set of MAC-level parameters, including CW[i]min, and CW[i]max, as well as AIFS[i]. At a given station, packets belonging to the same TC[i] flow are in fact an aggregation of streams emanating from different applications. Note that the different per-TC CW ranges (CW[i]min, CW[i]max) can be, in practice, highly customized to fit a particular WLAN deployment context by adjusting the tradeoff between strict QoS separation and high network utilization. CW[ i ]LB
CW[ i ]UB
....
SCW[i] size
CW[ i ]min
CW[ i ]UB+m∗SF[ i ]
CW[ i ] LB +m∗SF[ i ]
....
CW[i]max
Sliding CW scheme common to TC[i]'s flows
CW[ j ]
LB
CW[ j ]
UB
.... CW[ j ]min
....
CW[ j ] LB +m∗SF[ j ]
Sliding CW scheme common to TC[j]'s flows
CW[ k]
LB
CW[ k ] SCW[k] size
CW[ k ]min
CW[ j ]UB +m∗SF[ j ]
CW[ j ]max
UB
....
....
CW[ k]LB +m∗SF[ k]
Sliding CW scheme common to TC[k]'s flows
CW[ k]UB+m∗SF[ k]
CW[ k]min
Slot time [0 - 1024] Aggregated Contention Window CW [0 - CWmax]
Figure 1: Contention window sliding scheme for 3 different TCs.
0-7803-8966-2/05/$20.00 (C) 2005 IEEE
Rather than using persistence factors, a TC-dependent sliding factor, SF[i], is used. If PFs were used, the contention window would have a dimensioning problem, as its size changed with each window adjustment. Instead, SCW uses a Linear-increase Linear-decrease (LILD) model to adjust the SCW range. Compared to the conventional DCF, each time a flow in TC[i] experiences a high loss rate, SCW[i]’s range is increased by a SF[i] step, until the upper bound reaches the maximum window value. When losses are low and packets are transmitted successfully, rather than resetting the contention window, SCW[i]’s range is decreased with the same SF[i] step, until the lower bound reaches CW[i]min. The procedures for both decreasing and increasing the SCW range are as follows: 1)
SCW decreasing procedure: IF (oldCW[i]LB – SF[i] ≥ CW[i]min) { newCW[i]LB = oldCW[i]LB – SF[i] newCW[i]UB = oldCW[i]UB – SF[i] } ELSE { newCW[i]LB = CW[i]min newCW[i]UB = CW[i]min + size(SCW[i]) }
2)
SCW increasing procedure: IF (oldCW[i]UB + SF[i] ≤ CW[i]max) { newCW[i]LB = oldCW[i]LB + SF[i] newCW[i]UB = oldCW[i]UB + SF[i] } ELSE { newCW[i]LB = CW[i]max – size(SCW[i]) newCW[i]UB = CW[i]max }
At starting phase (t0), each traffic class i initializes its SCW[i] lower and upper bounds as follows: CW[i]LB = CW[i]min, and CW[i]UB = CW[i]min + 2·SF[i] Note that SCW[i]’s sliding granularity is determined exclusively by the sliding factor associated to the class i (TC[i]). Therefore, SF[i] basically represents the “stride” for adjusting the contention window up or down. For our purposes, the higher TC[i]’s priority, the smaller the sliding factor SF[i]. Lower priorities, such as best-effort traffic, get larger sliding factors. This ensures that fine grain adaptation for high priority traffic occurs, while low priority traffic is quickly limited when reacting to congested conditions. Since multimedia applications usually sustain losses by application-level error control mechanisms [9], we trade lower delay for slightly increased loss rate. By decreasing CW[i]max, the maximum backoff time can be limited during congestion, which significantly increases the collision probability, while confining the medium access deferring time.
B. Sliding Contention Window Fairness An important issue to tackle is the fairness among traffic belonging to the same priority class. This is, in part, insured by harmonizing the MAC parameters. However, when two flows of certain traffic class TC[i] (e.g., high priority) collides both of them increase their respective CW ranges, whereas the other flows belonging to TC[i] still maintain a lower CW ranges. As a consequence, the colliding flows are somehow penalized compared to other active flows belonging to the same priority class (TC[i]). The disparity between the QoS (throughput, loss rate, delays) received by different receivers becomes clearly noticeable after a short running period. To overcome this problem, previous
works have proposed to monitor the overall network conditions, and re-adjust the CW value of each traffic class accordingly [4][6]. However, estimates of metrics such as collision rate can only be made on the aggregate network conditions, concerning all traffic classes. The response to this signal will affect all traffic classes, which may result in the degradation of perceived QoS, particularly to higher bit-rate flows. Additionally, the different CWs ranges will tend to converge to the same area, resulting in a loss of differentiation and an increase in collisions and delays. Several approaches (see [8], and references therein), are based on Multiplicative-increase Linear-decrease (MILD) schemes. These protocols address the fairness problem by including the current CW size in the MAC header of transmitted packets. This allows overhearing nodes to compare this value with their own, and adjust if necessary. However, as well as increasing the overhead required for the header, this approach suffers from conformance and power issues. It is important to remember that we wish to provide constant QoS for each traffic class, rather than equality between different wireless stations. If traffic were balanced between nodes, achieving fairness between flows within the same traffic class would require the sliding contention windows on different nodes to remain harmonized. However, as traffic is frequently unbalanced, QoS metric thresholds combined with SCW smoothing rules are used to control the CW’s sliding process, in order to ensure that all traffic classes receive the required QoS. Thus, we propose to guarantee the same QoS metrics (e.g. loss rate, mean delay, mean jitter) for all flows belonging to the same TC. The objective is to maintain a sustained application-level perceived QoS; this is an imperative in most of existing and future operated wireless networks [2]. Consider the loss rate Lr[i] of a high priority flow, as it is perceived by the application. This loss rate account for the drop rate measured at LLC/MAC queue and the frames discarded after several failed retransmissions. If Lr[i] is too low than αi (a threshold value for the maximum tolerated loss rate for TC[i]), the smoothing rules cause the SCW to be linearly increased, in order to give more access opportunities to lower priority flows. This ensures that the lower priority TCs still receive adequate QoS, as well as improving the utilization of the medium. On the other hand, if the loss rate is too high, the SCW range can be decreased, giving the TC higher priority and reducing the loss rate. The above introduced scheme is implemented in EDCA-like architecture. Note that a single multimedia stream (e.g., H.264 video) is usually fragmented and mapped into different QoS classes [10]. Basically, the queue drop rate (Lr[i]) significantly influences the service-level perceived QoS. We rely on the Lr[i] metric to dynamically adjust the SCW[i] ranges of high priority flows. In fact, we use Lr[i] to independently vary the SCW[i] range at each station, and thus achieving long-term fairness among TC[i] flows of different stations. This way, the penalized flows of a certain traffic class i (i.e., stations having Lr[i] ≥αi) will be able to gain further transmission opportunities and consequently decrease their queue loss rates (see the sliding algorithm bellow). Here, αi represents the maximum tolerated loss rate by TC[i]. On the other hand, best-effort traffic does not require any QoS metric thresholds, the contention window must be adjusted slightly differently. We use the instantaneous network load B(T) to adjust the SCW range. B(T) is the fraction of slots that the medium was observed to be busy out of the previous T slots. This includes all slots where a transmission was successfully completed, or a collision occurred. If the network load drops below the threshold B(T)Threshold, then the SCW range for best effort traffic is decreased.
0-7803-8966-2/05/$20.00 (C) 2005 IEEE
If the load exceeds a throughput saturation threshold B(T)Saturation, then the SCW range is increased. Based on extensive simulations, we have found that 0.7 and 0.9 are appropriate values for B(T)Threshold and B(T)Saturation respectively. The entire sliding algorithm is shown below:
SLIDING ALGORITHM Sliding for high priority flows i (e.g., EF, AF11, AF12, etc): IF (Lr[i]≥α i) then Decrease(SCW[i]) Else IF (α i/2
