A new framework for QoS provisioning in WLANS using p-persistent 802.11 MAC

Computer Communications 31 (2008) 4035–4048

Contents lists available at ScienceDirect

Computer Communications journal homepage: www.elsevier.com/locate/comcom

A new framework for QoS provisioning in WLANS using p-persistent 802.11 MAC Kiran Anna *, Mostafa Bassiouni 1 School of Electrical Engineering and Computer Science, University of Central Florida, Orlando, FL 32816, USA

a r t i c l e

i n f o

Article history: Received 5 February 2008 Received in revised form 6 August 2008 Accepted 13 August 2008 Available online 29 August 2008 Keywords: p-persistent 802.11 QoS Service differentiation Admission control Scheduler

a b s t r a c t In this paper, we propose a new framework which consists of three components viz., a p-persistent 802.11-based core MAC layer which provides optimum channel utilization, a scheduler that provides service differentiation between different classes of traffic in the contention phase of 802.11 and an admission control which provides statistical QoS guarantees. The current 802.11e EDCF protocol uses the same set of parameters viz., CWmin, CWmax, CW, AIFS, and PF for both channel access and service differentiation. We show that using these parameters to achieve a specific desired throughput differentiation is a very cumbersome task. We present detailed performance results of our framework using the NS2 simulator to demonstrate the effectiveness of the scheduler and the core MAC layer. Compared to 802.11 adaptive EDCF, our scheduler consistently achieves the desired throughput differentiation and easy tuning. The core MAC layer achieves better delays in terms of channel access, average packet service time and one hop delay and also achieves higher system throughput than the AEDCF for a given service differentiation ratio. We also develop an analytical model for the p-persistent 802.11 in the core MAC layer under unsaturated load conditions and present closed forms for the average packet service time and one hop delay. We consider both the ‘‘no retry limit” and ‘‘finite retry limit” cases and validate the analytical model by extensive simulation results. Ó 2008 Elsevier B.V. All rights reserved.

1. Introduction In the recent past, there is a growing deployment of WLANs for its flexibility and non-wired capability especially, in hotspot locations like bookstores, airports, shopping malls, etc. With the increasing interest of the integrated multimedia services over high-speed networks including wireless LANs, it is natural to expect that WLANs will support real-time applications with quality of service (QoS) guarantees at the same level as its wired counter part. IEEE 802.11 employs a carrier sense multiple accesses with collision avoidance (CSMA/CA) mechanism called distributed coordinated function (DCF) as a basic access method and is designed for best effort services only. The lack of built-in mechanism to support real-time services makes it very difficult to provide QoS guarantees for the throughput-sensitive and delay sensitive multimedia applications. Therefore, modification of the existing 802.11 standards is necessary. The new 802.11e standard [4] provides an enhanced DCF (EDCF) protocol. In IEEE 802.11e, a hybrid coordination function (HCF) provides QoS guarantees through a centralized polling mechanism. However, HCF is far from deployment in the real world. Although, * Corresponding author. Tel.: +1 407 928 7169/+1 407 823 3364. E-mail addresses: [email protected] (K. Anna), [email protected] (M. Bassiouni). 1 Tel.: +1 407 823 2837. 0140-3664/$ - see front matter Ó 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.comcom.2008.08.013

IEEE 802.11e is being proposed as the upcoming standard for the enhancement of service differentiation, QoS guarantee in EDCF is still a challenging problem and needs further study [18]. In the current EDCF implementation, there are four access categories (ACs) which are mapped to different classes of traffic. Each AC uses a set of parameters for channel access viz., CWmin (minimum contention window), CWmax (maximum contention window), AIFS (arbitration inter-frame space), and PF (persistent factor). As a result, the ACs’ compete among themselves for the channel within a node while this node competes for the channel with the other nodes in the WLAN. The internal competition leads to virtual collisions and reduced system capacity. The service differentiation for the ACs is provided by prioritizing the values of the parameters in the ACs. The service differentiation and channel access are two different components supported by the same set of parameters. Therefore, a desired performance on one component may lead to a negative impact on the other component. The QoS support in EDCF is provided only in terms of service differentiation. In order to support QoS guarantees, an admission control is essential. As wireless LANs may be considered as just another technology in the communications path, it is desirable that the architecture for QoS follow the same principles in the wireless network as in the wire-line Internet, assuring compatibility among the wireless and the wire-line parts [15]. In this paper, we propose new framework to provide statistical QoS guarantees, service differentiation and efficient channel access

4036

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048

that leads to improved system throughput. Our framework consists of three components viz., a scheduler, a core MAC layer, and an admission control. The scheduler provides service differentiation according to the weights assigned to each AC. The core MAC concentrates on just the channel access to improve efficiency. The admission control provides statistical QoS guarantees. Following is a summary of our contributions in this paper: – Separating the channel access and service differentiation functionalities into two different components. – Simplifying the 802.11 channel access mechanism to a single parameter p by using p-persistent 802.11 at the core MAC layer. – Presenting comparison of the performance of our scheme in terms of service differentiation and channel access with adaptive EDCF (AEDCF). – Presenting and validating an analytical model for the p-persistent 802.11 under unsaturated load conditions with closed forms for the average packet service time, and average one hop delay. The rest of the paper is organized as follows. In Section 2, we discuss previously proposed mechanisms and the related work. We review the basics of the IEEE 802.11 standard and introduce p-persistent 802.11 in Section 3. We discuss the problem, motivation and our framework in detail in Section 4. In Section 5, we present the performance evaluation of our framework with the AEDCF scheme. Finally, we conclude our paper in Section 6.

throughput relies on the value of transmission probability p and the number of active stations [6]. Also, they suggest that the IEEE 802.11 protocol with its current settings can hardly achieve the theoretical capacity bound. They extended their work by dynamically tuning the value of p so as to achieve the capacity bound under varying network conditions [7]. However, it is very difficult to model the average packet service time and one hop delays in ppersistent 802.11 using Cali’s model as it is based on the number of idle slots, and the number of collisions between two successful transmissions at a system level. There have also been several efforts to model the average service time for the legacy 802.11 in both the saturated [22–26] and unsaturated load conditions [27–30], respectively. Chatzimisios et al. and Carvalho et al. presented analytical models for service time in legacy 802.11 MAC in [23,24] and [25], respectively. Chatzimisios did not consider the delay due to the dropped packet after the retry limit stage in the calculation of the service time. Carvalho’s analysis did not consider finite retry limits. In [27], Tickoo modeled each node as a G/G/1 queue and derived a service time distribution. Barowski [29] extended Bianchi’s model [8] for saturated conditions to unsaturated conditions using M/M/1/K analysis. However, the model could not describe the behavior accurately when the system is slightly below saturation. Most of the above models describe the legacy 802.11 DCF with binary exponential back-off. 3. The IEEE 802.11 MAC 3.1. The standard 802.11 MAC protocol

2. Related work IEEE 802.11 WLANs have received considerable attention in recent years [1–4]. Various distributed fair scheduling algorithms are proposed in the literature to provide fairness of the bandwidth and service differentiation by assigning different weights to flows based on their priority [15–17]. Ge [12] extended Cali’s work to multiple queues to provide throughput service differentiation while achieving the maximum system capacity. But, the scheme works only in infrastructure mode. Yang [13] proposed QPART (QoS protocol for ad hoc real-time traffic) which provides QoS guarantees by adapting the contention window sizes at the MAC layer. The architecture consists of a QoS aware scheduler and QoS manager above a core MAC layer. The scheduler realizes the functionality of QoS aware scheduling while the QoS manager implements admission control and conflict resolution. The Core MAC layer uses 802.11 EDCF and as a result, it does not provide maximum system capacity. Its architecture and implementation is completely different from ours which is based on ppersistent 802.11. Liu [14] proposed a MAC architecture where different adaptors can be added above the core MAC layer to provide service differentiation. Each adaptor works on adjusting only one of the parameters viz., CW, CWmin, CWmax, etc. But, the adaptors are very basic and do not tune based on the network conditions. There are several research efforts in the literature that model the performance of legacy 802.11 DCF [6–12]. Bianchi [8] models the behavior of a station with a Markov chain model and analyzes the saturation throughput of 802.11 MAC protocol for the no retry limit case (no bound on the number of frame retransmissions) in ideal channel conditions and finite number of terminals. Wu et al. extend Bianchi’s model to include finite packet retry limits as defined in the IEEE 802.11 standards and therefore, present a more accurate model [11]. Cali et al. analyze the performance of the legacy IEEE 802.11 DCF protocol through the p-persistent version of IEEE 802.11 protocol. Based on the analytical model, they observe that the system

IEEE 802.11 has become the standard medium access control (MAC) and physical layer (PHY) specification for the WLANS [11]. The 802.11 MAC works in two modes: (i) the DCF and (ii) the point coordination function (PCF). The DCF, also known as the basic access method, is a carrier sense multiple access with collision avoidance (CSMA/CA) protocol. On the other hand, the optional PCF is based on a polling mechanism and uses a centralized point coordinator to grant the channel access to each node in the WLAN. The PCF was basically designed as an extension to DCF to provide QoS guarantees in the WLANs. The DCF adopts an exponential back-off scheme. The back-off time is uniformly chosen in the range [0, CW], where CW is called contention window. In the first transmission attempt, the CW is initialized to CWmin and is doubled up to CWmax for each subsequent unsuccessful transmission attempt. The back-off timer is decreased as long as the channel is sensed idle, stopped when a transmission is in progress, and reactivated when the channel is sensed idle again for more than distributed inter-frame space (DIFS). As soon as the back-off timer expires, the node starts to transmit at the beginning of the next slot time. A positive acknowledgement (ACK) is transmitted by the receiver node to signal the successful frame reception. The receiver sends an ACK frame after a time period equal to short inter-frame space (SIFS), which is less than DIFS. If the transmitting station does not receive an ACK within a specified ACK timeout, the data frame is presumed to be lost, and after extended inter-frame space (EIFS), a retransmission is scheduled. According to the standard, a maximum number of retransmissions (default = 7) are allowed before the frame is dropped. The two-way handshaking technique outlined above for the frame transmission is called the basic access mechanism. In addition, DCF defines an optional four-way handshaking technique called request to send/clear to send mechanism (RTS/CTS) to solve the hidden terminal problem. In this mechanism, a node transmits a special short frame called RTS instead of the frame. When the

4037

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048

receiving node detects an RTS frame, it responds with a CTS frame after a SIFS. The transmitting station is allowed to transmit its frame only if the CTS frame is correctly received. The frames RTS and CTS carry the information about the length of the frame to be transmitted. This information can be read by any listening station, which is then able to perform virtual sensing through its network allocation vector (NAV) that contains information about the amount of time the channel will remain busy. The RTS/CTS mechanism is very effective in terms of system performance, especially when large frames are considered. 3.2. p-persistent 802.11 MAC protocol In the p-persistent 802.11 MAC protocol, a station at the beginning of each empty slot transmits with a probability p or defers with a probability 1  p. It is verified that it can approximate the standard 802.11 at least from the capacity point of view when the average back-off interval is the same as in the standard 802.11 [5,6]. The relation between p and CW is given in [7] as

p¼

2 : CW þ 1

ð1Þ

In the p-persistent 802.11, the back-off interval is sampled from a geometric distribution with parameter p as opposed to the binary exponential back-off used in the standard DCF version. It uses one parameter p in the back-off interval generation, whereas the standard 802.11e uses four parameters CWmin, CWmax, AIFS and PF. Also, it is well suited for analytical studies due to the memory less property of the back-off algorithm [6]. We shall refer to the p-persistent IEEE 802.11 protocol as PDCF.

4. Framework for QoS provisioning in WLANS The standard 802.11 standard was originally designed to support best effort traffic with only one queue or access category (AC). However, with the widespread use of WLANs in various hotspot locations, the demand to support other types of traffic especially, real-time traffic has also increased. Therefore, the IEEE 802.11e is the draft of amendments which introduces the QoS to the IEEE 802.11 standard and provides support for multiple classes of traffic [4]. The 802.11e provides HCF, which combines two access modes: EDCF and HCF controlled channel access (HCCA). The support for one AC in 802.11 is extended to four ACs, which are mapped to different classes of traffic. In the EDCF, each AC uses a separate set of parameters viz., CWmin, CWmax, AIFS and PF for channel access. Therefore, the four ACs within each node compete for the channel independently. As a result, there will be collisions between packets from different nodes as well as collisions between packets from different ACs within the same node. The internal collisions within a node are called virtual collisions. The virtually collided ACs will double their CW just as in the regular collisions between nodes and thereby leading to reduced channel utilization. This drawback is overcome by a virtual collision handler which gives channel access to a higher priority AC. The lower priority AC will double its CW. There are two drawbacks with this solution. When there is no real collision for the higher priority AC then the lower priority AC is unnecessarily penalized. Also, it is robbed of fairness in channel access when compared to the same priority ACs from other nodes. Second, the EDCF provides QoS in terms of service differentiation by prioritizing the channel access among the different classes of traffic within a node. The channel access is prioritized by using different AIFS, CWmin, CWmax, and PF values for each class of traffic in the MAC layer. As a result, the two separate functional compo-

nents, the QoS service differentiation and the channel access mechanism, are combined into one functional component by the same set of parameters increasing the complexity of tuning the 802.11e MAC. A desired performance on one component may lead to a negative impact on the other component. In IEEE 802.11e, the HCF provides QoS guarantees through a centralized polling mechanism. However, HCF is far from deployment in the real world. Although, IEEE 802.11e is being proposed as the upcoming standard for the enhancement of service differentiation, QoS guarantee in EDCF is still a challenging problem and needs further study [18]. In order to support QoS guarantees, an admission control is essential. As wireless LANs may be considered as just another technology in the communications path, it is desirable that the architecture for QoS follow the same principles in the wireless network as in the wire-line Internet, assuring compatibility among the wireless and the wire-line parts [15]. Most of solutions in the literature add some extensions or modify the components to address the problems discussed above. These solutions either increase the complexity or provide band aid solutions. We plan to solve this problem by separating the different functionalities into separate components. In an earlier conference paper [20], we proposed separation of these two functionalities by means of a scheduler-based architecture. We refine this idea in this paper. We propose a new framework to provide efficient channel access that leads to improved system throughput, service differentiation, and statistical QoS guarantees. Our framework consists of three components viz., a core MAC layer, a scheduler, and an admission control. The core MAC layer, implemented using p-persistent 802.11, concentrates on just the channel access to improve channel utilization. The scheduler provides service differentiation according to the weights assigned to each AC. The admission control provides statistical QoS guarantees in collaboration with the Core MAC layer and the scheduler. Fig. 1 shows the block diagram of our framework. One additional feature with this framework is that each component can be tuned and improved independently of the other. 4.1. The core MAC layer In the core MAC layer, there are no separate channel access parameters for different traffic classes as in the standard EDCF. It has only one parameter p, called transmission probability, which

Application Layer

Link Layer

Admission Control

Queues

Scheduler

Core P-Persistent 802.11 MAC Layer Fig. 1. New framework based 802.11 MAC.

4038

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048

the p-persistent 802.11 in the core MAC layer uses for the channel access. The core MAC layer receives only one packet from the scheduler from any of the four ACs for the channel access. Thus, it can just concentrate on optimizing the channel access irrespective of the traffic class and thereby improves the overall channel utilization. However, in the EDCF, each traffic class has different channel access time leading to reduced channel utilization. The core MAC layer plays an important role in the implementation of the admission control by providing information about channel access delays, network and load conditions. Since it is difficult to get this information in the distributed scenario, a complete understanding of the behavior of the p-persistent 802.11 used in the core MAC layer helps the implementation of an admission control. So, in the next sections, we define the calculation of the optimum p to achieve efficient channel access and also develop an analytical model to derive closed forms for various packet delays. 4.1.1. Optimum P for efficient channel access In order to achieve a maximum possible system throughput, the core MAC layer needs to use an optimum value for the channel access parameter p for a given network and load conditions. In [8], Bianchi derived a closed form for Popt under saturated conditions, which is given below:

Popt ¼

1 pffiffiffiffiffiffiffiffiffiffiffi N  T C =2

ð2Þ

where N = number of nodes and T C ¼ T C =r  r is the slot time and TC is the average time the channel is sensed busy due to collision. For the basic and RTS/CTS access methods, the TC is given by

T C;basic ¼ H þ L þ EIFS  DIFS T C;rts=cts ¼ RTS þ EIFS  DIFS

ð3Þ

where H (=PHYhdr + MAChdr) is the time to transmit a packet header, RTS is the time to transmit RTS frame and L is the packet transmission time. Our core MAC layer uses p calculated from (2) dynamically based on the network and load conditions at regular intervals. 4.1.2. Analytical model Our core MAC layer is fundamentally based on the p-persistent 802.11 MAC protocol. In an earlier conference paper [21], we developed an analytical model for p-persistent 802.11 under saturated conditions. In this section, we extend this model to the case of unsaturated conditions. We assume that there are n contending stations in our network. A station always resides in an empty state (E) when there are no packets to transmit. In the presence of at least one packet, it resides in one of following three states: at a non-empty state (NE), where a station contends for the channel; it moves to a transmission state (X) or an idle state (I) based on whether a packet is transmitted or not in a given time slot. After a successful transmission a station moves back to either E or NE depending on the presence of a packet. The NE state is a transient state; visiting the NE state means processing a new packet. Let q0 be the probability that a station has an empty queue. Let s(t) be the stochastic process that represents a station’s state at a time slot t. When a station is competing for the channel, the time interval between its two consecutive states is referred to as the slot transition time and is variable as it observes one of the three mutually exclusive events namely, a successful transmission, a collision or an idle slot (no transmission) from the rest of stations in the network. The idle slot is equal to the constant system parameter r based on the PHY layer used. We develop our model based on the notation given in [8] and use the same assumptions as in [8,21]. Let K denote the maximum queue size in terms of the number of packets submitted to the MAC layer for transmission and k be the data load rate at each node. We use M/M/1/K queue for the analysis.

4.1.2.1. State model for no retry limit. When there is no limit on the number of retransmissions, a packet at the head of the queue remains there until it is successfully transmitted. The state model in Fig. 2 shows the behavior of a station for the case of no retry limit. As described above, the model remains in one of the four states. A station transmits in a physical slot time with probability p and remains idle with probability q. It will succeed in its transmission with probability Vs and fail with probability Vc. The relationship between them is given below:

pþq ¼1

ð4Þ

Vs þ Vc ¼ 1

ð5Þ

Q If the notation fA j Bg denotes the probability that a station moves from state B to state A, the transition probabilities of the above station model are the following:  The station remains in the empty state E if the empty queue has no packet arrival in any slot time or the queue becomes empty after the successful transmission of the last packet in the queue.

PfE=Eg ¼ q0 PfE=Xg ¼ q0  V S  The station is in non-empty state NE when there is a new arrival or the queue is non-empty after a successful transmission.

PfNE=Eg ¼ 1  q0 PfNE=Xg ¼ ð1  q0 Þ  V S  The station is in idle state I when it has a packet but does not transmit it according to the p-persistent protocol.

PfI=NEg ¼ q PfI=Ig ¼ q PfI=Xg ¼ q  V C Notice that q is the probability that the station has a packet but does not transmit it in a given time slot whereas q0 is the probability that the station does not have any packet in its queue.  The station is in transmission state when it transmits a packet. It can transmit a packet either from NE, I or X. The transmission from state X happens when the previous transmission ended in a collision.

PfX=NEg ¼ p PfX=Ig ¼ p PfX=Xg ¼ p  V C

4.1.2.2. State model for finite retry limit. In this section we consider finite retransmissions. Let m  1 be the maximum allowed number of retransmission attempts and r(t) be the stochastic process representing the (re)transmission stage (1, . . ., m) of a station at time t, where stage 1 is the initial transmission and stages 2 through m

1-q0

q0

NE

E (1-q0)Vs p

q0Vs

q

p

pVc X

q I

qVc Fig. 2. State model for no retry limit.

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048

are retransmissions. We use the same value p in all the stages. Therefore, we have

pi ¼ p 1 6 i 6 m

ð6Þ

The empty and non-empty states are the same as in the no retry limit model. As in [8,21], the key assumption in our model is that in each transmission attempt, a packet collides with constant and independent probability Vc irrespective of the number of the retransmission attempts. The probability Vc is a conditional probability because it is the probability of collision seen by a station transmitting a packet, i.e., it is conditioned on the existence of a station(s) transmitting a packet in a given time slot. Thus, the two-dimensional process {r(t), s(t)} is a discrete-time state model, which is shown in Fig. 3. On careful observation, the finite retry limit model is the same as the no retry limit model except that infinite stages is replaced by finite stages. The transition probabilities for the finite retry limit are the following:  The station remains in the empty state E if the empty queue has no packet arrival in any slot time or the queue becomes empty after the successful transmission of the last packet in the queue at any stage. In stage m, the packet is dropped (removed from the queue) irrespective of success or collision.

PfE=Eg ¼ q0 PfE=X i g ¼ q0  V S 1 6 i 6 m  1 PfE=X m g ¼ q0  The station is in non-empty state NE when there is a new arrival in an empty queue or the queue remains non-empty after a successful transmission of a packet in any stage. As mentioned above, in stage m, the packet is removed from the queue irrespective of its transmission status. So, if the queue remains non-empty after this, then it moves to the state NE.

PfNE=Eg ¼ 1  q0 PfNE=X i g ¼ ð1  q0 Þ  V S 1 6 i 6 m  1 PfNE=X m g ¼ ð1  q0 Þ

4039

 The station is in idle state Ii when it does not transmit a packet from the states NE,Xi1 or Ii.

PfI1 =NEg ¼ q PfIi =Ii g ¼ q 1 6 i 6 m PfIi =X i1 g ¼ q  V C 2 6 i 6 m  The station is in transmission state Xi when a packet is transmitted in stage ‘i’. The transition from state Xi1 to Xi occurs when the transmission ends in a collision. The transition from NE to X1 occurs when a station transmits a packet

PfX 1 =NEg ¼ p PfX i =Ii g ¼ p 1 6 i 6 m PfX i =X i1 g ¼ p  V C 1 6 i 6 m  1

4.1.2.3. Service time in p-persistent 802.11. The service time of a packet is the time interval between the time the packet is at the head of its MAC queue ready for transmission and the time when an acknowledgement is received or the packet is discarded after the retry limit for collisions is exhausted. However, in the no retry limit case, the packet is never dropped irrespective of the number of collisions. We use ‘‘backoff step” and ‘‘state transition step” to mean the same thing and also assume that events in successive state transition steps are independent. So, when a collision occurs in a state transition step, the colliding nodes resolve the collision in the same transition step and are ready for transmission in the next transition step thus, avoiding any dependencies on the number of colliding nodes in the previous transition steps. Since each node is empty with probability q0, there are only a fraction of nodes which contend for the channel. Therefore for a given n nodes there are only (1  q0).n nodes that are active. In saturated conditions, the queue is never empty and so q0 = 0. (a) No retry limit Let r be the average number of state transition steps a node observes before it transmits a packet. Since a node transmits a packet with probability p, the average number of state transition (back-off) steps r before the node transmits a packet is given by

q0

r¼

1-q0 E

NE p

q0

q

1-q0 p

Vs X1

I1

q

qVc

pVc

X2

ð7Þ

At each transition step, a station observes one of the three events on the channel viz., a successful transmission Es, a collision Ec, or an idle slot Ei. Let gs = P{Es}, gc = P{Ec} and gi = P{Ei} be the probabilities for the above three events. Since these three events are independent and mutually exclusive at each state transition step, the probability that there are ri ‘‘idle slots”, rc ‘‘collision slots”, and rs ‘‘successful slots” in r slots is given by,

g i ¼ ð1  pÞð1q0 Þðn1Þ

p

Vs

1p p

I2

q

g s ¼ ð1  q0 Þ  ðn  1Þ  p  ð1  pÞð1q0 Þðn2Þ

ð8Þ

g c ¼ 1  fð1  pÞð1q0 Þðn2Þ gfðn  1Þð1  q0 Þp þ ð1  pÞð1q0 Þ g Then, ri = r.gi, rs = r.gs and rc = r.gc. Therefore, the total average back-off time or state transition time (TB) spent before transmitting a packet is,

pVc

T B ¼ r  r i þ T c  r c þ T s  r s ¼ r  ½r  g i þ T c  g c þ T s  g s 

p Xm Fig. 3. State model for finite retry limit.

Im

q

ð9Þ

Tc and Ts are the average time for a collision and successful transmission respectively. The symbol r denotes the PHY layer slot time parameter. After TB, a station will transmit a packet and will

4040

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048

succeed with probability Vs or fail with probability Vc = 1  Vs. The packet transmission will be successful only when none of the other nodes with non-empty queues transmit packets at the same time. Thus

V s ¼ ð1  pÞð1q0 Þðn1Þ

ð10Þ

Vc ¼ 1  Vs

ð11Þ

Since p is constant in all transmission stages, Vs is same in all transmission attempts and therefore the average number of transmission attempts ending with a collision sc before a packet is successfully sent is given by

sc ¼

Vc Vs

ð12Þ

Therefore the average service time is equal to

T ¼ sc  ðT B þ T c Þ þ ðT B þ T s Þ



 ð1  qÞ  qk ð1  qKþ1 Þ

06k6K

¼0

kavg ¼ ð1  qK Þ  k

Q avg ¼ ½q=ð1  qÞ  fðK þ 1ÞqKþ1 =ð1  qKþ1 Þg

ð22Þ

Thus, the average one hop delay ðDÞ for the unsaturated conditions is given by

Q avg kavg

ð23Þ

Under saturated conditions, the queue is always full to the size K. Therefore the one hop delay ðDÞ is given by

16k6m

PfXDg ¼ V m c

ð14Þ ð15Þ

D¼K T

q0 ¼

ð1  qÞ ð1  qKþ1 Þ

Calculate T init from ð19Þ using q0 ¼ 0 Do

T B ðkÞ ¼ k  T B þ ðk  1Þ  T c

Calculate T from ð19Þ using the new value q0

ð16Þ

Let Tk be the average service time corresponding to the event XSk and Tw be the average wasted time corresponding to the event XD.

16k6m

T w ¼ T B ðmÞ þ T c

ð17Þ ð18Þ

l ¼ 1=T T diff ¼ T  T init T diff ¼ jT diff j T init ¼ T Until ðT diff < 0:00000001Þ

l ¼ 1=T q ¼ k=l

Therefore, the average service time is given by, m X

f½PfXSk g  T k  þ ½PfXDg  T w g

D ¼ Q avg =kavg

k¼1

¼ ð1  V m c Þ  fsc  ðT B þ T c Þ þ ðT B þ T s Þg

ð25Þ

where q = k/l and K is the maximum queue size. Thus, for a given k, we observe that q0 and T are dependent on each other. To overcome the dependency, we use an iterative method to find T and thereby the one hop delay (D). We observed that our iterative method converges in three to four iterations.

l ¼ 1=T init q ¼ k=l q0 ¼ ð1  qÞ=ð1  qKþ1 Þ

16k6m

ð24Þ

However, from Section 4.1.2.3, we know that T is dependent on p, n, and q0. Also, for a given k; l ¼ 1=T and K, the value of q0 is given by,

Let TB(k) be the total time spent before transmitting a packet successfully in the stage k. Therefore TB(k) is given by

T¼

ð21Þ

The average queue size is given by,

ð13Þ

Let m  1 be the upper bound on the number of the retransmissions and the node behavior in each stage is same as that of the single stage in the ‘‘no retry limit” case. Also, the events that occur in a node at the kth transmission stage are independent of the events that occur in the (k  1)th stage. Since p remains constant in all stages, the average number of transition slots r before a node transmits a packet is the same as that in the ‘‘no retry limit” case in each of the stages 1 to m. The probability of success that a packet experiences when it is transmitted at the end of the kth back-off stage is Vs = 1  Vc. Let XSk be the event that the packet is successfully transmitted at the end of kth stage and XD be the event that the packet is discarded at the end of the mth stage.

T k ¼ T B ðkÞ þ T s

ð20Þ

Otherwise

where, q = k/l. Once the queue is full, the packets arriving at the queue are dropped and as a result the input load rate is not the effective load anymore. Therefore, the average load kavg is given by

D¼

(b) Finite retry limit

PfXSk g ¼ V ck1 V s

qk ¼

ð19Þ

4.1.2.4. One hop delay in p-persistent 802.11. A one hop delay in a single hop WLANS is defined as the time interval between the time at which a packet is placed in the queue and the time at which a corresponding ACK is received which is also end-toend delay in single hop WLANS. With increasing emphasis on QoS in WLANS, a thorough study of one hop delay in WLANS helps us identify its limitations and helps in designing an efficient admission control. The service rate l is the inverse of the average service time T. In [31], the probability to find k packets in a queue with a maximum size of K in M/M/1/K queue is given by

ð26Þ

4.2. The scheduler The scheduler is built above the core MAC layer and its basic functionality is to provide service differentiation to the different classes of traffic. The scheduler selects the next packet from one of the ACs after the core MAC layer sends a packet successfully and requests for another packet or at the arrival of a new packet when all the ACs are empty. In our scheme, the scheduler is provided with weights for each AC. It sends the next packet from an AC based on the weights assigned to it. Thus, it completely eliminates the virtual collisions. It also provides a much easier and more accurate control in providing the QoS service differentiation for all classes within a node. In

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048

addition, the weights in the scheduler can also be changed by an access point in an infrastructure based hotspot to optimize the revenue for the service providers. Since the same set of parameters is used for both the channel access and the service differentiation in the EDCF, there is no easy way in to enforce a certain differentiation ratio among different traffic classes in a node. In this paper, we used a packet-based weighted fair queuing algorithm in our scheduler to provide throughput service differentiation at the node level. When one of the queues is empty, it sends the packets available in the remaining queues while keeping track of the deficit in the empty queue. It tries to compensate the deficit once the empty queue is filled with packets. We also like to point out that throughput service differentiation means bandwidth reservation per class within a node. The scheduler divides the bandwidth among the classes based on the differentiation parameters which means it provides bandwidth reservation. For example, a service differentiation of 3:2:1 provides a bandwidth reservation of 1/2, 1/3 and 1/6 of the total bandwidth allocated to a node to the three classes within a node. 4.3. Admission control IEEE 802.11 is based on CSMA/CA access mechanism. Due to its inherent distributed nature, it is very difficult to provide QoS guarantees in an ad hoc scenario. In order to provide QoS guarantees, there needs some kind of check on the number of flows or load in the system and only admission control can do this function. Even with the admission control in the 802.11, we can only provide statistical QoS guarantees. The admission control in our framework can be implemented in both distributed and centralized scenarios. We developed a measurement assisted model based admission control in a distributed scenario and obtained encouraging results [34]. The implementation details of the admission control is beyond the scope of this paper. We will present a brief overview of it here. In our framework, the application layer interacts with the admission control to find out whether it can allow a new flow of traffic. The core MAC layer can measure various parameters like service time delay, channel capacity or utilization, collisions, collision length and nodes. The admission control unit interacts with the core MAC layer to get the measured information. We also estimate the current load in the system using a load estimation method we developed [34]. The analytical model presented in the Section 4.1.2 is based on equal load in all the nodes. The current load estimate in a node is obtained by dividing the system load with the number of nodes. This current load in the node along with the load from the new flow is applied in the analytical model to obtain the one hop delay estimate. If the one hop delay estimate satisfies the delay requirements of all the existing flows and also the new flow then it admits the new flow. Otherwise, it rejects admission to the new flow. We give the algorithm below. N = nodes knew = load for the new flow ktotal = total load in the system knode = load estimated in the node P = measured values on the channel Dmin = minimum one hop delay from all the flows Dnew = one hop delay requirement for the new flow Dest = one hop delay estimate with the new flow included Get P from the Core MAC layer Estimate ktotal knode = ktotal/N Estimate the Dest using (knode + knew) and P using Eq. (26) if (Dest < Dmin and Dest < Dnew)

4041

admit the new flow Dmin = min(Dmin,Dnew) else Reject the flow

5. Performance analysis and evaluation In this section, we validate the analytical model for the p-persistent 802.11 used in the core MAC layer. We also present the simulation results and analysis of our scheduler and the core MAC layer proposed in our framework. Our results for the model show the correctness of the proposed model for the p-persistent 802.11. We also show that our framework provides the desired throughput service differentiation according to the values set in the QoS parameters while providing the maximum possible system throughput. We compared the performance of our scheme with the adaptive EDCF (AEDCF) scheme [19]. AEDCF is a dynamic scheme that adapts to the varying network and load conditions to provide a best possible system performance for the given parameters. We use AEDCF scheme for comparison with our scheme for the following reasons. The main emphasis of this paper is to show the advantages of separating the channel access and scheduling components in the IEEE 802.11 MAC. The performance of the admission control in our scheme is beyond the scope of the paper and so we did not choose scheme which implements an admission control. The AEDCF uses the same set of parameters for both throughput QoS differentiation and channel access. Thus, it fits a good scheme to compare the efficiency of separating channel access and service differentiation in our scheme with a scheme which uses same set of parameters for both the functionalities. The NS2 implementation of AEDCF [33] was obtained with permission from the author Qiang Ni [19]. Our tests have shown that it is very difficult to set the AEDCF parameters to accurately achieve a desired throughput QoS differentiation among the traffic classes of a node. For example, there is no clear way of setting the parameters CWmin, CWmax, AIFS, and PF in order to achieve a throughput ratio of 2:1 (3:2:1) between two (three) classes of traffic. In contrast, achieving an exact throughput differentiation in our framework is straightforward. Since, we separated the channel access component from the service differentiation component the core MAC layer can concentrate on achieving efficient channel access and is therefore capable of achieving maximum possible system throughput than the AEDCF. Our performance tests also showed that for a given throughput differentiation, our scheme provides better channel access delay, packet service time delay and one hop delay than the AEDCF scheme. 5.1. Simulation setup The simulation results for all the tests reported in this paper are generated using the NS2 simulator [32]. The scheduler-based framework is incorporated into the NS2 simulator by modifying the protocol stack of the wireless node and its IEEE 802.11 EDCF implementation. The original 802.11e MAC can still be run without any change in its functional behavior. The scheduler decides the next frame to be sent from the queues associated with different classes of traffic and informs the Core MAC layer about the frame. Upon transmitting the frame, the MAC layer calls the scheduler to provide with the next frame. We also added a static ARP and a dumb routing protocol to the NS2 code to avoid any overhead due to the ARP and the routing protocol. The static ARP does not send any ARP packets and the

4042

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048

dumb routing protocol does not use any routing mechanism. It just sends the data packets down the stack to the link layer. In our study, we simulate a WLAN made up of wireless stations with no hidden nodes. We studied the performance of our framework for different sets of wireless stations varying from 5 to 100. However, we present results only for 10 nodes unless specified exclusively in the graphs. The packet sizes were also varied from 10 to 100 data transmission slots. The WLAN operates in IEEE 802.11a mode with each station operating at 54 Mbps data transmission rate. Our simulation scenario consists of a circular topology, where node ‘i’ sends traffic from all its classes to node ‘i + 1’ when i is less than 10, and the last node ‘10’ sends its traffic to the node ‘1’. The MAC layer system parameters are set as shown in Table 1.

‘‘finite retry limit” cases. In the finite retry case, we studied retry limits from 1 to 7. The simulation results for all the retry limits and no retry limit matched with analytical results. However, we present only the results for the retry limit of 6. In Figs. 4 and 5, we see a very close match between the analytical and simulation results (within 2% error) when the nodes are varied from 5 to 100 for both the saturated load and the unsaturated load (of 25 Mbps). In the next validation experiment, we varied the system load from 5 to 100 Mbps for a fixed number of 10 nodes and Figs. 6 and 7 show that the simulation results matched the analytical results with less than 2% error. Thus, we conclude that our model accurately defines the p-persistent 802.11 WLAN behavior. 5.3. Throughput differentiation

5.2. Model validation for the p-persistent 802.11 In the following experiment, we evaluated the performance of the scheduler from our framework and AEDCF in terms of throughput service differentiation among the traffic classes. We ran each simulation run five times for a duration of 50 s with traffic sources activated at t = 0.000001 s and deactivated at t = 50 s. We generated the CBR traffic at the same rate in each class in each station. We conducted the tests for two (and three) classes and the traffic rates are varied from 1 to 3.5 Mbps to generate a system load of 20 (30 Mbps) to 70 Mbps (105 Mbps). The data packet sizes are also varied from 10 to 100 transmission slots (100 slots is larger than the maximum 802.11 frame size but is used for evaluation purposes only). In our scheme, the desired throughput service differentiation is easily and accurately achieved by simply setting the weights of the different traffic classes in the desired proportions in the scheduler component. The AEDCF scheme, however, uses a set of parameters namely CWmin, CWmax, AIFS, and PF for each class to provide throughput differentiation as well as channel access. It becomes a very complex operation to set these parameters correctly in order to achieve a desired throughput differentiation among the classes. To further illustrate this important issue, let us assume we want to achieve a throughput service differentiation ratio of 3:2:1 between class1, class2 and class3 traffics. To do so, we simply set the weights of class1, class2 and class3 to 1, 2/3, and 1/3 respectively in our scheduler. In the AEDCF scheme, however, a tedious search process is needed to find the parameter values to achieve the required differentiation ratio. For example, setting each of

The analytical models were implemented in C++ with the same set of parameters as that used in the NS2 simulator. The results presented here are for a data packet size of 50 transmission slots. To verify our analytical models, we ran each simulation for 25 s and activated all the traffic sources at t = 0.000000001 s and deactivated them at t = 25 s. The system load is equally divided among all the given nodes. For instance when 10 nodes are used, the exponential traffic is generated with a load ranging from 0.5 to 10 Mbps in each node which is equivalent to a system load of 5–100 Mbps. We used single class traffic in our validation. We validate our model in terms of service time and one hop delay by varying the number of nodes and amount of the system load. We evaluated our analytical model for both the ‘‘no retry limit” and

Table 1 802.11 MAC parameters 9 ls 16 ls 34 ls 54 Mbps 6 Mbps 144 bits 48 bits 14 bytes 28 bytes

tslot SIFS DIFS Data rate PLCP data rate Preamble length bits PLCP header length ACK length bytes MAC header length bytes

70 anal. model m=6 sat load simulation m=6 sat. load anal. model m= 6, load= 25Mbps simulation m=6, load=25Mbps

Service Time Delay (m sec)

60

50

40

30

20

10

0 5

10

15

20

25

30

35

40

45

50

55

60

65

70

Nodes Fig. 4. Average service time-finite retry limit.

75

80

85

90

95

100

4043

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048

350 anal. model m=6 sat load simulation m=6 sat load anal. model m=6 load=25Mbps simulation m=6 load=25Mbps

One Hop Delay (m sec)

300

250

200

150

100

50

0 5

10

15

20

25

30

35

40

45

50

55

60

65

70

75

80

85

90

95

100

Nodes Fig. 5. Average one hop delay-finite retry limit.

7 anal. model m=6 Nodes=10 simulation m=6 Nodes=10

Service Time Delay (m sec)

6

5

4

3

2

1

5

10

15

20

25

30

35

40

45

50

55

60

65

70

75

80

85

90

95

100

75

80

85

90

95

100

Sys. Load (Mbps) Fig. 6. Average service time-finite retry limit. 30 anal. model m=6 Nodes=10 simulation m=6 Nodes=10

One Hop Delay (m sec)

25

20

15

10

5

0 5

10

15

20

25

30

35

40

45

50

55

60

65

70

Sys. Load (Mbps) Fig. 7. Average one hop delay-finite retry limit.

4044

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048

the parameters CWmin, CWmax, PF for class2 and class3 to two and three times the values of class1 parameters does not result in a throughput differentiation ratio of 3:2:1. To evaluate the behavior of AEDCF when we double and triple the parameter values of class2 and class3 with respect to class1, we used the MAC parameters given in Table 2 for the AEDCF scheme with three classes. In our results, we present the throughput ratio for classes as a fraction of system throughput. So, a throughput ratio of 3:2:1 is equivalent to 0.5, 0.33 and 0.17 in terms of fraction of the system throughput for class1, class2 and class3 respectively. Our framework with an underlying p-persistent Core MAC is referred to as SPDCF from now onwards. We use SPDCF3 to mean our scheme with a throughput differentiation of 3:2:1 for three classes of traf-

Table 2 AEDCF3A MAC parameters for the three classes Parameters

Class1

Class2

Class3

CWmin CWmax AIFS (ls) PF

5 200 34 2

10 400 34 4

15 600 34 6

Table 3 AEDCF3 MAC parameters for the three traffic classes (3:2:1) Load/class/node (Mbps)

Class

CWmin

CWmax

AIFS

PF

1, 1.25, 1.5, 1.75, 2.0

1 2 3 1 2 3 1 2 3 1 2 3 1 2 3

7 5 9 7 5 9 7 5 9 7 5 9 8 5 9

200 275 350 225 250 325 250 200 300 300 200 300 350 250 325

43 34 34 43 34 34 43 34 34 43 34 34 43 34 34

2.0 2.0 2.0 2.0 2.0 2.0 2.0 2.0 2.0 2.0 2.0 2.0 2.0 2.0 2.0

2.5, 2.75

3.0

3.25, 3.5

1 AEDCF3A CLASS1 AEDCF3A CLASS2 AEDCF3A CLASS3

0.9

Fraction of the System Throughput

2.25

fic respectively. We use the notation AEDCF3A to denote the AEDCF scheme with 3 classes and with an ad-hoc approach of setting the parameter values in proportion to the required throughput differentiation. Thus we use AEDCF3A to denote the AEDCF scheme with parameters set according to Table 2 for three classes of traffic. In order to get a throughput ratio of 3:2:1 for the AEDCF scheme, we found that the parameter values depend on the traffic load and we had to run several simulation tests with different parameter combinations at each load rate. Table 3 gives the values of the parameters that reasonably produced the desired throughput ratio 3:2:1 at the various traffic loads for 10 nodes. We refer to the AEDCF scheme that uses the parameters values given in Table 3 as AEDCF3. The parameter values for the classes given in Table 3 show that class2 and class3 were given higher priority than class1 in terms of AIFS. It should be noted that the virtual collision handler and the AEDCF behavior give skewed priority for class1 over class2 and 3. In order to achieve 3:2:1 differentiation, the parameter values were chosen so that they compensate the emphasis given by the virtual collision handler to class1. This explains why class2 was sometimes given more favorable parameter values compared to class1 but class1 still ended up getting 1/2 the throughput. Also, the throughput differentiation in AEDCF is not dependent on just one parameter but a combination of all the four parameters from each class. In summary, the ad-hoc scheme AEDCF3A is based on Table 2 that increases the value of the parameters CWmin, CWmax, and PF for a traffic class in proportion to the desired increase in the throughput of this class. The guided scheme AEDCF3 is based on Table 3 that sets the values of the channel access parameters guided by an exhaustive search to achieve the desired throughput differentiation. We next examine the performance of the two schemes under different CBR traffic loads. The load rate for each class per node is varied from 1 to 3.5 Mbps giving a total system load of 30 to 105 Mbps for three class traffic. We observe from Figs. 8–10 that there is no throughput differentiation for loads below 1.25 Mbps in all the schemes AEDCF3A, AEDCF3 and SPDCF3. This is because the traffic load is not sufficient enough to differentiate between the three classes. As the load is increased up to 2 Mbps, the class1, class2 and class3 are differentiated but not to the desired ratio as the total load is still not suf-

0.8 0.7 0.67 0.6 0.5 0.4 0.33

0.2 0.17 0.1 0 1

1.25

1.5

1.75

2

2.25

2.5

2.75

Load/Class/Node (Mbps) Fig. 8. AEDCF3A: fraction of the system throughput.

3

3.25

3.5

4045

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048 1 AEDCF3 CLASS1 AEDCF3 CLASS2 AEDCF3 CLASS3

0.9

Fraction of the System Throughput

0.8 0.7 0.67 0.6 0.5 0.4 0.33

0.2 0.17 0.1 0 1

1.25

1.5

1.75

2

2.25

2.5

2.75

3

3.25

3.5

3.25

3.5

Load/Class/Node (Mbps) Fig. 9. AEDCF3: fraction of the system throughput.

1 SPDCF3 CLASS1 SPDCF3 CLASS2 SPDCF3 CLASS3

0.9

Fraction of the System Throughput

0.8 0.7 0.67 0.6 0.5 0.4 0.33

0.2 0.17 0.1 0 1

1.25

1.5

1.75

2

2.25

2.5

2.75

3

Load/Class/Node (Mbps) Fig. 10. SPDCF3: fraction of the system throughput.

ficiently large enough. As the load is increased further, the AEDCF3 and SPDCF3 maintained the desired throughput differentiation of 3:2:1 (0.5, 0.33 and 0.17 fractions of system throughput). However, the ad-hoc AEDCF3A scheme gave less preference to class2 traffic and least preference to class3 traffic in accordance to the parameter values given to it. Thus, we show that achieving a desired throughput differentiation is not a straight forward process in the IEEE 802.11e. We achieve this very easily in our framework by separating the channel access and scheduling components. The desired throughput differentiation is achieved very easily by setting the differentiation parameters in the scheduler. 5.4. System throughput In the previous section, we saw the advantage of the scheduler in providing throughput differentiation. In the section, we look into

the performance of the core MAC layer in terms of system throughput. A higher system throughput means better channel access. Fig. 11 shows that the SPDCF3 achieved better system throughput than both the AEDCF3A and AEDCF3 schemes for a given throughput differentiation of 3:2:1. In SPDCF3, the core MAC layer’s function is to just provide the channel access to the packet selected by the scheduler. So, it sets its channel access parameter p to an optimum value, based on the network and load conditions without any concern to the differentiation requirement. However, in the AEDCF, the differentiation component and channel access component are catered to by the same parameters. Even though the AEDCF scheme dynamically adapts its channel access parameters, it is still bound by the initial input access parameters. The virtual collisions within a node also contribute to lower system throughput. The AEDCF3 seems to perform lower than the ad-hoc AEDCF3A because the MAC parameters for the AEDCF3 were set to provide a

4046

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048 55 AEDCF3A AEDCF3 SPDCF3

50

System Throughput (Mbps)

45 40 35 30 25 20 15 10 5

1

1.25

1.5

1.75

2

2.25

2.5

2.75

3

3.25

3.5

Load/Class/Node (Mbps) Fig. 11. System throughput.

specific throughput differentiation of 3:2:1. In the process, this setting affected the over all system throughput as the parameters are also used for channel access and may not be the optimum values to achieve optimum system throughput. This shows that using the same set of parameters for both channel access and service differentiation will not result in optimum performance in both the components at the same time. 5.5. Packet delay

tempt of this packet starts). Notice that the channel access delay does not depend on the length of the packet. The packet service time delay is the amount of time it takes from the time the packet is at the head of the MAC queue till it receives an acknowledgement or is dropped after its retransmission limit is exhausted. The one hop delay is the amount of time a packet takes from the time it arrives at the queue till it is successfully transmitted. The one hop delay is the sum of queue delay and service time delay for a packet.

In this section, we evaluate the performance of our scheme and AEDCF in terms of three types of packet delays viz. channel access delay, packet service time delay and one hop delay for a given throughput differentiation of 3:2:1. This is another look into the performance of the core MAC layer as in the previous section. We define the three delays as follows: The channel access delay is the amount of time a packet takes from the time it is at the head of the MAC queue ready for transmission till the time it starts accessing the channel (i.e., till the time the first transmission at-

5.5.1. Channel access delay We study the channel access delay for both the AEDCF and SPDCF schemes for a throughput differentiation of 3:2:1. The AEDCF uses the same parameters given in the Table 3. Since, we now have the same throughput ratios in both schemes; we compare their channel access delay for the three traffic classes. Fig. 12 shows the channel access delays for each of the three classes of traffic for AEDCF3 and SPDCF3 schemes. The SPDCF3 scheme performed better than the AEDCF3. In our scheme, the core

50 AEDCF3 CLASS1 AEDCF3 CLASS2

Avg. Channel Access Delay (msec)

45 AEDCF3 CLASS3 SPDCF3 CLASS1 SPDCF3 CLASS2 40 SPDCF3 CLASS3

35 30 25 20 15 10 5

1

1.25

1.5

1.75

2

2.25

2.5

Load/Class/Node (Mbps) Fig. 12. Channel access delay.

2.75

3

3.25

3.5

4047

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048 50 45

Avg. Service Time (msec)

40

AEDCF3 CLASS1 AEDCF3 CLASS2 AEDCF3 CLASS3 SPDCF3 CLASS1 SPDCF3 CLASS2 SPDCF3 CLASS3

35 30 25 20 15 10 5

1.25

1

1.5

1.75

2

2.25

2.5

2.75

3

3.25

3.5

Load/Class/Node (Mbps) Fig. 13. Packet service time delay.

MAC layer tries to improve the system capacity by using an optimal p (obtained from Eq. (2). A greater throughput implies lesser channel access latency. However, in AEDCF3 scheme, the parameters used, provide prioritized channel access without emphasizing on optimal throughput. When we set the parameters to achieve a throughput differentiation of 3:2:1, the parameters are not optimal enough to provide better channel access delay. Fig. 12 shows that the channel access delay for the AEDCF3 increases with the load for class3 and class2 traffic as it has least and less priority parameters for channel access. Each class in the AEDCF3 contends for the channel independently and thereby, the channel access delay will be different for different classes where as in our framework, there is only one channel access for all the classes and so all the classes will have the same channel access delay and our results in Fig. 12 show the same.

previous section would give us some idea about the packet service time. Since AEDCF3 is not optimized for channel access, it might take more time to access the channel and might involve more collisions. Fig. 13 shows that the packet service time for AEDCF3 is more than that of the SPDCF3 for all the three traffic classes. The packet service time for class3 traffic in AEDCF3 scheme is higher than that of class1 and class2 as it is given least priority channel access than the class1 and class2. Similarly, the packet service time for class2 traffic in AEDCF3 scheme is higher than that of class1 as it is given lesser priority channel access than the class1 where as the packet service time in the SPDCF3 is nearly same for all the three classes as they have one channel access and the scheduler decides the next traffic class which will compete for the channel access. 5.6. One hop delay

5.5.2. Packet service time delay The service time of a packet depends on the efficiency of its channel access. Therefore, the results from channel access in the

240 220

Avg. One hop Delay (msec)

200

The one hop delay is an important QoS metric which determines whether we can support real-time applications that require

AEDCF3 CLASS1 AEDCF3 CLASS2 AEDCF3 CLASS3 SPDCF3 CLASS1 SPDCF3 CLASS2 SPDCF3 CLASS3

180 160 140 120 100 80 60 40 20 0 1

1.25

1.5

1.75

2

2.25

2.5

Load/Class/Node (Mbps) Fig. 14. One hop delay.

2.75

3

3.25

3.5

4048

K. Anna, M. Bassiouni / Computer Communications 31 (2008) 4035–4048

stringent delay bounds. The one hop delay in our scheme depends on two components. The core MAC layer provides optimum one hop delay possible in a given network and load conditions. The scheduler provides the differentiation of the one hop delay based on the priorities of the traffic. Fig. 14 shows that the SPDCF3 scheme provides better one hop delay than the AEDCF3 scheme. At a load 3.5 Mbps, the one hop delay of class3 and class2 in SPDCF3 is about 2.91 and 1.6 times the one hop delay of class1, whereas the one hop delay of class3 and class2 in AEDCF3 is about 4.16 and 2.36 times the one hop delay of class1. In general when achieving a throughput differentiation of 3:2:1, the ratios of the one hop delay between the three classes in SPDCF3 is slightly lesser than 1:2:3 but is considerably closer to 1:2:3 than the ratio obtained by the AEDCF3. Therefore, we can use either the one hop delay metric or the throughput metric at the scheduler to represent the service differentiation. We observed similar results for the two class traffic [20]. Due to lack of space, we don’t show the results for two classes. 6. Conclusions and future work We presented a new framework for QoS provisioning in single hop WLANS. The QoS differentiation and channel access functionalities in the standard 802.11 MAC are separated into two functional components and assigned to the scheduler and the core MAC layer respectively. In this paper, the performance of the scheduler and the core MAC layer were evaluated with respect to AEDCF. Our scheme not only performed better in terms of service differentiation, system throughput but also demonstrated the ease at which it can be tuned. Our scheme also achieved better channel access delay, packet service time and one hop delay for a given throughput service differentiation demonstrating the advantage of separating service differentiation and channel access components. We also presented an analytical model for the p-persistent 802.11 in the core MAC layer. The closed forms for packet service time and one hop delay in unsaturated load conditions were developed and presented. Our model provides excellent approximation for the average service time and one hop delay. The simulation results for the no retry case and the finite retry cases are very close to the analytical results (within 2% error). In our future work, based on the analytical model, we plan to develop an admission control in both distributed and centralized modes and evaluate the efficacy of utilizing the one hop delays predicted by the analytical model in making admission decisions. References [1] IEEE Std. 802.11-1999, Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, Reference No. ISO/IEC 880211:1999(E), 1999 ed., IEEE Std. 802.11, 1999. [2] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, W. Weiss, An Architecture for Differentiated Services, RFC 2475, 1998. [3] IEEE 802.11 WLANS. Availabel from: . [4] IEEE 802.11e/D11.0, Draft Supplement to Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications: Medium Access Control (MAC) Enhancements for Quality of Service (QoS), 2004. [5] H. Zhu, M. LI, I. Chlamtac, A survey of quality of service in IEEE 802.11 networks, IEEE Wireless Communications 11 (4) (2004) 6–14.

[6] F. Cali, M. Conti, E. Gregori, IEEE 802.11 protocol: design and performance evaluation of an adaptive back-off mechanism, IEEE Journal on Selected Areas in Communications 18 (9) (2000) 1774–1786. [7] F. Cali, M. Conti, E. Gregori, Dynamic tuning of the IEEE 802.11 protocol to achieve a theoretical throughput limit, IEEE/ACM Transactions on Networking 8 (6) (2000) 785–799. [8] G. Bianchi, Performance analysis of the IEEE 802.11 distributed coordination function, IEEE Journal on Selected Areas in Communication 18 (3) (2000) 535– 547. [9] G. Bianchi, I. Tinnirello, Kalman filter estimation of the number of competing terminals in an IEEE 802.11 network, in: Proceeding of IEEE INFOCOMM 2003, vol. 22, no. 1, 2003, pp. 844–852. [10] C.H. Foh, M. Zukerman, Performance analysis of the IEEE 802.11 MAC protocol, in: Proceedings of European Wireless 2002, 2002, pp. 184–190. [11] H. Wu, Y. Peng, K. Long, S. Cheng, J. Ma, Performance of reliable transport protocol over IEEE 802.11 wireless LAN: analysis and enhancement, in: Proceedings of IEEE INFOCOM 2002, vol. 21, no. 1, 2002, pp. 599–607. [12] Y. Ge, QoS Provisioning for IEEE 802.11 MAC Protocols, Ph.D. Dissertation, Ohio State University, 2004. [13] Y. Yang, R. Kravets, Distributed QoS guarantees for reatime traffic in ad hoc networks, in: Proceedings of SECON 2004, 2004, pp. 118–127. [14] W. Liu, W. Lou, X. Chen, Y. Fang, A QoS-enabled MAC architecture for prioritized service in IEEE 802.11 WLANS, in: Proceedings GLOBECOM 2003, vol. 7, 2003, pp. 3802–3807. [15] N. Vaidya, P. Bahl, S. Gupta, Distributed fair scheduling in a wireless LAN, in: Proceeding of MOBICOM 2000, 2000, pp. 167–178. [16] A. Banchs, X. Perez, Distributed weighted fair queuing in 802.11 wireless LAN, in: Proceedings of ICC 2002, vol. 5, 2002, pp. 3211–3127. [17] W. Atikom, S. Banerjee, P. Krishnamurthy, Starvation prevention and quality of service in wireless LANs, in: Proceedings of the IEEE 5th Intl. Symposium on Wireless Personal Multimedia Communications (WPMC), October 2002, vol. 3, pp. 1078–1082. [18] H. Kim, J. Hou, Improving protocol capacity with model-based frame scheduling in IEEE 802.11-operated WLANs, IEEE Journal on Selected Areas in Communication 22 (10) (2004). [19] L. Romdhani, Q. Ni, T. Turletti, Adaptive EDCF: enhanced service differentiation for IEEE 802.11 wireless ad-hoc networks, in: Proceedings of IEEE WCNC 2003, vol. 4, no.1, 2003, pp. 1373–1378. [20] K. Anna, M. Bassiouni, A scheduler-based architecture for QoS provisioning in IEEE 802.11 MAC protocol, in: Proceedings of IEEE TeNe’05, 2005. [21] K. Anna, M. Bassiouni, Analytical modeling and QoS application of p-persistent 802.11 MAC protocol, in: Proceedings of IEEE GLOBECOM’06, 2006. [22] N. Gupta, P.R. Kumar, A performance analysis of the 802.11 wireless LAN medium access control, Journal on Communications in Information and Systems 3 (4) (2004) 279–304. [23] P. Chatzimisios, A.C. Boucouvalas, V. Vitsas, IEEE 802.11 packet delay a finite retry limit analysis, in: Proc. IEEE Global Telecommunications Conference (GLOBECOM’03), San Francisco, California, USA, 2003, vol. 2, pp. 950–954. [24] P. Chatzimisios, A.C. Boucouvalas, V. Vitsas, IEEE 802.11 wireless LANS: performance analysis and protocol refinement, EURASIP Journal on Applied Signal Processing (2005) 67–78. [25] M.M. Carvalho, J.J. Garcia-Luna_Aceves, Delay analysis of IEEE 802.11 in single hop networks, in: Proc. 11th IEEE International Conference on Network Protocols (ICNP’03), Atlanta, GA, USA, 2003, pp. 146–155. [26] E. Ziouva, T. Antonakopoulos, CSMA/CA performance under high traffic conditions: throughput and delay analysis, Computer Communications 25 (3) (2002) 313–321. [27] O. Tickoo, B. Sikdar, Queueing Analysis and Delay Mitigation in IEEE 802.11 Random Access MAC based Wireless Networks, IEEE INFOCOM, Hong Kong, China, 2004. [28] D. Malone, K. Duffy, D.J. Leith, Modeling the 802.11 distributed coordination function in non-saturated heterogeneous conditions, IEEE/ACM Transactions on Networking 15 (1) (2007) 159–172. [29] Y. Barowski, S. Biaz, Towards the performance analysis of IEEE 802.11 in multihop ad-hoc networks, in: Proceedings of IEEE WCNC‘05, no. 1, 2005, pp. 100– 106. [30] M. Garetto, C.-F.Chiasserini, Performance Analysis of the 802.11 Distributed Coordination Function under Sporadic Traffic, Networking 2005, Waterloo, Ont., Canada, 2005. [31] L. Klienrock, Queuing Systems, vol. 1, John Wiley & Sons, 1976. [32] The Network Simulator (NS2). Available from: . [33] http://www-sop.inria.fr/planete/qni/Research/AEDCF/. [34] K. Anna, M. Bassiouni, A distributed admission control to provide statistical QoS in p-persistent 802.11 MAC protocol, in: Proceedings of IEEE Wireless Telecommunication Symposium (WTS), April 2008.