2 BACKGROUND. J. Yu and S. Choi

Transcription

1 Comparison of Modified Dual Queue and for VoIP over IEEE.11 WLAN JEONGGYUN YU AND SUNGHYUN CHOI Multimedia & Wireless Networking Laboratory (MWNL) School of Electrical Engineering and INMC, Seoul National University, 151-7, Korea and Abstract. The popular IEEE.11 WLAN today does not provide any quality-of-service (QoS) because of its contention-based channel access nature of the medium access control (MAC). Therefore, we proposed a simple software upgrade-based solution, called a dual queue scheme, to provide QoS to real-time services such as Voice-over-IP (VoIP) in our previous work [7]. The dual queue scheme operates on top of the legacy MAC. In this paper, we present a modified dual queue (MDQ) scheme by considering a practical implementation limitation. On the other hand, the emerging.11e MAC includes a contention-based channel access mechanism, called enhanced distributed channel access (). Prioritized channel accesses are provided to different traffic classes under. In this paper, we comparatively evaluate the performances of MDQ and via simulations in terms of delay, jitter, and throughput for various scenarios. We also compare the two access schemes when multiple transmission rates are used via rate adaptation mechanism for reliable transmission over erroneous wireless channel. The simulation results show that apparently provides a better performance in most situations. However, MDQ provides an acceptable performance close to that of so that it can be used when the is not available and/or the hardware upgrade to.11e is not desirable for the cost reason. 1 INTRODUCTION In recent years, IEEE.11 WLAN has gained a prevailing position in the market for the (indoor) broadband wireless access networking. The IEEE.11 standard defines the medium access control (MAC) layer and the physical (PHY) layer specifications [1]. The mandatory part of the.11 MAC is called the distributed coordination function (DCF), which is based on Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA). Today, most of the.11 devices implement the DCF only. Because of the contention-based channel access nature of the DCF, it supports only the best-effort service without guaranteeing any Quality of Service (QoS). Recently, the needs for real-time (RT) services such as Voice over IP (VoIP) and audio/video (AV) streaming over the WLANs have been increasing drastically. However, the current.11 devices are not capable to support the RT services properly, which are delay-sensitive while tolerable of some losses. The emerging IEEE.11e MAC, which is an amendment of the existing.11 MAC, will provide the QoS [3, 5,,, 13, 1]. The new MAC protocol of the.11e is called the Hybrid Coordination Function (HCF). The HCF contains a contention based channel access mechanism (), which is an enhanced version of the legacy An earlier version of this paper was presented at European Wireless Conference 5. This work was in part supported by the Korea Research Foundation Grant funded by Korean Government (MOEHRD) (R ). DCF, for a prioritized QoS support. With, a single MAC contains multiple queues with different priorities that access channel independently in parallel. Frames in each queue are transmitted using different channel access parameters. In our previous work in [7, 1], we proposed the dual queue scheme with the legacy.11 MAC, which is a software upgrade-based approach to provide a QoS for real-time applications such as VoIP. This is proposed as a short- and mid-term solution to provide QoS in the.11 WLAN since it does not require any WLAN device hardware (HW) upgrade. Note that for many of today s.11 MAC implementations, the.11e requires a HW upgrade. However, replacing existing.11 HW devices to provide QoS should be quite costly, and hence, may not be desirable to many WLAN owners, especially, hotspot service providers with a huge number of deployed APs [1]. The dual queue scheme basically implements two queues in the device driver of the.11 WLAN devices. Therefore, these queues are conceptually located on top of the.11 MAC controller running the legacy DCF. Realtime (RT) and non-real-time (NRT) packets 1 are classified and enqueued into one of these two queues. Then, the RT queue is always served prior to the NRT queue via strict priority queuing. We showed that the dual queue approach provides a good QoS to the VoIP packets [7]. However, it 1 In the.11 MAC terms, frame is a better word than packet. However, we use the terms frame and packet inter-changeably in this paper since we consider the issue across layers, i.e., VoIP over WLAN. Submission 1

2 J. Yu and S. Choi is identified that the performance of the dual queue scheme could be compromised due to the queuing delay within the MAC HW queue, and in many MAC implementations, the queue size cannot be adjusted by the system integrator. Accordingly, we, in this paper, propose a modified dual queue scheme to minimize the impact of the MAC HW queue delay. In this paper, we compare the modified dual queue scheme and the emerging.11e. We show that the modified dual queue scheme provides an acceptable performance, and the performance is quite comparable with that of the in most situations. We also compare the two access schemes when multiple transmission rates are used via rate adaptation mechanism for reliable transmission over erroneous wireless channel. The VoIP performance of MDQ is degraded due to its MAC HW queue, which is a FIFO queue, when non-real-time packet transmission time increases. On the other hand, can protect existing VoIP traffic from such non-real-time traffic thanks to a novel mechanism, called TXOP. Therefore, we conclude that the dual queue scheme can be a good solution for QoS provisioning until the.11e devices prevail. However, we eventually need to use.11e MAC to maximize channel utilization in multirate WLAN. The biggest advantage of the.11e is the capability to adjust the channel access parameters depending on the underlying network condition, e.g., the traffic load. We in this paper investigate a possible access parameter adaptation for the AP s fast channel access. However, it seems not to be very easy to find the proper access parameters for a given condition in general, and hence the development of algorithms to adjust the access parameters deserves more future research efforts. The rest of the paper is organized as follows. Section briefly introduces legacy IEEE.11 DCF and VoIP. Sections 3 and present the modified dual queue scheme and.11e, respectively. Section 5 discusses the characteristics of MDQ and in the multiple rate supporting WLAN. After comparatively evaluating the dual queue scheme and scheme via simulations in Section, we conclude in Section 7. BACKGROUND.1 IEEE.11 LEGACY MAC The IEEE.11 legacy MAC [1] defines two coordination functions, namely, the mandatory DCF based on CSMA/CA and the optional point coordination function (PCF) based on poll-and-response mechanism. Most of today s.11 devices operate only in the DCF mode. We briefly overview how the DCF works here as the dual queue scheme proposed in [7] runs on top of the DCF-based MAC and the.11e is also based on it. The.11 DCF works with a single first-in-first-out (FIFO) transmission queue. The CSMA/CA of the DCF works as follows: when a packet arrives at the head of Figure 1: IEEE.11 DCF channel access scheme Table 1: MAC Parameters for.11b Parameters SIFS DIFS Slot CWmin CWmax (µs) (µs) (µs).11b PHY transmission queue, if the channel is busy, the MAC waits until the medium becomes idle, then defers for an extra time interval, called the DCF Interframe Space (DIFS). If the channel stays idle during the DIFS deference, the MAC then starts the backoff process by selecting a random backoff counter. For each idle slot time interval, the backoff counter is decremented. When the counter reaches zero, the packet is transmitted. The timing of DCF channel access is illustrated in Figure 1. Each station maintains a contention window (CW), which is used to select the random backoff counter. The backoff counter is determined as a random integer drawn from a uniform distribution over the interval [, CW ]. If the channel becomes busy during a backoff process, the backoff is suspended. When the channel becomes idle again, and stays idle for an extra DIFS time interval, the backoff process resumes with the suspended backoff counter value. For each successful reception of a packet, the receiving station immediately acknowledges by sending an acknowledgement (ACK) packet. The ACK packet is transmitted after a short IFS (SIFS), which is shorter than the DIFS. If an ACK packet is not received after the data transmission, the packet is retransmitted after another random backoff. The CW size is initially assigned CWmin, and increases to (CW + 1) 1 when a transmission fails. All of the MAC parameters including SIFS, DIFS, Slot Time, CWmin, and CWmax are dependent on the underlying physical layer (PHY). Table 1 shows these values for the.11b PHY []. The.11b PHY supports four transmission rates, namely, 1,, 5.5, and 11 Mbps. We assume the.11b PHY in this paper due mainly to its wide deployment base even if the proposed dual queue scheme and the.11e should work with any PHY.. VOICE-OVER-IP (VOIP) There are many types of voice codecs used for IP telephony, namely, G.711, G.73.1, G.79, and so on. These codecs have different bit rates and complexities. Table lists the characteristics of several commonly used voice codecs. More recent voice codecs like G.79B and G.73.1A have the ability to detect talk-spurts and silence ETT

3 Comparison of Modified Dual Queue and for VoIP over IEEE.11 WLAN Table : Voice Codec Characteristics Codec Voice Data Rate Data Size Packets/s UDP/RTP/IP/SNAP.11 Overhead Total Bandwidth (kbits/s) (Bytes) Overhead (Bytes) (Bytes) without VAD a (kbits/s) G.711 1/ 5/1 5 1/1 G /.3 3 / 5 7./. G.79 3/ 33/ /. a VAD stands for Voice Activity Detection. gaps within a conversation. During a silence time, the codec stops transferring data, in order to save bandwidth. This mechanism is referred to as voice activity detection (VAD). In this paper, we consider G.711, the simplest voice codec. G.711 generates a kbps stream, based on an -bit pulse coded modulation (PCM), with the sampling rate of samples/second. Even though it achieves the worst compression among peer voice codecs, it is often used in practice thanks to its simplicity. The number of samples per a VoIP packet is another important factor. The codec defines the size of a sample, but the total number of samples conveyed in a packet affects how many packets are generated per second. There is basically a trade-off since the larger a packet size (or more samples carried per packet), the longer the packetization delay, but the lower the packetization overheads as analyzed below. In this paper, we assume that a VoIP packet is generated every msec, i.e., with 1-byte (= kbytes/s ms) voice data. We also assume that RTP over UDP is used for the VoIP transfer. When an IP datagram is transferred over the.11 WLAN, the datagram is typically encapsulated by an IEEE. Sub-Network Access Protocol (SNAP) header. Note that all these assumptions are very typical in the real world. Accordingly, the VoIP packet size at the.11 MAC Service Access Point (SAP) becomes: 1-byte DATA + 1-byte RTP header + -byte UDP header + -byte IP header + -byte SNAP header = bytes per VoIP packet 3 MODIFIED DUAL QUEUE SCHEME WITH LEGACY.11 MAC We proposed a simple dual queue scheme in [7, 1] to provide a QoS for the VoIP service enhancement over.11 WLAN. The biggest advantage of this scheme is that it can be implemented in the existing.11 hardware. The dual queue approach is to implement two queues, called RT and NRT queues, inside the AP as shown in Figure. Especially, these queues are implemented above the.11 MAC controller, i.e., in the device driver of the.11 network interface card (NIC), such that a packet scheduling can be performed in the driver level. Packets from the higher layer or from the wireline port (in case of Note that we can easily extend this scheme by implementing more queues depending on the desired number of traffic types to support. Figure : Device driver structures the AP) are classified to transmit into RT or NRT types. The port number as well as UDP packet type is used to classify a RT packet. Packets in the queues are served by a simple strict priority queuing so that the NRT queue is never served as long as the RT queue is not empty. It turns out that this simple scheduling policy results in a surprisingly good performance. We have also implemented the dual queue scheme in the HostAP driver [11] of Intersil Prism.5 chipset [1]. The MAC controller itself has a FIFO queue (referred to as MAC HW queue ). The performance of the dual queue scheme is compromised due to the queuing delay within the FIFO queue when the FIFO queue is large [7]. Unfortunately, the size of the MAC HW queue cannot be configured in many chipsets. To handle this, we have implemented a NRT packet number controller (marked as flow control in Figure ), which restricts the number of outstanding NRT packets in the MAC HW queue. We refer to this modified scheme as Modified Dual Queue (MDQ). For the simulation of the modified dual queue in this paper, we assume that the number of NRT packets in the MAC HW queue is limited to two thanks to the flow control unit. This is the smallest number, which can be practically implemented..11e ENHANCED DISTRIBUTED CHAN- NEL ACCESS () The.11 legacy MAC does not support the concept of differentiating packets with different priorities. Basically, the DCF is supposed to provide a channel access with Submission 3

4 J. Yu and S. Choi Table 3: User priority to access category mappings User Priority Access Category Designation 1 AC BK Background AC BK Background AC BE Best-effort 3 AC BE Best-effort AC VI Video 5 AC VI Video AC VO Voice 7 AC VO Voice equal probabilities to all stations contending for the channel access in a distributed manner. However, equal access probabilities are not desirable among stations with different priority packets. The is designed to provide differentiated, distributed channel accesses for packets with different priorities (from to 7) by enhancing the DCF [3]. Basically, a channel access function uses AIFS[AC], CWmin[AC], and CWmax[AC] instead of DIFS, CWmin, and CWmax, of the DCF, respectively, for the contention process to transmit a packet belong to access category AC. AIFS[AC] is determined by AIF S[AC] = SIF S + AIF SN[AC] SlotT ime (1) where AIFSN[AC] is an integer greater than 1 for STAs and an integer greater than for APs. The backoff counter is selected from [, CW[AC]]. Figure 3 shows the timing diagram of the channel access. The values of AIFSN[AC], CWmin[AC], and CWmax[AC], which are referred to as the parameter set, are advertised by the AP via Beacons and Probe Response frames. The AP can adapt these parameters dynamically depending on the network condition. Basically, the smaller AIFSN[AC] and CWmin[AC], the shorter the channel access delay for the corresponding priority, and hence the more capacity share for a given traffic condition. However, the collision probability increases when operating with smaller CWmin[AC]. These parameters can be used in order to differentiate the channel access among different priority traffic. It should be also noted that the AP can use the parameter values different from the announced ones for the same AC. The.11 DCF originally is designed to provide a fair channel access to every station including the AP. However, since typically there is more downlink (i.e., AP-to-stations) traffic, AP s downlink access has been known to be the bottleneck to the entire network performance. Accordingly, the, which allows the differentiation between uplink and downlink channel accesses, can be very useful to control the network performance. Figure shows the.11e MAC with four channel access functions, where each access function behaves as a single enhanced DCF contending entity. Each access AIFS[AC] Immediate access when medium is idle >= AIFS[AC] PIFS AIFS[AC] Busy Medium SIFS Defer Access Contention Window from [, CW[AC] ] Backoff Window Slot Time Next Frame Select Slot and decrement backoff as long as medium stays idle Figure 3: IEEE.11e channel access scheme Figure : Four channel access functions for function employs its own AIFS[AC] and maintains its own Backoff Counter (BC). When there are more than one access functions finishing the backoff at the same time, the highest priority frame among the colliding frames is chosen and transmitted, and the others perform a backoff with increased CW values. 5 MDQ AND IN MULTI-RATE.11 WLAN Most of.11 Network Interface Cards (NICs) implement an automatic rate control scheme in which the sending stations adaptively change the data rate based on perceived channel conditions. Often, users can also manually set the data rate. The.11b standard defines four different data rates, namely, 1,, 5.5 and 11 Mbits/s. This makes the competing nodes within a BSS use different data rates. The.11 DCF basically provides equal long-term transmission opportunity to contending stations. Specially, when the data rate and frame size of all the stations are the same, DCF achieves the equal throughput among contending nodes [1]. However, if there is at least one station with a lower rate, a performance anomaly can occur [15]. That is, the throughput of all higher rate stations is degraded to the level of the lower rate station. Since MDQ scheme is implemented in the legacy.11 DCF-based device, which supports multiple rates and typically implements a automatic rate fallback (ARF) scheme [17], the VoIP performance of MDQ scheme can be degraded. Moreover, MDQ involves a MAC HW queue, which is a FIFO queue. It can lead to a head-of-line blocking problem when the channel quality towards the desti- ETT

5 Comparison of Modified Dual Queue and for VoIP over IEEE.11 WLAN Table : Default parameter set AC CWmin CWmax AIFSN AC BK acwmin acwmax 7 AC BE acwmin acwmax 3 AC VI (acwmin+1)/ - 1 acwmin AC VO (acwmin+1)/ - 1 (acwmin+1)/ - 1 nation of the head-of-line packet is degraded, and AP accordingly reduces the data rate to transmit the packet via its rate adaptation scheme. For example, when there are the large TCP packets preceding voice packets in the MAC HW queue, the delay performance of the voice packets can be severely degraded. On the other hand, the performance anomaly problem can be solved in.11e thanks to the support of transmission opportunity (TXOP), which is an interval of time when a particular station has the right to initiate transmissions. During a TXOP, there can be a set of multiple frame exchange sequences, separated by SIFS, initiated by a single station. On the other hand, if the TXOP is too small to transmit an entire frame, the frame should be fragmented into multiple small frames, and a contention is needed for each fragment transmission. A TXOP can be obtained by a successful contention, and it is called TXOP in this case. The TXOP mechanism limits the time interval during which a station can transmit its frames, thus providing the temporal fairness with contending stations [1]. We can eliminate the performance anomaly of.11 WLAN via temporal fairness guarantee. We will evaluate this capability of later via simulations. COMPARATIVE PERFORMANCE EVALU- ATION In this section, we comparatively evaluate the performance of the original DCF, the modified dual queue (MDQ) scheme, and the.11e using ns- simulator [9]. We use the following traffic models for our simulations: two different types of traffic are considered for our simulations, namely, voice and data. The voice traffic is modeled by a two-way constant bit rate (CBR) session according to G.711 codec [1]. The data traffic application is modeled by a unidirectional FTP/TCP flow with 1-byte packet size and 1-packet (or 175-byte) receive window size 3. This application corresponds to the download of a large file via FTP. We use the.11b PHY for our simulations. The 11 Mbits/s transmission rate (out of 1,, 5.5, and 11 Mbits/s of the.11b PHY) is used in the simulations unless specified otherwise. Table shows the parameter set used for each traffic type along with the corresponding priorities and ACs. This is the default parameter set in [3]. We use the default values of Table in our simulations unless 3 This window size corresponds to the TCP implementation of MS Windows XP. Figure 5: Network topology for simulations Figure : Delay of VoIP packets with single queue, MDQ, and specified otherwise. We use the queue size of 5 packets at the AP, which is large enough to ensure that there is no buffer overflow in our simulation environments [7]. Moreover, we assume that the MAC HW queue size for the MDQ scheme is equal to two packets as discussed earlier. The network topology for our simulations is shown in Figure 5. Each station involving with a VoIP session generates and receives only voice traffic. The other stations receive only TCP packets, and each of them treats only one TCP flow, i.e., the number of TCP flows corresponds to downstream TCP stations. This topology can be often found in the real WLANs with mixed VoIP and Internet traffic..1 COMPARISON OF SINGLE QUEUE, MDQ, AND We simulate with a single VoIP session and various numbers of downstream TCP flows in order to compare the VoIP performance of single queue (i.e., the original DCF), MDQ, and. Figure presents the delay performance of these three schemes. As was presented in [7], the downlink delay of the single queue increases linearly as the number of TCP flows increases, and hence cannot be used for VoIP in the mixed traffic environments. On the other hand, both MDQ and provide reasonable delay performance virtually independent of the TCP flow Submission 5

6 J. Yu and S. Choi 1 1 down delay with MDQ up delay with MDQ down delay with up delay with 5 MDQ Delay (msec) 1 Aggregated throughput (Mbits/s) Number of VoIP sessions Number of VoIP sessions Figure 7: VoIP delay performances of MDQ and number. This is because both schemes provide higher priority to the VoIP packets over the TCP packets. The understanding of the detailed behavior of the MDQ scheme should be referred to [7]. Figure shows that the voice delay of is superior to that of the MDQ though both of them show good delay performance. The reason can be understood as follows: first, the uses smaller values of the channel access parameters than the MDQ, based on the legacy DCF, namely, CWmin[AC VO] = 7 and CWmax[AC VO] = 15 for AC VO [3], and CWmin = 31 and CWmax = 13 for the legacy DCF [], respectively, in the case of the.11b PHY. Smaller channel access parameters imply a faster channel access. Second, in the case of the MDQ, the MAC HW queue of packets introduces an extra delay for the downlink VoIP packets since a VoIP packet at the head of the RT queue in the MDQ scheme should wait until all the preceding packets in the MAC HW queue are transmitted. This makes the VoIP downlink delay of the MDQ larger than the uplink delay. It should be also noted that the uplink delay performance of both single queue and MDQ are the same since there is no difference between two schemes in case of uplink in our simulation scenarios. That is, in our simulations, a station transmits only a single type of traffic, i.e., either VoIP or TCP-ACK. Figure 7 presents the delay performance comparison for the MDQ and as the number of VoIP sessions increases. We simulate with 1 downstream TCP flows and various number of VoIP sessions. The provides a better VoIP delay performance than the MDQ scheme with multiple VoIP sessions while both of them still provide acceptable delay performances. As discussed above, the main reasons should be the smaller access parameter values and the MAC HW queue of the MDQ scheme. However, we observe that the delay of, especially, the downlink delay, increases as the VoIP session number increases. This must be a negative effect of small access parameters, i.e., these small values result in some Figure : Aggregated TCP throughput of MDQ and collisions among VoIP packets from different STAs. Figure shows the aggregated throughput performance of downstream TCP flows, which are measured at the AP, with both MDQ and. We observe that the provides a better throughput performance than the MDQ. Note that it takes a shorter time for the to transmit a VoIP packet due to a lower channel access delay. As a result, the allows more time resource for TCP packet transmissions. Moreover, TCP under can get more transmission opportunities than that under the MDQ because it contends in parallel with VoIP under. On the other hand, with MDQ scheme, TCP packets are not served when a VoIP packet exists in the RT queue by strict priority queuing. This is the reason why the TCP throughput with the is a bit larger than that with the MDQ.. DEFAULT ACCESS VS. PIFS ACCESS Though we have observed that the provides a good VoIP delay performance through Figures and 7, the performance of can be even more enhanced via the access parameter control. AIFSN[AC] is an integer greater than one for STAs and an integer greater than zero for APs. Moreover, the values of CWmin[AC] and CWmax[AC] can be set to zero [3]. Therefore, we can use access parameter values of AIFSN[AC VO] = 1, CWmin[AC VO] =, and CWmax[AC VO] =, which are the smallest access parameter values for the AP, for the downlink AC VO of the AP. This allows the AP to transmit its pending VoIP packet after a PIFS channel idle time. The other parameter values are subject to Table, i.e., the default values. We refer to the AP s channel access after PIFS for VoIP packets as PIFS access in the rest of this paper. On the other Basically, the PIFS access of QAP is for the polling or beacon frame transmissions. Therefore, we need to schedule the VoIP packet transmissions properly in order to prevent the beacon or polling frames from being delayed due to the VoIP packet transmissions. Moreover, we need to design the HCCA scheduler in order to defer VoIP packet transmissions via the PIFS access during both Contention Free Period (CFP) and ETT

7 Comparison of Modified Dual Queue and for VoIP over IEEE.11 WLAN 1 Down - Default Up - Default Down - PIFS Up - PIFS 5 - Default - PIFS Average delay (ms) Aggregated throughput (Mbits/s) Number of VoIP sessions Number of VoIP sessions Figure 9: VoIP delay performance of default and PIFS accesses Figure 1: Aggregated TCP throughput of default and PIFS accesses hand, the usage of the default access parameters for both STAs and AP, as we have simulated thus far, is referred to as default access. Figure 9 compares the VoIP delay performance of the default access and the PIFS access as the number of VoIP sessions increases. We simulate with 1 downstream TCP flows and various number of VoIP sessions. As expected, we observe that the downlink delay performance of VoIP is enhanced significantly through the PIFS access in Figure 9. In all cases except for the downlink VoIP delay with the PIFS access, the delay of VoIP packets increases as the number of VoIP sessions increases. The reason can be understood as follows: downlink VoIP packets with the PIFS access hardly experience channel access delay because the AP can transmit its pending VoIP packets after a PIFS idle time without any contention from STAs. On the other hand, with the default access as well as the uplink case with the PIFS access, the value of CWmin[AC VO] is equal to seven, which is too small to avoid collisions in our simulations. This results in large channel access delays. Moreover, as discussed above, TCP under the is more aggressive than the MDQ with strict priority queuing because it contends in parallel with VoIP packets. At the same time, the number of contending TCP STAs to transmit TCP ACK packets is more than the MDQ case due to aggressive downlink AC BE for TCP data packets. Accordingly, downlink VoIP packets experience larger queuing delays with the default access. Figure 1 shows the aggregated throughput performance of downstream TCP flows at the AP, with both PIFS access and default access. We can see that the PIFS access provides a better throughput performance than the default access. Note that collision between downlink voice packet transmissions and uplink packet transmissions never occurs in the PIFS access. As a result, the PIFS access can reduce channel waste due to the collision and hence TCP stations Controlled Access Phase (CAP). can use more channel resource. This is the reason why the TCP throughput with the PIFS access is a bit larger than that with the default access..3 JITTER PERFORMANCE COMPARISON The jitter 5 of VoIP is another important performance measure along with the delay. In order to evaluate the jitter performance of four different access schemes, namely, DCF, MDQ, default access, and PIFS access, we simulate with 1 or 1 VoIP sessions along with 5 downstream TCP flows. Figure 11 shows the jitter performance for both 1 and 1 VoIP session cases. We can imagine that there are two major factors, which increase the VoIP jitter in the.11 WLAN, namely, contention/collision with other stations and random delay inside the queue. First, when one VoIP session exists, i.e., Figure 11(a), schemes (i.e., default and PIFS accesses) demonstrate better jitter performances than DCF schemes (i.e., DCF and MDQ) thanks to their smaller channel access parameter values. The reason is explained as follows: with 1 VoIP session, TCP flows can use a large fraction of the total bandwidth, and hence more TCP stations contend for channel. In this situation, AC VO of schemes, which use small channel access parameter values, can reduce the contention with TCP stations and AC BE in the AP. Accordingly, the jitter becomes smaller. On the other hand, with 1 VoIP sessions in Figure 11(b), we find the result quite different from 1 VoIP session case. In this situation, CWmin[AC VO] value of schemes is not large enough for collision avoidance. Accordingly, many collisions can occur, thus increasing the jitter considerably. However, the jitter of downlink VoIP packets in PIFS access still remains small because downlink AC VO can perfectly avoid the contention with 5 Jitter is the variation (i.e., standard deviation) in the time between packet arrivals at destination because traffic is originally CBR. Submission 7

8 J. Yu and S. Choi 7 7 Down Up Down Up 5 5 Jitter (ms) 3 Jitter (ms) DCF MDQ Default Access PIFS Access DCF MDQ Default Access PIFS Access Access schemes Access schemes (a) 1 VoIP session (b) 1 VoIP sessions TCP stations and AC BE in AP. The reason why the jitter of downlink VoIP packets in all schemes except for PIFS access increases is because VoIP packet generation times of each VoIP sessions are randomized in our simulations, and hence a VoIP packet arriving at the AP queue experiences random queuing delay. From the jitter performance evaluation thus far, we can conclude that when there are a smaller number of VoIP sessions, the jitter performance of the is better than that of the DCF/MDQ while they perform about the same when there are many VoIP sessions. Figure 11: Jitter performances of four access schemes. COMPARISON OF MDQ AND WITH RATE DIVERSITY As described in Section 5, the performance anomaly can frequently occur in the real environment since most of WLAN devices implement the rate adaptation mechanism. In this section, we compare the VoIP performance of MDQ and when the TCP stations in a BSS use a lower data rate than VoIP stations. This, for example, represents a situation when TCP stations reduce their data rate due to the bad channel condition which can be known from the received signal strength or ACK timeout. In the simulations, we basically assume that the transmissions of the low-rate stations do not suffer from channel errors because they can adapt their transmission rates properly based on the channel conditions. We simulate with VoIP stations and 1 downstream TCP stations. All unicast data frames are transmitted at 11 Mbits/s between AP and stations except for three AP-TCP station pairs. In the three pairs, unicast data frames are transmitted at 1 Mbit/s. Figure 1 shows the voice packet delay distribution of three different schemes, i.e., without TXOP, with TXOP, and MDQ. The with TXOP scheme uses TXOP mechanism for only AC BE (i.e., TCP packet) in this scenario. We can see that VoIP delay performances of both without TXOP and MDQ schemes are Figure 1: VoIP delay performances of MDQ and with low-rate TCP stations severely degraded. This is due to the head-of-line blocking problem and the performance anomaly as discussed in Section 5. The size of a TCP data packet (15 bytes at the MAC SAP) is larger than a voice packet ( bytes at the MAC SAP). The time to finish the transmission of a TCP data packet at 1 Mbits/s is 1.79 ms, which is about eight times longer than the finishing time at 11 Mbits/s (i.e., 1.57 ms). The increased transmission time degrades the VoIP delay performance of both without TXOP and MDQ schemes. The reason why the downlink delay performance of without TXOP scheme is worse than MDQ is that TCP packets in get more service opportunities than MDQ as mentioned in Section.1. However, we can observe that the VoIP delay performance of with TXOP scheme basically does not degrade. Because the transmission time of TCP data frame is limited by TXOP Limit[AC BE] (in this scheme, we set the TXOP Limit[AC BE] to 1.57 ms, which is the transmission finishing time at 11 Mbits/s.), the voice packets does not suffer from the increased TCP packet transmission time. In the.11e specification [3], it is default not to use the TXOP mechanism for AC BE. However, we can see that we need ETT

9 Comparison of Modified Dual Queue and for VoIP over IEEE.11 WLAN to use TXOP mechanism for AC BE or AC BK to protect delay-sensitive traffic from AC BE and AC BK traffic..5 EFFECT OF VOICE PACKET ARRIVAL PATTERN ON CHANNEL UTILIZATION A VoIP application is not the saturated traffic source. Therefore, we can basically imagine that the number of concurrently active stations is small, and hence collisions hardly occur when several non-saturated VoIP stations exist and the sum of their required bandwidths does not exceed the capacity of the BSS. However, in Figure 13, we can observe that the number of concurrently contending stations and the collision among them depend on the pattern of the traffic arriving at the BSS. In this scenario, we simulate with only seven VoIP stations for both and MDQ, and measure the number of active stations which have more than one pending packet to transmit. Figure 13 shows four combinations of artificial traffic arrival patterns (Figure 1(a) (d)) to observe the impact of traffic arrival patterns. In the caption, U and D mean uplink (i.e., station to AP) and downlink (i.e., AP to station), respectively. E means even distribution of voice packet arrival times at the BSS while S means synchronized arrivals that voice packets from all VoIP traffic source arrive at almost the same time. For example, USDE means that times that voice packets from all VoIP traffic sources arrive at the MACs of VoIP stations are almost the same and the times that all downlink voice packets directed to the VoIP stations arrive at MACs of AP are evenly distributed. In case of UEDE and UEDS, the number of concurrently contending stations is one or zero. It means that there is basically no collision and no performance degradation. However, in case of USDS and USDE, the large number of stations contend for a long time. It means that the collision probability is high, thus degrading the performance. By the way, in all the sub-figures, especially, in Figure 13(c), we can observe that the time that more than one active station contend under MDQ scheme is longer than. It is due to larger CWmin and CWmax sizes of MDQ, i.e., legacy DCF. In MDQ scheme, stations choose larger backoff counter values than stations. It means that a DCF station should wait for longer time than a station. Moreover, when a collision occurs, the DCF stations involved with the collision double their contention window size, and then choose a new backoff counter value. Therefore, the waiting time gap between and DCF stations becomes larger. Figure 13(e) shows the average delay performance of and MDQ. We can observe that the delay performance of MDQ is worse than that of. It is also due to larger CWmin and CWmax sizes of MDQ as discussed above. Accordingly, we can conclude from Figure 13 that we need to make VoIP traffic arrival pattern as even as possible such as UEDE (ideal case) in order to prevent the waste of resource (i.e., channel time) via reducing collisions. We plan to develop a scheme to ensure this as a future work. 7 CONCLUSION In this paper, we have comparatively evaluated the modified dual queue (MDQ) scheme, based on the legacy.11 DCF, and the.11e in terms of their QoS provisioning capability. We have presented a modified version of the originally-proposed dual queue scheme by considering a practical implementation limitation. Under the access parameter adaptation rule, we also consider a PIFS access rule, which allows the AP to send its VoIP packets after a PIFS idle period. From extensive simulations considering the VoIP delay/jitter and TCP throughput, we find that the surely provides a better performance than the MDQ scheme thanks to the flexible channel access parameter control of the depending on the underlying network condition, e.g., the traffic load. However, the MDQ also provides a good performance, which is acceptable in most situations, and is comparable with that of the. Finally, we compared the performances of MDQ and when TCP stations transmit large packets at low data rate. The VoIP performance of MDQ is degraded due to its MAC HW queue which is a FIFO queue when TCP packet transmission time increases. On the other hand, can protect existing VoIP traffic from performance degradation thanks to its TXOP mechanism. Accordingly, we conclude that the MDQ scheme can be practically a good solution in order to provide QoS for VoIP services when the.11e is not available or where the HW upgrade is not desirable. However, we eventually need to use.11e MAC to solve some practical problems such as performance anomaly and utilize channel more efficiently. Our future work includes the development of the algorithm to adjust the optimal channel access parameters of the as well as the admission control algorithm for acceptable QoS provisioning depending on the network condition and applications in service. REFERENCES [1] IEEE Std , Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications, Reference number ISO/IEC -11:1999(E), IEEE Std..11, 1999 edition, [] IEEE.11b-1999, Supplement to Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications: Higher-speed Physical Layer Extension in the. GHz Band, [3] IEEE.11e/D13., Draft Supplement to Part 11: Wireless Medium Access Control (MAC) and physical layer (PHY) specifications: Medium Access Control (MAC) Enhancements for Quality of Service (QoS), January 5. [] IEEE Std..1d-199, Part 3: Media Access Control (MAC) bridges, ANSI/IEEE Std..1D, 199 edition, 199. [5] Stefan M, Sunghyun C, Guido H, Ole K, Bernhard W. Analysis of IEEE.11e for QoS Support in Wireless LANs. IEEE Wireless Communications Magazine 3; 1():-5. [] Sunghyun C, Javier P, Sai S, Stefan M. IEEE.11e Contention- Based Channel Access (EDCF) Performance Evaluation. in Proceedings of IEEE ICC 3, Alaska, USA, May 3. Submission 9

10 J. Yu and S. Choi MDQ (a) UEDE MDQ (b) UEDS MDQ (c) USDS MDQ (d) USDE (e) Average delay Figure 13: The number of contending STAs with different voice packet arrival patterns AUTHORS BIOGRAPHIES [7] Jeonggyun Y, Sunghyun C, Jaehwan L. Enhancement of VoIP over IEEE.11 WLAN via Dual Queue Strategy. in Proceedings of IEEE ICC, Paris, France, June. [] Javier P, Sai S. Impact of Frame Size, Number of Stations and Mobility on the Throughput Performance of IEEE.11e. in Proceedings of IEEE WCNC, Georgia, USA, March. [9] The Network Simulator - ns-. online link. [1] Youngkyu C, Jeongyeup P, Sunghyun C, Gowoon L, Jaehwan L, Hanwook J. Enhancement of a WLAN-Based Internet Service in Korea. in Proceedings of ACM WMASH 3, San Diego, USA, September 3. [11] Jouni M. Host AP driver for Intersil Prism/.5/3, online link. [1] Daniel C. Carrier Grade Voice over IP. McGraw-Hill, September. [13] Yang X. Enhanced DCF of IEEE.11e to Support QoS. in Proceedings of IEEE WCNC 3, March 3. [1] Yang X, Haizhon L, Sunghyun C. Protection and Guarantee for Voice and Video Traffic in IEEE.11e Wireless LANs. in Proceedings of IEEE INFOCOM, Hong Kong, March. [15] Martin H, Franck R, Gilles BS, Andrzej D. Performance Anomaly of.11b. in Proceedings of IEEE INFOCOM 3, March 3. [1] Illenia T, Sunghyun C. Temporal Fairness Provisioning in Multi- Rate Contention-Based.11e WLANs. in Proceedings of IEEE WoWMoM 5, Taormina, Italy, June 5. [17] Ad K, Leo M. WaveLAN-II: a high-performance Wireless LAN for the Unlicensed Band. Bell Labs Technical Journal 1997; (3): Jeonggyun Yu received a B.E. degree in School of Electronic Engineering from Korea University, Seoul, Korea in. He is currently working toward his Ph.D. degree in the School of Electrical Engineering at Seoul National University (SNU), Seoul, Korea. His research interests include QoS support, algorithm development, performance evaluation for wireless networks, in particular, IEEE.11 wireless local-area networks (WLANs). Sunghyun Choi is an assistant professor at the School of Electrical Engineering, Seoul National University (SNU), Seoul, Korea. Before joining SNU in September, he was with Philips Research USA, Briarcliff Manor, New York, USA as a Senior Member Research Staff and a project leader for three years. He received his B.S. (summa cum laude) and M.S. degrees in electrical engineering from Korea Advanced Institute of Science and Technology (KAIST) in 199 and 199, respectively, and received Ph.D. at the Department of Electrical Engineering and Computer Science, The University of Michigan, Ann Arbor in September, His current research interests are in the area of wireless/mobile networks. He authored/coauthored over technical papers and book chapters. He also holds four US patents and one Korea patent, and has several patents pending. He has served on the TPC of many conferences, and has been involved with many conference/workshop organizations. He is also an area editor of ACM SIGMOBILE Mobile Computing and Communications Review (MCR) and an editor of KICS Journal of Communications and Networks (JCN). Since year, he is an active participant and contributor of the IEEE.11 WLAN standardization committee. From to 7, his research on IEEE.11(e) WLAN QoS is supported by Korea Ministry of Science and Technology (MOST) under Young Scientist Award program. Dr. Choi was a recipient of the Korea Foundation for Advanced Studies (KFAS) Scholarship and the Korean Government Overseas Scholarship during and , respectively. He also received a Bronze award at Samsung Humantech Paper Contest in He is a senior member of IEEE, and a member of ACM, KICS, KISS, and IEEK. 1 ETT