Remarks On Per-flow Differentiation In IEEE

Remarks On Per-flow Differentiation In IEEE 82.11 Imad Aad and Claude Castelluccia PLANETE project, INRIA Rhône-Alpes ZIRST - 655, Avenue de l Europe - Montbonnot. 38334 Saint Ismier Cedex - France [imad.aad, claude.castelluccia]@inrialpes.fr, http://www.inrialpes.fr/planete.html ABSTRACT Service differentiation at the MAC sub-layer is a recent research topic which is getting more attention as wireless Internet is becoming a reality. Our previous work [1] introduced some differentiation mechanisms for IEEE 82.11 and showed the need for per-flow differentiation mechanisms. In this paper we extend the perhost differentiation mechanisms to per-flow differentiation mechanisms, apply them to UDP and TCP flows respectively, and present some simulation results. We show that closed-loop traffic flows such as TCP above differentiated MAC sub-layers reduce the differentiation effect of the proposed mechanisms. In contrast open-loop traffic flows such as UDP tend to have a better behavior. We analyze these issues and draw general conclusions for further differentiation design. 1 INTRODUCTION Wireless networks are being increasingly deployed to extend wired networks to mobile users. Therefore the network architectures should also be extended to wireless parts: Differentiated services (e.g. DiffServ) are proposed to provide users with different quality of service (QoS) parameters, such as throughputs, delays and jitters. However, DiffServ operates at the network layer, regardless of the underlying infrastructure. Due to variable channel conditions, wireless environments do not support DiffServ optimally (e.g. data rate proportions available on the wireless link do not match the desired proportions on the network layer) and therefore should be adapted accordingly. Several works have been done for this kind of QoS support [2]. In [1], we proposed and analyzed three different differentiation mechanisms for IEEE 82.11 wireless networks: one based on Backoff differentiation, the other on DIFS differentiation and the last on maximum frame length differentiation. With each of these mechanisms, we used User Datagram Protocol (UDP) and Transport Control Protocol (TCP) traffic flows separately. The differentiation effect was clear when applied to UDP flows, however, TCP flows showed a major problem due to the feedback s (ACK) priority. This closed-loop feedback causes a shared receiver to control the flow priorities. Therefore, when two or more TCP senders share the same receiver, they all receive the TCP-ACKs with the same priority (the receiver priority). This tends to reduce the differentiation effect. Furthermore, if the shared receiver is slow, the observed relative priority is also reduced. This motivated the use of per-flow differentiation where a shared node uses different priorities for different flows. In this paper we investigate and analyze the use of per-flow differentiation using NS (Network Simulator) [3]. Section 2 briefly describes the IEEE 82.11 wireless LANs MAC protocol. Section 3 recalls the differentiation mechanisms we defined in [1] and introduces a new one. In section 4 we extend some of these differentiation mechanisms to per-flow ones and analyze them. Last, section 5 concludes this paper. 2 IEEE 82.11 In this section we present the IEEE 82.11, a widely deployed wireless LAN protocol, which is essentially based on CSMA/CA. IEEE 82.11 works in two different modes: Polling coordination function (PCF) and distributed coordination function (DCF). The former, PCF, consists of an access point (AP) synchronously polling the wireless terminals (WT) to transmit their packets, hence there is no contention or collision problem. The latter function, DCF is an asynchronous data transmission function, totally distributed, which best suits delay insensitive data. Each WT gets an equal share of the channel through contention, i.e. a WT contends for channel use before each frame waiting for transmission. The basic scheme for DCF is Carrier Sense Multiple Access (CSMA)[4]. This protocol has two variants: Collision Detection (CSMA/CD) and Collision Avoidance (CSMA/CA). A collision can be caused by two or more stations using the same channel 1 at the same time after waiting for the channel to become idle, or (in wireless networks) by two or more hidden terminals 2 [5] transmitting simultaneously. CSMA/CD is used in Ethernet (IEEE 82.3) wired networks. Whenever a node detects that the signal it is transmitting is different from the one on the channel, it aborts transmission, saving useless collision time. This mechanism is not possible in wireless communications because a WT cannot listen to the channel while it is transmitting, due to the big difference between transmitted and received power levels. To deal with this problem, the sender should wait for an acknowledgment (ACK) from the receiver after each frame transmission, as shown in 1 On the physical layer, in spread spectrum technologies, a channel is the pseudo-random sequence used to spread data. 2 Hidden terminals are terminals which cannot hear each other.

Fig. 1. Source axis shows the data transmitted by the source. The destination replies with an ACK, shown on the Destination axis. The third axis shows the network status, as seen by Other WTs. Note that transmission delays are not shown. The Interframe Spacings DIFS and SIFS will be explained later in this section. If no ACK is returned, a collision may have occurred and the frame is retransmitted. This technique may waste a lot of time if frames are long, keeping the transmission going on while collision is taking place (caused by a hidden terminal for example). To solve the hidden terminal problem, an optional RTS/CTS (Request To Send / Clear To Send) scheme is used in addition to the previous basic scheme: a station sends an RTS before each frame transmission to reserve the channel. Note that a collision of RTS frames (2 octets) is less severe and less probable than a collision of data frames (up to 2346 octets). The destination replies with a CTS if it is ready to receive and the channel is reserved for the packet transmission duration. When the source receives the CTS, it starts transmitting its frame, being sure that the channel is reserved for itself during all the frame transmission duration. All other WTs in the BSS (base station sub-system) update their Network Allocation Vector (NAV) whenever they hear an RTS, a CTS or a data frame. NAV is used for virtual carrier sensing, as detailed in the next paragraph. Not all packet types have the same priority. For example, ACK packets should have priority over RTS or data frames. This is done by assigning to each packet type a different Inter Frame Spacing (IFS), after the channel turns idle, during which no packet can be transmitted. In DCF two IFSs are used: Short IFS (SIFS) and DCF IFS (DIFS), where SIFS is shorter than DIFS (See Fig. 1). As a result, if an ACK (assigned with SIFS) and a new data packet (assigned with DIFS) are waiting simultaneously for the channel to become idle, the ACK will be transmitted before the new data packet (the first has to wait SIFS whereas the data has to wait at least DIFS). Carrier sensing can be performed on both physical and MAC layers. On the physical layer, physical carrier sensing is done by sensing any channel activity caused by other sources. On the MAC sub-layer, virtual carrier sensing can be done by updating a local NAV with the value of other terminals transmission duration. This duration is declared in data, RTS and CTS frames. Using the NAV, a WT s MAC sub-layer knows when the current transmission ends. NAV is updated upon hearing an RTS from the sender and/or a CTS from the receiver, so the hidden node problem is avoided. The collision avoidance part of CSMA/CA consists of avoiding packet transmission right after the channel is sensed idle for DIFS time, so it does not collide with other waiting packets. Instead, a WT with a packet ready to be transmitted waits the channel to become idle for DIFS time, then it waits for an additional random time (backoff time), after which the packet is transmitted, as shown in Fig. 1. Collision avoidance is applied to data packets in the basic scheme, and on RTS packets in the RTS/CTS scheme. The backoff time of each WT is decreased as Source (Tx) Destination (Tx) Other DIFS Data SIFS NAV ACK Defer access = NAV+DIFS DIFS Figure 1: Basic access scheme. Time Contention Window Backoff long as the channel is idle (during the so called contention window, CW). When the channel is busy, backoff time is freezed. When backoff time reaches zero, the WT transmits its frame. If the packet collides with another frame (or RTS), the WT times out waiting for the ACK (or the CTS) and computes a new random backoff time with a higher range to retransmit the packet with lower collision probability. This range increases exponentially as 2 4+i 1 where i (initially equal to 1) is the transmission attempt number. Therefore, the backoff time equation is Backoff_time = 2 4+i rand() 1 Slot_time, where Slot_time is function of physical layer parameters, and rand() is a random function with a uniform distribution in [,1]. There is a higher limit for i, above which the random range remains the same. The packet is dropped after a given number of retransmissions. With this mechanism, all WTs have equal probabilities to access the channel and thus share it equally. 3 DIFFERENTIATION MECHANISMS As we mentioned in the introduction, [1] describes and analyzes three MAC differentiation mechanisms: Backoff differentiation, DIFS differentiation and maximum frame length differentiation. With each of these mechanisms we analyze UDP and TCP traffic flows behavior separately. The simulation topology, using NS, was simple: Three WTs are placed around an AP, all using DCF. These WTs send their packets at full data rate to a sink node, wire connected to the AP. No possible congestion is possible on the wired link. The major observations are briefly described in the following subsections. 3.1 Backoff differentiation The main idea in Backoff differentiation is to assign different Backoff increase factors to different WTs. Therefore the Backoff sequence for W T j is given by Backoff_time = P 4+i j rand() 1 Slot_time. WTs with high P j backoff more than WTs with small P j and get, on the average, less chance to access the channel before the others, leading to lower data rates. When using UDP flows, the data rate proportions amongst WTs are equal to the probabilities of accessing the channel before other WTs (i.e. choosing a lower backoff value), computed over all CW value combinations, each combination weighted by its occurrence probability. However, when using TCP flows, no differentiation effects were observed and the data rates were almost the

same, independent of P j. The following explanation can be given: The AP is sending back the TCP-ACKs with the same priority, the AP s one, to all WTs. This factor surely reduces the differentiation effect. Whether high or low, a common P j used to return different destination TCP-ACKs reduces the differentiation between the closed-loop flows in comparison to UDP flows differentiation. If the AP is relatively slow (i.e. high P j ) a secondary effect exists: TCP-ACKs from the AP will be transmitted at a low rate. This would reduce the number of newly generated TCP packets at the WTs (that wait for the TCP-ACKs) and less collisions would be observed. Since the Backoff differentiation mechanism only works if collisions occur, differentiation is reduced. Differentiation effect increases as the AP speed increases (i.e. when P j decreases). 3.2 DIFS differentiation DIFS differentiation consists of assigning different DIF S j values to different W T j, therefore some WTs will have to wait longer than others before trying to transmit. This mechanism provides a wide range of differentiation: WTs can equally share the available data rate by choosing equal DIF S j, or they can have absolute priorities over each other if, for instance, DIF S j is greater than DIF S k +CW k. In the last case, W T j has no chance to transmit before W T k because it cannot choose a random transmission time shorter than the one chosen by W T k. The main observations on TCP flows differentiation are: Data rates are much more stable than with Backoff differentiation. Like in Backoff differentiation, the AP uses equal DIFSs for sending its TCP-ACKs to different WTs. Therefore the shared node priority problem still exists. However, since the mechanism is not based on Backoff, the secondary effect of the slow AP, which is avoiding collisions, does not exist. Therefore, the differentiation obtained is slightly better than with Backoff differentiation. 3.3 Maximum frame length differentiation This differentiation mechanism is rather simple. Different priority WTs are allowed to transmit frames with different maximum frame sizes. The WTs with high priority can send larger frames than WTs with low priority. This showed good differentiation results for UDP and TCP flows. Data rate shares are proportional to the assigned maximum frame sizes. The major drawback of this approach is that UDP packet sizes are controlled by the application, so differentiation is not really controlled by the MAC sub-layer. 3.4 CW min differentiation Working on Backoff differentiation led us to the fourth differentiation mechanism, not explored in [1], the CW min differentiation. The main motivation is that, with a small number of WTs contending to access the channel, CW values are at their minimum value (CW min ) most of the time. Therefore, a Backoff differentiation mechanism won t be applied correctly as the CWs are rarely increasing, and high CW values are rarely used. This led us to differentiate the most utilized CWs: CW min. The simulation scenario is the following: W T 1 starts transmitting at second 5, then W T 2 starts at second 1, then W T 3 starts at second 15. Simulation ends at second 25. Packet sizes are 11 bytes long, sent at.5 second intervals when using UDP flows. Results for both UDP and TCP flows are shown in Fig. 2 to 5 for comparison convenience. The sets of values shown as w/x/y/z indicate the values of CW min for AP/W T 1 /W T 2 /W T 3 respectively. When we use TCP flows, with 31/35/5/65 CW min values (Fig. 2), there is no noticeable differentiation effect visible. This is due to the slow TCP-ACKs transmissions by the AP. In fact, the AP uses a CW min which is close to that of W T 1. As each WT has to wait for a TCP-ACK before starting a new transmission, a slow AP makes each of the closed-loop flows much slower and the different CW min values assigned to the WTs do not have any real effect. When we use a faster AP, with CW min = 2 (Fig. 4), the TCP-ACKs are sent much faster, so the different WTs do not have to wait as in Fig. 2 before transmitting, and the different CW min values they have show much more effect. On the other hand, when we use UDP flows (Fig. 3 and 5), accelerating the AP from CW min = 31 to CW min = 2 has absolutely no effect on the differentiation scheme. This is obviously true as WTs do not wait for any feedback from the AP, so the data rate shares remain the same, whatever is the CW min value the AP has. Comparing the figures vertically (compare Fig. 2 to 3, and Fig. 4 to 5), shows that, for the same sets of CW min values, UDP flows get more differentiation effect than TCP flows. Consequently, the data rate shares that UDP flows get can be thought of as the maximum data rate shares that TCP flows can get, when we accelerate the AP indefinitely. 4 PER-FLOW DIFFERENTIATION All the differentiation mechanisms described in the previous section suffered one major common problem when trying to differentiate TCP flows: The AP always uses its own priority 3 to send back TCP-ACKs to different WTs. This reduces the differentiation effect. In fact, when W T 1 and W T 2 want to transmit a packet, they wait on average t 1 and t 2 respectively before transmitting their packets, the resulting data rate ratio is proportional to t 2 /t 1. However, when we use TCP flows, the TCP-ACK transmission introduces an additional delay t to the 3 Either the DIFS, the backoff factor or the CW min. The maximum frame size differentiation does not really apply to TCP-ACKs

1 1 1 1 1 1 1 1 5 1 15 2 25 5 1 15 2 25 Figure 2: CW min diff.: 31/35/5/65, using TCP flows Figure 4: CW min diff.: 2/35/5/65, Using TCP flows 1 1 1 1 1 1 1 1 5 1 15 2 25 5 1 15 2 25 Figure 3: CW min diff.: 31/35/5/65, using UDP flows closed-loop flows. So the data rate ratio between W T 1 and W T 2 becomes proportional to (t 2 + t )/(t 1 + t ). When the AP is slow, t is high, which reduces the fraction (t 2 + t )/(t 1 + t ) considerably. If we try to compensate the differentiation loss by increasing t 1 and t 2, we loose efficiency by making W T 1 and W T 2 wait more than necessary. Therefore, the AP must be as fast as possible, so t is as small as possible, but t 1 and t 2 still have to compensate the small differentiation loss. Another alternative is to make the AP use different priorities for different destinations, i.e. per-flow differentiation. So, instead of waiting a fixed time t before transmitting a packet, the AP should wait t 1, t 2 or t 3 according to the destination of the packet. The resulting data rate ratio between W T 1 and W T 2 becomes (t 2 + t 2 )/(t 1 + t 1 ) which is equal to t 2 /t 1 when t 2 /t 1 = t 2 /t 1, so no differentiation loss needs to be compensated. In other words, optimally, the AP should send back the TCP-ACKs with different priorities, proportional to the priorities of the destinations. In the following subsections, we are going to apply the perflow differentiation to the mechanisms briefly described in section 3. The waiting times t j and t j defined above would designate the expected waiting time when we apply: different DIF S j in DIFS differentiation. different CW minj values in CW min differentiation, Figure 5: CW min diff.: 2/35/5/65, using UDP flows and different backoff increase factors P j in Backoff differentiation. Among these mechanisms, CW min and DIFS differentiations show similar behavior when the AP uses per-flow differentiation. However, backoff differentiation do not show much effect due to the low number of collisions. Therefore the effect introduced by per-flow differentiation is even less visible. In the following we only show DIFS per-flow differentiation. The same reasoning applies to the other mechanisms. 4.1 Per-flow DIFS differentiation We ran ten simulations with per-flow DIFS differentiation and we show the quantitative results in this subsection. Table 1 shows the values of DIFS the AP and the WTs use respectively: The column AP j shows the DIF S j values the AP uses to send the TCP-ACKs to W T j, while the column W T j shows the DIF S j values each W T j uses to send its data packet. The last line describes the observations made on data rate differentiation using each set of DIFS values. Set I shows the case where we wanted to check out the elementary effect of data rate differentiation due to the AP only. Simulation showed no differentiation at all. Two possible reasons:

Table 1: Per-flow DIFS differentiation with TCP Set I Set II Set III Set IV Set V j AP j W T j AP j W T j AP j W T j AP j W T j AP j W T j 1 5 9 5 5 5 5 9 5 5 5 2 9 9 9 9 5 9 9 9 9 5 3 9 9 9 9 5 9 9 9 9 5 Obs. Bad diff. Bad diff. Good diff. Bad diff. Bad diff. Set VI Set VII Set VIII Set IX Set X j AP j W T j AP j W T j AP j W T j AP j W T j AP j W T j 1 5 5 3 5 7 5 3 5 3 5 2 5 15 5 15 5 15 5 12 5 1 3 5 15 7 15 3 15 7 15 7 15 In fact, AP 1 is relatively fast, but AP 2 and AP 3 are slow. As the TCP-ACKs are sent serially by the AP, the fast TCP-ACK still have to wait for the slow one and the overall AP speed is still slow. W T 1 TCP data are sent slowly. Accelerating the corresponding AP 1 would not accelerate the loop flow unless we accelerate W T 1. Set II eliminates the second possible reason: Even though we accelerated W T 1, no differentiation took place, the AP is still globally slow, fast packets still have to be queued with slow ones to be transmitted. Set III shows the situation where the WTs require differentiation, and the AP is sufficiently fast. The resulting data rate differentiation is good. W T 2 and W T 3 get equal data rates, lower than the data rate W T 1 gets. Set IV and Set V provide redundant information: the AP is slow in Set IV, therefore the differentiated W T j do not result in differentiated data rates. In Set V, the differentiated AP does not result in flow differentiation because it still is globally slow. Sets VI to X show the AP DIFS values, where the AP is always fast, so we can observe the per-flow differentiation effect. Set VI shows a situation similar to the one in Set III, we replaced the DIFS = 9 by DIFS = 15, so the differentiation is bigger. Keeping the AP with the same average speed (DIFS = 5), we changed the AP j values to 3/5/7 as in Set VII. Even though the AP average speed is the same 4 as in Set VI, we observe more differentiation because the AP sends the TCP-ACKs with a speed proportional to the destination speed. We keep the same AP average speed, but we inverse them, 7/5/3, as in Set VIII. The resulting differentiation is lower than in Set VII, because the AP uses differentiated DIFS values, but in the wrong order. Sets VII, IX and X show the case where W T 2 DIFS decreases from 15, 12 then to 1. All the rest of AP j, W T 1 and W T 3 remain the same. The throughput ratio (differentiation) between W T 1 and W T 2 4 This is not really the same average, 5. W T 1 has a short DIFS, therefore it sends more packets than the other two WTs, so the AP 1 = 3 is more used than the AP 2 = 5 and AP 3 = 7, and the resulting AP average DIFS is lower than 5. decreases also. One interesting observation is that, in Set X, AP 1 and AP 2 has very different values, as in W T 1 and W T 2. However, the throughputs are almost equal, which was not the case in Set IX. In fact, when we reduced W T 2 from 12 to 1, we invoked more utilization of AP 2 = 5, which led to a slower AP than in Set XI, so the differentiation is less visible. We should also note that the throughput of W T 3 remained almost the same in all the three simulation sets, VII, IX and X. Mathematical models which better describe the differentiation behavior function of the differentiation parameters are under construction. 4.2 MAC sub-layers with per-priority queues. In the previous subsection we saw that all packets are put in the same queue, independent of their priority. This introduced mutual interferences between priorities: When the AP serves a low priority flow, the AP global speed depends on the utilization of this flow. If it is highly used, the AP gets slow, and differentiation gets lower. A possible solution is to assign to each priority (or to each WT) a different queue. Simulation showed a total independence between priorities: Even if a low priority flow exists, it won t slow down the AP (the shared node), and differentiation is much more clear. Note that when using this approach with CW min differentiation, the shared node (e.g. the AP in our scenarios) will be avoiding collisions less than other WTs do. In fact, when a single queue per MAC sub-layer is used, we just have one packet/node contending to access the channel, during a CW period. However, when a node uses n queues (for n TCP connections), we have n packets per CW period, as if in a shared node, collision avoidance decreases as the number of connection increases. 5 CONCLUSION AND FUTURE WORK In this paper we first described previously introduced differentiation mechanisms and introduced a new one: CW min differentiation. Within these mechanisms, TCP flows do not perform very well due to the TCP-ACKs the access point (AP) sends back to the WTs with equal priorities. We then explored per-flow differentiation on the MAC sub-layer, and we applied it to the previously introduced differentiation mechanisms. We drew several remarks concerning the shared node transmission speed and the mutual interference between flows which

motivated the use of MAC sub-layers with per-priority queues. The remarks made above are not restricted to the differentiation mechanisms we simulated. They can be generalized to any shared node trying to differentiate between its outgoing flows. REFERENCES [1] Imad Aad and Claude Castelluccia, Differentiation mechanisms for IEEE 82.11, in Proceedings of IEEE Infocom 21, Anchorage - Alaska, April 21. [2] Michael Barry, Andrew T. Campbell, and Andras Veres, Distributed control algorithms for service differentiation in wireless packet networks, in Proceedings of IEEE Infocom 21, Anchorage - Alaska, April 21. [3] Network simulator, http://www.isi.edu/nsnam/ns/,. [4] L. Kleinrock and F. A. Tobagi, Packet Switching in Radio Channels: Part 1: CSMA Modes and Their Throughput Delay Characteristics., IEEE Trans. on Comm. COM-23, 1975. [5] Bob O Hara and Al Petrick, IEEE 82.11 handbook. A designer s companion., IEEE Press, 1999.