Achieving Scalable Performance in Large-Scale IEEE Wireless Networks *

Transcription

1 chieving Scalable Performance in Large-Scale IEEE Wireless Networks * Ping Chung Ng, Soung Chang Liew, and Li Bin Jiang Department of Information Engineering, The Chinese University of Hong Kong {pcng3, soung, lbjiang3}@ie.cuhk.edu.hk bstract In large-scale wireless networks, interferences among nodes limit channel spatial re-use and are a main hurdle for scalable performance. There are two types of interferences: 1) physical interferences due to the receiver s inability to decode a signal when the powers received from other signals are large; and ) protocol interferences imposed by the specific multi-access protocol being used. This paper models interference types 1 and in terms of a set of inequality constraints for the IEEE CSM/C protocol. Based on the inequalities, we investigate the impact of the parameters basic-rate, data-rate, and physicalpreamble-rate, on channel spatial re-use. Regardless of the parameter settings, the total capacity in an network reaches a ceiling as the number of node n increases. We identify the fundamental causes for the non-scalable performance, and show that can be made to be scalable with a simple modification, achieving O(n) throughput without adaptive power control. We believe that this is the first paper that has demonstrated this. Keywords Wireless Networks, d hoc Networks, IEEE 80.11, Performance, nalysis, Capacity, Scalable Performance. I. INTRODUCTION driving force behind the popularity IEEE 80.11b/g wireless network is that the frequency band used is unregulated. nyone can set up a WLN quickly. This brings up a critical problem as deployments continue to spread - we may soon reach a point of radiation pollution in which densely populated wireless stations interfere with one another to cause severe performance degradations. The performance of such densely populated networks is non-scalable. The total capacity reaches a ceiling as the number of nodes increases. question is whether networks can be made to be scalable. To answer this question, we need to 1) identify the fundamental causes for the nonscalable performance; and ) to devise a strategy to overcome them. This paper does both. In a previous work of ours [1], two types of interferences have been isolated and examined separately: 1) physical interferences due to the receiver s inability to decode a signal when the powers received from other sources are large; and ) protocol interferences imposed by the specific multi-access protocol being used. This paper elaborates the impact of these interferences on scalability. * This work was sponsored by the reas of Excellence scheme established under the University Grant Committee of the Hong Kong Special dministrative Region, China (Project Number oe/e-01/99). The rest of this paper is organized as follows. In Section II, we review the model of physical constraints used in this paper. Section III derives the constraints imposed by due to its virtual and physical carrier-sensing mechanisms as well as its stop-and-wait RQ. Section IV examines the implications of the physical and protocol constraints around a link, and presents numerical results showing that there is a fundamental trade-off in networks that is preventing it from achieving scalable performance. Based on the insights obtained, we devise a scheme that make performance scalable in Section V. Section VI concludes this paper and presents promising future research arising from this work. II. PHYSICL CONSTRINTS To model physical interference, this paper adopts the protocol model in []. Consider two links, links 1 and. Let T i and R i be the transmitter and receiver of link i, respectively. The model in [] assumes link can interfere with link 1 if < ( 1 + ) and link 1 can interfere with link if (1) where is the location of node i, and > 0 is related to a i power margin. Simultaneous transmissions on links 1 and are not possible if either (1) or is true. Inequalities (1) and can be derived from a signal-to-interference requirement and typical signal power as a function of distance relationship [3]. lthough reference [] names the model a protocol model, it may not capture all interferences imposed by a specific protocol. In addition to (1) and, MC protocols may introduce additional protocol-specific constraints to further limit channel spatial re-use. III. IEEE PROTOCOL CONSTRINTS To avoid simultaneous transmissions that cause collisions, a distributed multi-access protocol can be used. multi-access protocol s function is (i) to avoid simultaneous transmissions that may lead to collisions under the physical constraints, while (ii) allowing simultaneous transmissions that are collision-free under the physical constraints. The MC protocol may fail to achieve (i) or (ii). We first derive the protocol constraints based on its carrier sensing mechanism. We assume the RTS/CTS access mode. RTS/CTS can be decoded if the distance of transmission

2 is less than the virtual carrier sensing range ( VCSRange ). Consider links 1 and, and suppose link 1 s transmission is already in progress when link has a packet to transmit. Link cannot transmit if any of the following inequalities is satisfied: R (5) R (6) If inequality were satisfied, would have received the RTS of T 1 and the NV contained in the RTS would prevent T from transmitting. Similarly, if inequality were satisfied, T would have received the CTS of R 1. In inequalities (5) and (6), R would have received the RTS of T 1 and CTS of R 1 respectively, and it would not reply to the RTS of if T sent one. Besides virtual carrier sensing, physical carrier sensing may also prevent simultaneous transmissions. For physical carrier sensing, the preamble of the PHY header is decoded. The length field in the PHY header informs the receiver of the duration of the payload that follows. When T or 1 T can decode the PHY preamble transmitted by each other, simultaneous transmissions will be prevented: < PCSRange (7) where PCSRange is the physical carrier-sensing range. Inequalities -(7) are related to simultaneous transmissions that are disallowed by through its carrier sensing mechanism. If none of -(7) is satisfied, carrier sensing cannot prevent simultaneous transmissions, but the simultaneous transmissions may not be successful if one of the following is true: R < PCSRange (8) < PCSRange (9) For (8), if T 1 transmits first followed by, then R will not attempt to receive the signal form. Similar argument applies for (9). In addition, the CK of one link may collide with the DT or CK of the other. Each atomic communication on a link i consists of a DT frame in the forward direction and an CK frame in the reverse link i in the other direction. So, simultaneous communications on two links i and j are interference-free if the link pairs i and j, i and j, i and j; and i and j are interference-free. Thus, collisions due to the RQ of could occur if one of the following is true: (10) (11) (1) (13) (14) (15) The relative size of PCSRange and VCSRange matters as far as which of the inequalities (1)-(15) are distinct. n ideal protocol will be one in which the protocol constraints overlap exactly with the physical constraints so that only (1) and are distinct. Unfortunately, that is generally not the case. IV. LOCL EFFECTS OF PHYSICL ND PROTOCOL CONSTRINTS UNDER VRIOUS TRNSMISSION RTES FOR DT, RTS/CTS/CK, ND PHY PREMBLES In this section, we evaluate the effectiveness of the protocol around a local link under various transmission-rate settings. We assume the use of 80.11b, which supports four data transmission rates, 1,, 5.5 and 11Mbps. The same methodology is applicable to 80.11g and 80.11a. TBLE I. TRNSMISSION RNGES FOR VRIOUS DT RTES. Data Rate (Mbps) TxRange (m) TBLE II. RELTIONSHIPS BETWEEN TxRange, VCSRange ND PCSRange UNDER VRIOUS TRNSMISSION RTES. PHY Preamble Rate Data Rate Basic Rate 1Mbps Mbps 1Mbps 1Mbps (1) --- Mbps 5.5Mbps 11Mbps 1Mbps Mbps 1Mbps Mbps 5.5Mbps 1Mbps Mbps 5.5Mbps 11Mbps --- (1) Table I shows the maximum transmission ranges (TxRange), adopted from [4], which assumes omni antennae. Table II shows the PCSRange, VCSRange, and transmission range (TxRange) under various transmission rates for PHY, RTS/CTS/ CK (i.e., basic rate) and DT (i.e., data rate). Typically, PHY _ rate basic _ rate data _ rate, implying TxRange VCSRane PCSRange. The settings in Table II can be grouped into four main cases: (1) TxRange = VCSRange = PCSRange; TxRange = PCSRange; TxRange < PCSRange; TxRange = VCSRange < PCSRange. Let us assume the use of the two-ray propagation model with the SIR requirement set to 10dB. From [5], the interference range due to physical constraints for R for link ( T, R ) is InRange d where d = T. That R is, there should not be another transmitter within the InRange

3 of R when T is transmitting to R. Note that InRange depends on the distance d. The VCSRange and PCSRange for the protocol constraints, on the other hand, are independent of d. Thus, may not approximate the physical constraints well. Case (1): For case (1), Fig. 1 shows the interference region, virtual carrier-sensing region and physical carrier-sensing region when T and R are separated by d = TxRange. simultaneous transmission by another node in the shaded region will cause physical interference at R but is allowed by the protocol. Figure 1. Effects of physical and protocol constraints when TxRange = VCSRange = PCSRange Consider link B inside the shaded region with T B in the shaded region. Transmitter T B (node 3) and receiver R B (node 4) cannot hear the transmission of link. transmission from node 3 to node 4 will cause a collision at R. In this case, fails to prevent collisions. Consider link B in the reverse direction in which node 4 is the transmitter and nodes 3 is the receiver. In this case, the CK of node 3 to node 4 will interfere with the DT from T to R (inequality (10)). gain, fails to prevent collision-causing simultaneous transmissions. The cross-lined region on the right side in Fig. 1 is covered by the protocol constraints but not the physical constraints. This implies simultaneous transmissions are allowed in the real physical situation, but are restricted by the protocol. Note that the smaller the d, the smaller is the shaded region; but the larger is the cross-lined region (see Fig. 3). This means that the ability for to allow proper simultaneous transmission decreases with its ability to disallow improper simultaneous transmissions. This is a fundamental deficiency of the current protocol that is preventing scalable performance. Case : For PCSRange(Mbps)=VCSRange(Mbps) or PCSRange(1Mbps) = VCSRange(1Mbps) > TxRange(11Mbps), the interference region is fully covered by the carrier-sensing region. This means that can successfully prevent collisions. However, it over-estimates the region in which simultaneous transmissions are not allowed, limiting channel spatial re-use. With the same PCSRange and VCSRange, if we have TxRange(Mbps) or TxRange(5.5Mbps), we again have the same scenario as in Case (1) in which the probability of failing to prevent collisions trade-off against the probability of preventing legitimate simultaneous transmissions. Case : For case, again, depending on the parameter setting, the interference region may or may not be fully covered by carrier sensing, and the same trade-off observed before applies. In addition, the fact that VCSRange ( Mbps) < PCSRange ( 1Mbps) means that inequalities (7)-(9) are not included in other inequalities and are distinct. In this case, the directionality of links matters. We introduce the straight dotted-line region (see Fig. ) in which if only the receiver (e.g., node 3 in the figure), but not the transmitter, is within the physical carrier-sensing range but outside the virtual carrier-sensing range of R, simultaneous transmissions from node 4 to node 3 and from T to R are still allowed by the protocol and they will succeed. However, if node 3 is transmitting to node 4 when T sends an RTS to R request for transmission, R will not reply to T because R has already heard the physical preamble of node 3. Thus, another receiver in the straight-dotted line region will not interfere with the transmission of link, but a transmitter may. Figure. Effects of physical and protocol constraints when TxRange < PCSRange Case : Discussion similar to the other cases applies here. Figure 3. reas of cross-lined and straight-dotted lined regions, and shaded region of case (1) versus d in typical settings for the four cases: (1) TxRange(Mbps) = VCSRange(Mbps) = PCSRange(Mbps)

4 TxRange(11Mbps) (Mbps) = PCSRange(Mbps) TxRange(11Mbps) (Mbps) < PCSRange(1Mbps) TxRange(11Mbps) = VCSRange(11Mbps) < PCSRange(1Mbps) Figure 3 shows the areas of cross-lined and straight-dotted lined regions against d in the four cases for typical parameter settings. s d decreases, the cross-lined and straight-dotted lined regions expand, and the difference between the network capacity under and the potential network capacity under only (1) and increases. This means the smaller the d, the more spatial-reusable resources are wasted with 80.11, a cause for non-scalable performance. Numerical Results To evaluate this non-scalable capacity, we generated random network topologies with various numbers of links in a circular plane of km radius. Figure 8 plots the NS [6] simulation results of the capacities and transport capacities [] as a function of the number of nodes in a circular plane of km radius. The total network capacities in all four cases reach a limit after 60 nodes. When node density increases, the throughput does not increase. In fact, the transport capacity goes down after 60 nodes. This is because links of shorter distance in are favored because shorter-distance links have a smaller carrier-sensing region, and the RTS/CTS of fewer other nodes can reach them to bar them transmitting. The average transport distance of a typical transmission is therefore smaller as number of nodes increases. whole area, no other transmissions are allowed. The theoretical results of [] based on physical constraints (1) and alone, however, indicate that the total network capacity should not reach a limit as the number of nodes increases, even without adaptive power control. This means that the standard fails to scale in a situation that is potentially scalable. We have focused on capacity as a performance measure in the above. In general, fairness is also a performance measure of interest. The degree of fairness achievable given a network topology can also be subjected to similar study as above. For example, reference [7] proposed an algorithm to allocate fair shares to links under the objective of max-min fairness. For details of the algorithm, we refer interested reader to [7]. V. SELECTIVE DISREGRD OF NVS For scalable performance, we need to find a way to make use of the carrier-sensing property of without having it limit spatial re-use. That is, we need to eliminate inequalities (9). We present a scheme below. We assume that VCSRange and PCSRange > Max InRange. This would be the case, for example, if the data rate is 11Mbps, basic and preamble rates are Mbps or 1Mbps. To eliminate the effect of the inequalities (9) on the protocol operation, each node could construct an interference graph (i-graph) [1] with i-edges due to (1), and (10) (15) incidents on the node s links. In an i-graph, an arrow-shape vertex represents a wireless link with the arrowhead pointing toward the receiver. There is an i-edge between vertices 1 and if (1), or (10)-(15) is satisfied. Figures 5 show an example of mapping a network topology to an i-graph. a) b) Figure 4. a) Total network capacity and b) total transport capacity verses number of nodes in a circle of km radius Both the virtual and physical carrier-sensing ranges are fixed. No other transmissions are allowed within the carrier sensing ranges of a transmitting link. The total network capacity does not scale with the number of nodes because once the carrier-sensing regions of existing transmissions cover the Figure 5. Mapping of a network topology a) to b) link-interference graph The construction of the i-graph could be based on the signal powers it hear from other links in the neighborhood. When transmitting or receiving, the nodes of a link can ignore the NV in the RTS/CTS that they hear from another nearby link if there is no such i-edge between the two links. The basic idea is to remove constraints to (9) in the protocol operation. The first algorithm below eliminates the effects of -(6); the second algorithm constructs the neighborhood i-graph for each node. The third is a feature of I. Selective Disregard of NV (SDN) (i) Each node a keeps an NV set, NVa = {NV(a, k)}, consisting of the NVs node a hears in its neighborhood, where k is the label for links. NVs heard are contained in the RTS and CTS sent by other nodes.

5 (ii) Suppose there is a packet for transmission from node a to node b (or from nodes b to a). Node a will send an RTS to b (or will reply a CTS to the RTS of b) if and only if for all k with NV(a, k) > 0, there is no i-edge between the following four vertex pairs: Vertex (a, b) and vertex k Vertex (a, b) and vertex k (link k in the reverse direction) Vertex (b, a) and vertex k Vertex (b, a) and vertex k In particular, NVs from links that do not have an i-edge with vertex (a, b) or (b, a) will be ignored II. Power Exchange for i-graph Construction (PE) ssumption: Same transmit power and receiver sensitivity for all nodes. (i) Each node a measures the powers from the RTS/CTS it hears in the neighborhood and keep it in a power set, Pa = {P(c, a)} where c is the node label of the sender. (ii) Each node a broadcasts Pa periodically using basic rate. (iii)node a constructs i-edges affecting it as follows. There is an i- edge between: Vertex (a, b) and vertex (c, d) if P(a, b) < K P(c, b) or P(c, d) < K P(a, d); Vertex (a, b) and vertex (d, c) if P(a, b) < K P(d, b) or P(d, c) < K P(a, c); Vertex (b, a) and vertex (c, d) if P(b, a) < K P(c, a) or P(c, d) < K P(b, d); Vertex (b, a) and vertex (d, c) if P(b, a) < K P(d, a) or P(d, c) < K P(b, c), where K > 1 should be sufficiently large (e.g., K = 10 for 10dB power margin) To see that PE does not miss any i-edge, without loss of generality, suppose that there should be an i-edge according to the first condition above, but node a misses it. If P(c, b) > P(a, b)/k, according to the assumption of VCSRange > Max InRange > TxRange, node b should have been able to hear the RTS/CTS sent by node c and therefore associate the measured received power with c. This is because P(a, b)/k > Po(TxRange)/K > Po(Max InRange) > Po(VCSRange), where Po(.) denotes the received power at a specific distance from the transmitter. If this is the case, in step (ii) node a should have received the values of P(a, b) and P(a, b) from node b with its periodic broadcast, and node a could not have missed the detection of this i-edge. If P(a, d) > P(c, d)/k, then P(d, a) = P(a, d) = P(c, d)/k > Po(TxRange)/K > Po(Max InRange) > Po(VCSRange). Therefore, P(c, d) and P(a, d) transmitted by node d periodically using the basic rate in step (ii) could not have been missed by node a, since the transmission range for the basic rate is VCSRange. III. Receiver Restart Mode (RS) In some commercial chips, there is a so-called re-start mode in the receiver design. If the receiver is in the midst of receiving a signal, another signal with sufficiently large power margin arrives, the receiver will switch to receive the new signal. This feature can be used to lift the limitations imposed by inequalities (7)-(9) to allow proper simultaneous transmissions. The power margin could be set to K in the PE algorithm above. If the power received from the transmitter is K times more than that from an interfering source, (7) and (9) will not come into effect. If it is less than K time, the link should not transmit simultaneously anyway by the restriction of (1),, (10)-(15). Detailed simulation of the algorithms and receiver operation on the NS platform will be explored in future work. Here, we present results from a simplified MTLB simulation experiment to study the potential advantage of SDN compared with regular Table III shows the average throughputs and transport capacities [] obtained from 500 simulation runs. In each run, a random network is created, with n nodes are randomly placed inside a circle of radius 150m. The transmit powers are kept constant as n varies. Each node could potentially transmit to all other nodes within its TxRange. That is, there is a link between a node and all other nodes within TxRange. The data rate is 11Mbps and the RTS/CTS/PHY transmission rate is Mbps. The corresponding VCSRange and TxRange ranges are respectively 437m and 3m, assuming omni antenna. With this setting, is effective in avoiding all potential collisions. But it is overly aggressive in disallowing simultaneous transmissions. SDN removes this defect. Given a random network topology, we then choose links to transmit either (i) in random order; (ii) in ascending order of the link distance; or (iii) in descending order of the link distance. For example, for (i), we randomly order the links in the network, and then we scan though the links in that order, each time allowing the link to transmit if inequalities (1),, (10)-(15) are not violated by other links already chosen to transmit. The operation of the contention-backoff mechanism in the standard is closest to (i). (ii) and (iii) simulates scenarios where priorities are given to small-hop links and large-hop links, respectively. t the end of each run, the throughput (i.e., number of transmitting links) and the transport capacity (i.e., sum of link distances of all transmitting links) are computed. s shown in Table III, the throughput and transport capacity quickly reach a limit as n increases, while those of SDN do not (for cases (i) and (ii)). Indeed, if priorities are given to links of small hops, SDN performance scales very well with n. One could argue intuitively that for very large n, the small-hop SDN throughput should be of order n + f(n), where f(n) is an increasing function of n, if the boundary effect of the area containing the nodes can be ignored. The corresponding transport capacity should be of order n + f(n) / n. To see this, suppose that the circle in the above simulation is a small area a within a large expanse of area. Given a certain node distribution in, suppose we map the position of node i, i to i = k*i, where k < 1. Suppose we also reduce the TxRange to k*txrange. The nodes in area a are then mapped to an area of a. However, the scaling does not affect the form of inequalities (1),, (10)-(15) and the throughput of the nodes within a is the same as the throughput within a previously. For the original area a, the throughput has increased by a factor of 1/k. So does the number of nodes in a. Now this is assuming TxRange has been reduced to k*txrange.

6 If this reduction had not occur and priority were given to link of small-hop as in (ii), then all the links with distances smaller than k*txrange would be considered for transmission first according to (ii). By the time we scan through all these links, the throughput in a would have increased by 1/k. The additional throughput from links with distance larger than k*txrange are icing on the cake if we can pack their transmissions in. It is also not difficult to understand the non-scalability of intuitively. The areas of the exclusion regions, a concept introduced in [] and [8], do not decrease as the node density increases. If we estimate the area of the exclusion region of a link as π (VCSRange/) and divide the area of the circular plane by this value, we get a value of 16.4, which is slightly higher than the throughput result obtained in Table III, possibly due to the slack in the packing problem. One could of course use power control and decrease the transmit power as n increases. Current standard and products, however, do not have power control. Indeed, if power control is considered, the i-edges in SDN could be decreased for additional performance improvement. That will be an interesting subject for future study. Before leaving this section, we note that SDN could also be modified to capture and work with the physical interferences from multiple nodes beyond the pair-wise physical interference model []. in networks to scale as node density increases, without the need for adaptive power control. REFERENCES [1] P. C. Ng, S. C. Liew, L. B. Jiang, Performance Evaluation Framework for IEEE d-hoc Networks, CM Workshop on Performance Evaluation of Wireless d Hoc, Sensor, and Ubiquitous Networks (PE-WSUN 04), Venice, Italy, Oct. 004 [] P. Gupta, P. R. Kumar, The Capacity of Wireless Network, IEEE Trans. on Information Theory, Vol. 46, No., pp , Mar [3] T. Rappaport, Wireless Communications: Principles and Practice, Prentice Hall, New Jersey, 00. [4] S. Bansal, R. Shorey,.. Kherani, Performance of TCP and UDP Protocols in Multi-Hop Multi-Rate Wireless Networks, IEEE WCNC 04, pril 004. [5] K. u, M. Gerla, S. Bae, How Effective is the IEEE RTS/CTS Handshake in d Hoc Networks?, IEEE GLOBECOM '0, Vol. 1, pp. 17-1, Nov. 00. [6] The Network Simulator ns, [7]. L. Huang, B. Bensaou, On Max-min Fairness and Scheduling in Wireless d-hoc Networks: nalytical Framework and Implementation, CM MobiHoc 01, pp. 1-31, Oct [8]. garwal, P. R. Kumar, Capacity bounds for ad-hoc and hybrid wireless networks. CM SIGCOMM Computer Communication Review, pp , Vol. 34, No. 3, July 004 Table III. SIMULTION RESULTS FOR ND SDN WITH RNDOM / SMLL-HOP/ LRGE-HOP TRNSMISSION PRIORITIES: T = THROUGHPUT (IN 11 MBPS) ND C = TRNSPORT CPCITY (IN 11 MBPS*KM). n T (80.11) 7./8.0/5.0 1/14/11 14/15/1 14/14/1 T (SDN) 9./11/7.0 18/5/11 6/5/13 34/93/13 C (80.11) 1.1/.66/ /.79/.0 1.9/.63/.6 1.9/.41/.5 C (SDN) 1.4/1.0/1.3./1.7/.3 3.0/.8/.9 3.6/3.6/.7 VI. CONCLUSION This paper has formulated a general framework to study the limits against spatial reuse imposed by multi-access protocols. The framework consists of isolating intrinsic physical interferences from protocol interferences imposed by the multiaccess protocol. By identifying the causes for the extraneous protocol constraints, we could devise strategies to overcome them. We have applied the framework to study the multiaccess protocol. Specifically, we have examined the local behavior around a link and global behavior of the overall network, and conclude that there is a fundamental trade-off between the protocol s ability to prevent collisions and its ability to allow legitimate spatial re-use. This trade-off cannot be removed by simply adjusting the settings for the transmission rates of RTS/CTS, DT and CK. The key to overcoming the fundamental limitations is to remove the extraneous constraints imposed by the protocol. We have demonstrated this with a scheme that allows the capacity