TCP with delayed ack for wireless networks

Transcription

1 Available online at Ad Hoc Networks 6 (8) with delayed ack for wireless networks Jiwei Chen *, Mario Gerla, Yeng Zhong Lee, M.Y. Sanadidi University of California, Los Angeles, CA 995, United States Received 5 May 6; received in revised form 25 October 7; accepted 3 October 7 Available online 6 November 7 Abstract This paper studies the performance with delayed ack in wireless networks (including ad hoc and WLANs) which use IEEE MAC protocol as the underlying medium access control. Our analysis and simulations show that throughput does not always benefit from an unrestricted delay policy. In fact, for a given topology and flow pattern, there exists an optimal delay window size at the receiver that produces best throughput. If the window is set too small, the receiver generates too many acks and causes channel contention; on the other hand, if the window is set too high, the bursty transmission at the sender triggered by large cumulative acks will induce interference and packet losses, thus degrading the throughout. In wireless networks, packet losses are also related to the length of path; when traveling through a longer path, a packet is more likely to suffer interference. Therefore, path length is an important factor to consider when choosing appropriate delay window sizes. In this paper, we first propose an adaptive delayed ack mechanism which is suitable for ad hoc networks, then we propose a more general adaptive delayed ack scheme for ad hoc and hybrid networks. The simulation results show that our schemes can effectively improve throughput by up to 25% in static networks, and provide more significant gain in mobile networks. The proposed schemes are simple and easy to deploy. The real testbed experiments are also presented to verify our approaches. Furthermore, a simple and effective receiver-side probe and detection is proposed to improve friendliness between the standard and our proposed with adaptive delayed ack. Ó 7 Elsevier B.V. All rights reserved. Keywords: ; Congestion control; Adaptive delayed ack; Wireless medium access/contention; Friendliness 1. Introduction The Transport Control Protocol () is the most widely used reliable transport protocol over the Internet. was originally designed for wired links where buffer overflows account for most packet losses. However, in multihop ad hoc wireless * Corresponding author. addresses: cjw@ee.ucla.edu (J. Chen), gerla@cs. ucla.edu (M. Gerla), yenglee@cs.ucla.edu (Y.Z. Lee), medy@cs. ucla.edu (M.Y. Sanadidi). networks, several other inherent factors attribute to the performance deterioration, including unpredictable channel errors, medium access contention complicated by hidden/exposed terminal problems, and frequent route breakages caused by node mobility. All these factors pose great challenges to the design of protocols to provide efficient and reliable end-to-end communications. Many researchers have made valiant effort to propose various methods to make survive in such challenging environments. In this paper, we focus on studying the effect of delayed acks on /$ - see front matter Ó 7 Elsevier B.V. All rights reserved. doi:1.116/j.adhoc.7.1.4

2 J. Chen et al. / Ad Hoc Networks 6 (8) performance. Then based on our findings, we propose adaptive schemes that address the aforementioned factors effectively. In standard the receiver generates one ack for each data packet, or more popularly, two inorder data packets with the standard delayed ack option [1]. This mechanism works well in wired networks. In multihop wireless networks, however, this mechanism can be further improved due to: Since the data and ack packets usually take the same route (or spatially close routes), they interfere with each other. The interference increases with the number of acks generated. Generating acks wastes scarce wireless resources. Though acks are essential to provide reliability, generating more acks than necessary is not desirable in wireless networks. Ideally, the receiver should generate minimal number of acks required for reliable data recovery. Recently, the delayed ack strategy has been studied to improve performance [2,3]. However, this field is not fully exploited and many issues remain unsolved. Some important questions include: what is the relation between delayed acks and performance, how to choose the proper delay window (the number of in-order data packets to be waited for before generating an ack) for, and how to deal with the loss of acks in multihop wireless networks. In this paper we carry out a systematic study to understand the effect of delayed acks on over wireless links. We investigate performance and delayed ack related packet loss characteristics, under various wireless network scenarios and flow patterns. Our objective is to clearly identify the relationship between throughput and delayed ack over multihop wireless links. We reveal several interesting findings, which are tremendously beneficial to deeply understand behavior in wireless multihop networks, and to design enhanced protocols. First, we found that does not always benefit from arbitrary delaying of acks. In fact, for a given network topology and flow pattern, there exists an optimal delay window size at the receiver, at which achieves maximum throughput. Second, since does not guarantee collision free packet transmission, a packet is more likely to be interfered with when going through a long path. If a large delay window is used over a long path, it causes large bursts of packets in transit, and this in turn causes severe packet interference with each other. When packet loss rate becomes high, the benefit of delaying acks, via reducing ack packets, disappears. Consequently, performance degrades after reaching the peak at the optimal delay window size. Armed with the deep understanding of delayed ack impact on performance over multihop wireless links, we propose an adaptive scheme based on the hop count of a path. The basic idea behind this scheme is, for a short path where interference problem is minimal, we delay the ack as much as possible to maximally improve throughput; while for a long path, we apply an appropriate delay window size restriction to avoid high packet loss to achieve optimal performance. Furthermore, we propose an end-to-end delay based scheme tailored for hybrid wired/wireless networks. These two schemes can be deployed separately, but are also compatible with each other to provide better performance in heterogenous networks. It is also worth noting that our proposed schemes only reside in transport layers at the end hosts, thus no modification on intermediate nodes is needed. However, with untraditional delayed ack (with delay window larger than 2) would cause friendliness to the standard. Since sender increases the congestion window (cwnd) by usually counting the number of acks received in the current popular implementations, the with untraditional delayed ack is slower in increasing the cwnd than the standard. If they coexist, the with untraditional delayed ack would not get fair share with the standard. The unfriendliness problem was largely avoided in previous research and stays unsolved. In the paper, we address the problem with a simple and intelligent end-to-end receiver-side detection mechanism. If the receiver detects other competing traffic, it switches to the standard, and vice versa. With this adaptive switching scheme, the friendliness of the standard and our proposed adaptive delayed ack scheme is significantly improved. Another challenge for with adaptive delayed ack comes from mobile ad hoc networks. Although delayed acks may potentially improve throughput regardless of underlying routing protocols. Different routing mechanisms, however, have great impact on performance [4], and to what degree delayed acks can benefit performance. The problem of more packet losses due to

3 11 J. Chen et al. / Ad Hoc Networks 6 (8) mobility could have more impact on with untraditional delayed ack since fewer acks are generated. In this paper, we investigate the (Delayed Cumulative Ack) which addresses the ack loss issue in mobile networks. Meanwhile, the performance of various schemes are presented over popular ad hoc routing protocols, namely, Ad- Hoc On-Demand Distance Vector Routing (AODV) [5], Dynamic Source Routing [6] and Greedy Perimeter Stateless Routing (GPSR) [7]. The reason we include GPSR in our study is that Geo-routing is a recently developed routing scheme promising to be scalable, and robust to node mobility. While our work shares the common concept with previous research in that the receiver delays ack up to a certain number of data packets, we undergo a more complete and optimized study to understand the delayed ack impact and propose more effective solutions suitable for various scenarios. To the best of our knowledge, our proposed schemes are the first existing mechanisms suitable for both static/mobile wireless and hybrid networks, and taking the ack loss and friendliness into account as well. The remainder of this paper is organized as follows. Background work is provided in Section 2. A thorough study of delayed ack impact on performance is presented in Section 3 under various network topologies and traffic patterns. We present our adaptive delayed ack () scheme in Section 4. The issue of the loss of ack packets is discussed. Section 5 evaluates on static and mobile ad hoc networks, and hybrid wired/wireless networks as well. Several enhancements are presented to improve s efficiency and friendliness. Section 6 discusses a few issues to further improve with untraditional delayed ack. We conclude the paper in Section 7. The impact of different routing protocols on is also considered in proper sections. 2. Related work The standard assumes that a packet loss is invariably due to buffer overflow and reduces the cwnd by half when packet loss happens. However, in ad hoc network, packet loss caused by buffer overflow is rare. Instead, it is more likely due to medium contention as shown in [8]. The fundamental problem resides in the limitations of IEEE MAC. Since the interference range is usually longer than the transmission range, request-to-send (RTS) and clear-to-send (CTS) in MAC cannot ensure collision free transfer of packets. This causes hidden/exposed terminals leading to packet loss in ad hoc multihop wireless paths [8]. Many research efforts have been made to adapt to the unique characteristic of wireless ad hoc network, e.g. Transport layer Fixed RTO in [4], delayed ack in [2,3], network layer support [9 11] and even lower-layer assistance, for example, MAC support in [8]. Although a ack packet is small, typically 4 bytes, the transmission of ack packets may require the same overhead as that of data packets in MAC depending on the RTS threshold. If interference from acks could be reduced, data packets would suffer less collisions resulting in higher throughput. Several approaches to delay acks have been proposed [2,3,12]. - ADA (Adaptive Delayed Ack) [12] proposed to decrease the number of acks to improve performance. They claimed that maximum throughput is achieved when one ack acknowledges the full cwnd. However, the scheme did not address several important issues, such as packet loss and out-of-order packet reception. In fact, as we will show in this paper, does not always benefit from delaying ack as much as possible. Altman and Jimenez [2] presented a basic delayed ack scheme which was further improved in [3] where (Dynamic Adaptive Acknowledgement) was proposed. In, the receiver adjusts the delay window according to the channel condition (packet loss event). A receiver delays acking until it receives a certain number of data packets, ranging from 2 to 4 packets. When there is no packet loss, the receiver waits for more data packets (up to 4) before generating an ack, but reduces the number to 2 in case of out-oforder packet arrival. However, since a sender automatically cuts the cwnd when packet loss occurs, i.e. automatically adapting to the channel state at the sender side, the receiver side adaptation provides little extra improvement. We will show that in our adaptive delayed ack scheme, the receiver does not respond dynamically to packet loss, yet achieves better performance. In [3], the ack time is set to be one average packet inter-arrival time. That is, an ack is generated when no data packet arrives after one average packet inter-arrival time since last unacknowledged data packet. In wireless networks, the inter-arrival time

4 J. Chen et al. / Ad Hoc Networks 6 (8) is highly variable due to random MAC contention and back-off, so it is very difficult to get accurate enough statistics in a complex large system. Another impact of this timer implementation is that the receiver operates insensitively to the number of acks (2 4) to be delayed because any unexpected extra delayed data packet will trigger an ack. An important aspect of with untraditional delayed ack is the delay window size selection. In [3], the receiver may delay up to four ack packets and this number is limited by the sender cwnd, which is fixed at four packets. The similar delay window size of 3 4 packets is also picked in [2] heuristically. There are issues in this scheme that desires further clarification and improvement. First, although a small cwnd limit helps operation in wireless networks, it is not suitable for hybrid wired and wireless network where a high bandwidth delay product exists. Second, if the cwnd is not limited by four packets, the choice of a delay window size of 4 may no longer be suitable. In this paper, we address the above issues and provide more effective and general solutions that apply to various environments. The delayed ack has impact on the sender. Since the receiver does not ack as frequently as in the standard, the congestion window increasing rate for delayed ack is slowed down. Such slower probing rate improves performance in ad hoc networks, as reported in [13]. The conservative window increase, however, hinders efficient transmission in wired network where delay bandwidth product may be large. In this paper, we solve this problem and allow to cope with ad hoc and hybrid wired/wireless networks via adaptive schemes. We also study the impact of cwnd limit at the sender. A small sender cwnd can decrease interference and maximize pipeline effect. [8] revealed that there exists an optimal cwnd size that maximizes spatial channel reuse. Further increasing the window size does not lead to better performance. On the contrary, it results in increased link layer contention and degraded throughput. In [14], the optimal congestion window limit is determined based on the hop count for maximum pipeline effect and the interference/transmission range ratio. In the paper, we will show that in our scheme, the sender s cwnd is not limited, and yet, performs better than the case of the limited cwnd used in [3]. resides at layer, and it has potential to be combined with other enhancements for further improvement. As it is a primarily receiver-side modification, it could be combined with other sender-side schemes or low layer approaches. For example, at sender-side, [15] studied the performance of on three different routing protocols and proposed a modification called fixed RTO, which essentially freezes the RTO value whenever there is a route loss. Recent work [9 11] are approaches at the network layer. Chandran et al. [9] discussed a mechanism called -Feedback, which uses route failure and re-establishment notifications to provide feedback to, and thus reduce the number of packet retransmissions and back-offs during route calculation, to improve throughput. Holland and Vaidya [1] evaluated an explicit link failure notification technique (ELFN) in the context of improving performance over multi-hop ad-hoc networks. They studied the effect of link failures due to mobility on throughput and showed through simulations that ELFN improves the performance of. Xu et al. [11] proposed a Neighborhood RED (NRED) scheme, which extended the RED concept to the distributed neighborhood queue, to address significant unfairness in ad hoc wireless networks. These mechanisms are promising schemes to be integrated with our mechanism for more robustness against packet loss and fairness. 3. throughput vs. delay window In this section, we investigate the impact of the receiver ack delay window size on performance. The conclusion we draw is that, when the path length is short, achieves better performance with a delay window as large as the entire sender cwnd. When the path length increases, however, a large delay window triggers bursty transmissions, which results in mutual data packet contention and higher losses. In fact, for long paths, there exists an optimal delay window size that achieves maximum throughput. We verify the delay ack impact using various network topologies, including cross and grid topologies with various flow patterns. Real testbed measurement results are also presented to reinforce our conclusions. Firstly, we study a basic scheme with fixed delay window limit to illustrate the impact of the delay window size on performance, named as - FDCA. We use static routing to investigate delayed ack performance without any routing interference. Routing impact will be shown in Section 5.

5 112 J. Chen et al. / Ad Hoc Networks 6 (8) with fixed delay window limit (-FDCA) -FDCA maintains a variable delay window w representing the number of data packet arrivals before acking at the receiver. This delay window w is bounded by a delay window limit, and also limited by the cwnd size to prevent the delay window from exceeding the cwnd which results in delaying the ack, but for possibly non-existent packets. If data packets arrive in order, the receiver generates one cumulative ack for every w data packets. If an out of order packet is received, or a packet that fills a gap in the sequence space of packets in the receiver buffer (that is, recovery from earlier packet loss), the receiver acks immediately to inform the sender of the packet loss/recovery in a timely manner. The receiver also keeps a fall-back delay ack timer which estimates the expected time to receive w data packets. At the sender, when it receives a cumulative ack acknowledging w data packets, it updates the cwnd and sends out a burst of at least w packets, provided such packets are ready for transmission in the sender buffer. To get the delay timer period, the receiver monitors the data packet inter-arrival interval and computes a smoothed interval through a low-pass filter. The average inter-arrival interval is used to set the ack delay timer based on the current delay window size. In -FDCA, an accurate delay timer is not needed and the timer is solely for fall back purpose. In fact, our inter-arrival time computation is rather simple and loose: the receiver samples the inter-arrival time of any consecutively arrived data packets, not necessarily in-order packets. Therefore, the inter-arrival sample is an inflated value compared with inter-arrival between in-order packets. A -FDCA receiver smoothes such inflated inter-arrival samples through a low-pass filter I iþ1 avg ¼ bi i avg þð1 bþi s; ð1þ where I avg is the average of inflated packet inter-arrival time (I s ). I avg is used to set delay ack timer (T w ) which defines the total time for receiving a whole delay window of in-order data packets T w ¼ awi avg ; ð2þ where w is the delay window size, a is a parameter to tolerate high dynamic packet delay. Obviously, the delay ack timer is rather inflated and quite robust to high inter-arrival variation. In the paper, we choose b as.8 and a as 1.5. The receiver also generates an ack within 5 ms of the arrival of an unacknowledged data packet in accordance with RFC2581 [1]. One drawback to this approach is that the - FDCA receiver needs to be informed of sender s cwnd size to prevent delay window larger than cwnd leading to unnecessary delaying at the receiver. To address this problem, we propose two possible methods. One is to imbed the cwnd size in an option field of the packet header. The other is to reuse some header field. For example, the sender could reuse the advertised window field in data packet header for advertising its cwnd size to the receiver. This method is not unrealistic due to the fact that current connection is mostly used for one direction, i.e. there is only data packets from the sender to the receiver and no backward data, so the advertised window field from the sender is usually wasted. Even if two-way data is included, it is still possible to reuse the advertised window field, for example every other packet with one bit in the reserved field to indicate whether cwnd or advertised window size is used. The lowed frequency of cwnd notification is not expected to affect -FDCA much since -FDCA receiver does not need to know the exact cwnd so often, as we will see later in the paper Chain topology In order to understand -FDCA, its performance is shown over a chain topology in Ns2 [16]. The chain topology is general used to study performance since typically runs over a single path. Fig. 1 shows a chain example. The connection is sourced at the first node (node 1) and packets travel over the chain to the end node (node 6). The interference range is 55 m and transmission range is 25 m. We place the nodes at m intervals, so that nodes are 4 hops away can transmit simultaneously without interference. Notice that a node that is 3 hops away is a hidden node. IEEE is the underlying MAC. Data rate on the wireless channel is 2 Mbps and one simulation run lasts s. Each data point represents an averaged result of 5 simulation runs with different random seeds. The packet size is 1 bytes and the - Newreno is adopted as it is the most popular protocol nowadays. The delay window limit applied at the receiver ranges from 1 to the current cwnd size.

6 J. Chen et al. / Ad Hoc Networks 6 (8) hop 2hop 3hop Fig. 1. Chain topology. The solid-line circle denotes a node s valid transmission range. The dotted-line circle denotes a node s interference range. Node 4 s transmission will interfere with node 1 s transmissions to node CW Delay Window Limit (packets) (a) Path Length h Single flow performance over wireless multihop inevitably depends on the path length (hop count). The longer the path is, the lower the throughput would be. We present -FDCA throughput as a function of delay window limits on a path with different lengths in Fig. 2. Fig. 2a shows that when a sender and a receiver are within 3 hops, -FDCA gets steady performance gain by increasing the delay window size up to the entire cwnd size. For the one hop case, the fastest throughput increase can be seen when the delay window is small, say less than 5. The increasing trend becomes slower for delay windows larger than 5 and the throughput approaches the limit when the delay window reaches the cwnd size (denoted by CW). The same trend is also observed when the path length is 2 and 3. The reason for this performance gain is that MAC provides nearly collision free packet transmissions for short paths. Since the interference range is larger than 2 times the transmission range, when the sender and receiver are within 2 hops, every node along the path can sense all other nodes transmissions. In this case, no hidden nodes exist, thus packet loss is minimal. When the hop length is 3, the receiver is the only node hidden from the sender. If the receiver acks only after receiving all packets in a cwnd, no data packet can be interfered. The maximal performance gain is achieved when the receiver acks after receiving all packets in a cwnd, and attains up to 25% throughput gain relative to the with delay window at 1. Since the interference range is larger than the transmission range, RTS/CTS cannot completely hop 5hop 6hop 7hop 8hop 9hop CW Delay Window Limit (packets) (b) Path Length 1 > h >3 1hop 12hop 14hop 16hop 18hop 2hop CW Delay Window Limit (packets) (c) Path Length 2 h 1 Fig. 2. -FDCA throughput vs. delay window on chain topology. (a) Path length h 6 3. (b) Path length 1 > h > 3. (c) Path length 2 P h P 1. solve the hidden/exposed terminal problem. As the path becomes longer than 3 hops, packet collision

7 114 J. Chen et al. / Ad Hoc Networks 6 (8) is unavoidable. For example, node 1 and node 4 are interfering nodes in Fig. 1, so simultaneous packet transmissions on them will be interfered. The hidden terminal results in packet loss and this problem becomes more severe for longer paths because a packet has more chances to be interfered with. Another problem is that when the sender receives a cumulative ack, it immediately sends a burst of packets. The larger the delay window, the larger the burst and the more packet loss due to increased interfering. (Note that the packet burst size is directly related with the receiver delay window since it is at least equal to the size of the delay window.) When the packet loss becomes high, the - FDCA throughput gain from delaying ack is lost due to the increased packet losses. Fig. 2(b) shows this tradeoff of -FDCA performance gain with delay window size for paths larger than 3 hops. When the hop count is 4 or 5, we do observe unsuccessful packet transmissions caused by interference, however, since a sender is able to recover packet loss rather rapidly due to the small round trip time (RTT), -FDCA maintains performance gain by delaying ack for more data packets. After peak performance at certain delay window, the -FDCA performance gets worse due to more packet losses. Similar trend also exists for paths longer than 5 hops, where -FDCA achieves throughput gain only when the delay window size is small; for large delay window size, delaying ack cannot maintain throughput gain because of excessive data packet losses. Further, now that RTT is larger, -FDCA spends more time detecting packet loss and recovering lost packets by entering fast retransmit/recovery in which only one packet is recovered per RTT. Therefore, for long paths, large delay window is not preferred. We also show -FDCA performance over a very long path h P 1 in Fig. 2(c) which has similar results with Fig. 2(b). -FDCA only gets performance gain for small delay window size. For large delay window size, -FDCA gets poor throughput Real testbed verification We investigate the effectiveness of our - FDCA in an actual ad-hoc network testbed. Our testbed consists of six laptops equipped with Orinoco 82.11b PCMICA cards with channel rate of 2Mbps. The laptops run Mandrake Linux distribution 7 with kernel version Linux PCMCIA package version 3.2. and Orinoco wavelan2-cs driver are used for 82.11b devices and the devices are set to ad-hoc mode. The topology of the testbed is a chain and the route is statically configured. There is one source laptop and one selected destination laptop among other 5 laptops depending on the number of hops in our experiments. We use Iperf to generate FTP traffic with packet size of 1 bytes. From the experiment results shown in Fig. 3, we confirm that the -FDCA throughput vs. receiver delay window follows the same trends as in the simulation results shown in Fig. 1. Fig. 4 compares -FDCA throughput for 1 hop and 5 hops paths from simulation and actual testbed, when delay window limit is equal to 1 and cwnd respectively. We note that the average difference between the testbed throughput and the simulation results is below 1%. Such difference is hop 2hop 3hop CW Delay Window Size (packets) hop 5hop (a) 1 to 3 hop CW Delay Window Size (packets) (b) 4 and 5 hop Fig. 3. Real testbed measurement. (a) 1 to 3 hop. (b) 4 and 5 hop.

8 J. Chen et al. / Ad Hoc Networks 6 (8) More complex topologies and flow patterns Total d 1 d CW d 1 d CW Simulation (a) 1 hop Experiment We expand our study to scenarios of more complex topologies and flow patterns, including cross and grid topologies. For all cases, we observe the similar results to what is described above, i.e., when the path is short, -FDCA achieves optimal throughput by delaying acks as much as possible, up to the entire sender s cwnd. On the other hand, when the path becomes longer, a large delay window hinders performance gain and deteriorates throughput gradually. We show that there exists a certain delay window size, at which - FDCA achieves optimal throughput performance. The following provides a summary of results on multiple flows and various topologies. The performance over random topologies will be shown in Section 5.2. Total Simulation (b) 5 hop Experiment Fig. 4. Simulation vs. experiment. (a) 1 hop. (b) 5 hop. mainly caused by parameter setting in the testbed measurements. In our testbed experiments, RTS/ CTS control packets are adopted for unicast data packets to overcome the hidden terminal problem only if the packet size is above the minimum threshold 256 bytes. Since the size of ack packet (4 bytes) is below the minimum RTS/CTS threshold in Linux settings so that no RTS/CTS handshake is performed when transmitting acks in testbed experiments. Compared with simulation results where RTS/CTS is always performed regardless of packet size, such zero MAC overhead in testbed measurements has impact on the performance: the throughput gain derived from -FDCA in the testbed measurements is lower than the corresponding results in the simulations because of the lack of RTS/CTS reduction for acks in simulations. Other than this slight difference, the results are similar in simulation and in testbed measurements Multiple flows Fig. 5 exhibits result for 5 concurrent - FDCA flows on a chain topology with hop count varying from 3 to 7, and has similar results to Fig Cross and grid topology Fig. 6 shows examples of cross and grid topologies with a 4 hop path for each -FDCA flow, and the results with different delay window size are shown in Fig. 7. Additionally the results from varying path lengths are shown. Similar trend is observed again. We also run extensive simulations on cross and grid topologies with multiple overlapped flows instead of one flow. The results, not included here, also confirm the same trends discussed above. Aggregate hop 5hop 7hop CW Delay Window Limit (packets) Fig. 5. -FDCA throughput vs. delay window on chain topology (5 flows).

9 116 J. Chen et al. / Ad Hoc Networks 6 (8) flow 1 7 Flow Flow 2 flow (a) Cross To pology Flow 3 (b) Gri d Topology Fig. 6. More complex topologies. Left: cross topology with 4 hop path on each direction. m distance between two adjacent nodes. Right: 5 5 grid topology, m distance between horizontal and vertical adjacent nodes. (a) Cross topology. (b) Grid topology hop 6hop 8hop x5 5x7 5x9 Aggregate Aggregate CW Delay Window Limit (packets) (a) Cross Topology CW Delay Window Limit (packets) (b) Grid Topology Fig. 7. -FDCA throughput vs. delay window on complex topologies. (a) Cross topology. (b) Grid topology. 4. with adaptive delayed cumulative ack () In this section we propose with adaptive delayed window according to the underlying path information to optimize performance. The issue of ack loss is also discussed and a simple receiver retransmission scheme is proposed with adaptive delayed ack Up to now we have presented the tradeoff between -FDCA throughput and delay window size on different path lengths. Inspired by this observation, we dynamically determine the delay window size based on the hop count of a connection and call it as. For a short path (h 6 3), could delay ack for a whole cwnd to get best performance; otherwise a small delay window for long paths. Since still gets slight benefit by delaying more for moderate long paths (3 < h 6 9), therefore the delay window could be slightly larger for such paths. receiver can get the path length information from the TTL field in packet header or from the routing layer at the receiver node Table 1 lists the delay window size applied in according to the path length.

10 J. Chen et al. / Ad Hoc Networks 6 (8) Table 1 Delay window at receiver Path length (h) Delay window limit h 6 3 Congestion window 3 < h h P Ack loss The standard generates more acks than, due to the cumulative property of the ack, one ack successfully received at the sender can makes up for all the preceding lost acks. Thus, the ack loss in standard is not serious. However, the number of acks in is much less than standard, and thus an ack loss has more negative impact on performance. In order to be robust to ack loss, a retransmission timer at the receiver is adopted to retransmit a cumulative ack if necessary. A challenge here is to estimate RTT at the receiver which is needed for setting the ack retransmission timeouts. If includes two-way data, the receiver can get the accurate RTT measurement since the sender acks the data immediately. If only has one-way data, RTT measurement at the receiver may not be straightforward. If the timestamp option is used, the receiver could use it to estimate RTT, though such an estimate may be inflated if the sender does not send data packets immediately after receiving an ack [17]. However, in, the ack retransmission timer is only a coarse timer for predicting when to retransmit acks, an accurate RTT measurement is not required and thus an inflated RTT can be tolerated. In the paper, the timestamp option is used to estimate the RTT for a flow which has one-way data. The receiver computes the ack retransmission time based on this RTT measurement. We use the following formula to calculate the retransmission timer period: SRTT r kþ1 ¼ 7 8 SRTTr k þ 1 8 RTTr kþ1 ; RTT var kþ1 ¼ 3 4 RTTvar k þ 1 4 jrttr kþ1 SRTTr kþ1 j; RTO r kþ1 ¼ SRTTr kþ1 þ 4 RTTvar kþ1 ; where RTT r is the RTT estimation at the receiver, SRTT r is the smoothed RTT. RTT var and RTO r are RTT variance and retransmission timer period at the receiver. The minimum value of the retransmission timer period is set to 1 s. This ack retransmission timer is started after sending an ack, and canceled after receiving a data packet. Until now all previous results were obtained without ack retransmission. This is due to two facts: first, the results with and without ack retransmission are almost the same for static networks. Mostly because the ack loss is rare, and the ack retransmission timer period is large (at least 1 s). Second, without the ack retransmission timer, it is easy to observe the sole impact of the delay window on performance. In the following simulations, the ack retransmission timer is enabled unless explicitly mentioned. 5. Performance evaluation This section presents the performance over wireless and hybrid networks. First, we show performance on static multihop wireless networks and compare it with in [3]. Second, we demonstrate performance on mobile ad hoc network, and show performance over several ad hoc routings. Third, we extend the hop count based approach to an endto-end delay based scheme to further improve in hybrid wired/wireless networks. Last, we propose a simple approach to address the friendliness problem of to -Newreno with standard delayed option Static ad hoc network In this section, we show performance in static ad hoc networks. Meanwhile, we compare with and -Newreno (with and without standard delayed ack option [1]). More specially, we use the name original for -Newreno without standard delayed option, and for -Newreno with standard delayed option. AODV routing is used unless otherwise stated. [3] is an interesting extension of [2] and has shown good performance in static ad hoc networks. It is the most recent work and the major differences between and are listed in Table 2. For fair comparison, most simulations scenarios presented in this subsection were shown in [3]. Each result is the average of five simulation runs. In Fig. 8, we compare to in the chain topology with different hop count and

11 118 J. Chen et al. / Ad Hoc Networks 6 (8) Table 2 Design difference Sender Duplicate acks threshold to trigger retransmission is 2; CW upper limit is 4; Retransmission (RTO) timer is increased fivefold Puts CW into advertised window field in packet header Receiver Delay window is adaptive from 2 and 4 based on loss event Delay window is adaptive on the path length number of concurrent flows. Although is designed without cwnd limit, we include with cwnd limit at 4 in this experiment, named - DCA-CWL, to show a fair comparison with - DAA. Note that in, the cwnd limit is set to 4 by default. Fig. 8a shows the performance of over 3 hop chain. The aggregate throughput of decreases when the number of flows increases. When there are few flows, since each generates as few acks as possible, the advantage of is obvious. When more flows coexist, the cwnd of each flow becomes smaller since they compete for the bandwidth, thus the benefit of delayed ack reduces because more acks are generated. If the cwnd is below the cwnd limit in -CWL, they have similar performance. The performance of is better than original, but worse than and. and -CWL provides better performance than other variants. Overall, achieves best performance, up to 15% improvement over, and 25% over respectively. Fig. 8b c show the performance of on a 5 and 7 hop path. The same trend persists and outperforms all other algorithms in most situations. Note that the cwnd limit on does not bring considerable benefit. For a short path in Fig. 8a, the cwnd limit could considerably restrict the performance as the number of flows is small. For a long path in Fig. 8b c, the cwnd limit only provides slightly more performance gain for 2 flows, and this advantage disappears for more flows. From Fig. 8, the performance of without cwnd limit is better than that of -CWL in most situations. It indicates that can achieve the distinguished performance without setting cwnd limit. We also give results for the grid topology. This grid topology is shown in Fig. 9a and 6 flows running on the grid. Fig. 9b compares the performance of original,,, - DCA-CWL and, and we observe the similar results as before., - CWL and achieve best performance and their difference is small. They provide better performance, about 1 25%, than the standard (with and without delayed ack option). With static topology and steady traffic pairs, the variance of results is very small (less than 5%) so we omit its error bar results. Also note that - CWL does not bring significant benefit over - DCA. Since the ack flow is decreased, the optimal cwnd size could actually increase (for details, please refer to [14]). Although it may be interesting to investigate the optimal cwnd for, we do not pursue it because our ultimate goal to design without cwnd limit. Moreover, this property makes distinct since selecting an optimal cwnd is non-trivial, which needs to know a lot of underlying information to accomplish. For example, it requires the knowledge of transmission/interference ratio [14] at the physical layer which is usually hard to obtain. It could also potentially limit the application of in a variety of wireless networks, such as that with cwnd limit in wired and wireless scenario does not perform well (will be shown in Section 5.4). Taking this into account, we do not choose cwnd limit in - DCA and thus in the following experiments, only (no cwnd limit) is adopted Mobile ad hoc network We have demonstrated that applying delayed ack scheme improves performance in a static network, now we show that it can improve performance in a mobile network as well. Though our delayed ack scheme is not specifically designed for the in mobile network, we show that our scheme works well in mobile scenarios. In Fig. 1, we show the performance of the original,, and - DCA running over three routing schemes: GPSR, AODV and DSR. The experiment consists of 4 nodes randomly moving within a 1 m 1 m area. Each node moves with the same speed without pause. The Random Waypoint mobility model [18]

12 J. Chen et al. / Ad Hoc Networks 6 (8) CWL Flow 1 Flow 4 Flow 5 Flow 6 Aggregate 5 45 Flow 2 Aggregate Number of Flows (a) 3 Hop Path -CWL Flow 3 - DCA - DCA- CWL - DAA - STD- DA (a) Grid Topology Aggregate Number of Flows (b) 5 Hop Path -CWL Number of Flows (c) 7 Hop Path Fig. 8. Aggregate throughput for variants over chain topology. (a) 3 hop path. (b) 5 hop path. (c) 7 hop path. is used. All results displayed are calculated as the average of 4 runs with the same traffic pattern, Aggregate (b) Aggregate Throughput for variants over Grid Topology Fig. 9. Performance comparison on grid topology. (a) Grid topology. (b) Aggregate throughput for variants over grid topology. but with different randomly generated mobility traces. We run the set of experiments with 4 different mobility traces and plot the 95% confidence interval as error bars on the figures. We illustrate results on speed 1 m/s and 2 m/s in Fig. 1 with one flow, for other speeds and multiple flows, the results are similar. From Fig. 1 we observe that the delayed cumulative ack contributes to the performance gain compared with other variants. produces at least 1% performance gain over other variants regardless of routing protocols and mobility speed, especially more gain for over DSR. is unsurprisingly consistently better than the original without delayed ack. However,

13 111 J. Chen et al. / Ad Hoc Networks 6 (8) ACK-LOSS ACK-LOSS GPSR AODV DSR GPSR AODV DSR (a) 1 m/s (b) 2 m/s Fig. 1. performance over mobile network. (a) 1 m/s. (b) 2 m/s. does not show consistent improvement across different routing protocols. It performs similarly as - STD-DA over GPSR, but does not show better performance than (with or without standard delayed option) over DSR. In, the receiver would probe the sender by retransmitting acks in the event of high packet loss (see Section 4.2). It prevents the sender from entering long timeout due to the severe packet loss event. In Fig. 1, we also show the result of without ack retransmission (- DCA-ACK-LOSS) to confirm the effectiveness of ack retransmission. When packet loss is severe (e.g. over DSR), ack retransmission considerably boosts the s performance. On the other hand, the ack retransmission does not help much when packet loss is not severe (e.g. over GPSR). Therefore, only incurs more ack retransmission overhead when needed, and works well with packet losses in mobile networks. Comparing with static ad hoc networks in Section 5.1, mobility poses more challenges for. In a mobile network, the loss of packets (or acks) may be caused by temporary route loss as well as network congestion. Higher mobility causes more route loss and thus more packet losses. Since routes are likely to be lost frequently in high node mobility environments, and different routing algorithms have different capability to repair broken routes quickly, thus different performance is observed over different routing schemes. Moreover, The frequent route breakages could put into long idle state leading to poor performance, especially for with untraditional delayed ack where fewer acks are generated and more timeouts are likely to happen. Fig. 11 illustrates the idle time of during the whole simulation time (4 s) for different variants and routing schemes. As expected, (considering ack loss) has least idle time over each routing scheme, which helps maintain better throughput in a mobile network. Note that -ACK-LOSS has similar idle time with other variants, which confirms that retransmission scheme is effective in improving performance in mobile networks Lossy network In wireless networks, the wireless link in fact could become unreliable and error-prone. After studying the performance of under packet losses in mobile networks, we investigate its performance under lossy wireless link. BER (bit error rate) is used as the loss metrics at the wireless link and the error is assumed to be uniformly distributed and independent on each link. It is worth pointing out that when a packet is received with bit error, the IEEE MAC layer helps in recovering packet loss by retransmissions until a maximum retransmission limit is reached [19] or retransmission is successful. Therefore, when BER is low, the MAC retransmission scheme can easily recover the packet and would not cause much problem for. While the BER is high, the heavy packet loss would knock down performance. Fig. 12 illustrates a scenario where a single flow goes through 5 hop chain path under varying BER from very low BER to high BER. Each bar

14 J. Chen et al. / Ad Hoc Networks 6 (8) ACK-LOSS 35 -ACK-LOSS Idle Time (second) 15 1 Idle Time (second) GPSR AODV DSR (a) 1 m/s GPSR AODV DSR (b) 2 m/s Fig. 11. idle time. (a) 1 m/s. (b) 2 m/s. Throughput(Kbps) e-6 6e-6 2e-5 6e-5 Bit Error Rate Fig. 12. performance under random loss. is an averaged result of 2 random runs. Overall, the performance of is no worse than other variants, and quite robust to wireless loss. - DAA also works very well unless the loss is too heavy. With the increasing loss rate, all variants gracefully reduce their performance and the variance of results increase Hybrid (wired and wireless) network In this section we study in the wired and wireless ad hoc environment. Since wired network has abundant bandwidth, it demands a much larger cwnd than the pure ad hoc network for effective performance. In current implementation, sender increases the cwnd only based on the number of acks received while not taking into account the acknowledged data covered in the acks. Using delayed ack could potentially cause slow cwnd increase and thus hurt performance. Therefore, the pure adaptive delay window based on the hop count appears not adequate for this scenario. For example, in WLAN where 1 hop wireless client can delay the cumulative ack for the whole cwnd, the cwnd at the sender could increase slowly, particularly if a large propagation delay is encountered in the wired part. Although slow increase for cwnd is helpful for improving performance in ad hoc networks as shown in our previous results, it is not desirable for the hybrid network. In RFC3465 [2] Allman proposed to increase the cwnd based on the number of bytes acknowledged by the arriving acks rather than based on the number of acknowledgments that arrive at the sender in RFC2581 [1]. However, this approach potentially causes large bursts if the delay window is large. An appropriate delay window limit needs to be investigated in this scenario. Another challenge of the hybrid network is that the hop count information from the packet header is no longer accurate for the wireless hops since it includes the wired path. It is possible for a wireless node to get hop count from the routing table to the base station (or gateway) to the wired network, however, this information is not guaranteed to be available all the time. To address the problems in hybrid networks, we tailor and propose a novel receiver-side probe mechanism based on CapProbe in [21], and extend the previous hop-count based approach via the receiver-side probe in hybrid networks.

15 1112 J. Chen et al. / Ad Hoc Networks 6 (8) The receiver-side probe works as follows: First the receiver passively estimate the wireless bandwidth and minimum RTT. Second it computes the equivalent hop count based on the minimum RTT and wireless bandwidth, then the previous hop count approach can be applied afterwards. Now we address the problem to estimate the endto-end hybrid path bandwidth at the receiver. CapProbe [21] is a recently proposed bottleneck capacity and path rate estimation technique shown to be both fast and accurate. It combines the use of time interval measurements and end-to-end delay measurements to filter out bad packet pair probing samples. It is proved to be suitable for wired and last-hop wireless networks. Inspired by CapProbe, we proposed a new approach at the receiver without triggering any probing traffic. The idea is that after the receiver sends a cumulative ack, the sender would send a burst of packets. The burst of packets can just serve as the packet pair probing. Since the receiver knows the sequence number of the first burst packets, it can identify the packet pair and estimate the bandwidth when packet pair arrives. Our approach is based on the CapProbe, but it is different with CapProbe. First our approach is imbed at the receiver without any additional probing traffic. Moreover, our approach is a one-way (instead of round-trip in CapProbe) estimation technique, and it is expected to work for wired, hybrid and purely wireless networks. The other challenge is to estimate the minimum RTT at the receiver, but this have been solved as described in Section 4.2. After the wireless bandwidth and minimum RTT is obtained, the receiver derives the equivalent hop count by considering the data and ack packet sizes. For example, with wireless bandwidth of 2 Mbps, data packet size of 1 bytes and ack size of 4 bytes, the minimum RTT for a 1 wireless hop path is about 6 ms considering various overhead, such as header overhead at (2 bytes), IP (2 bytes) and MAC (34 bytes) and MAC layer fixed overhead shown in Table 3 (for details, please refer to [22]). In Table 4, we listed the delay window size in a wireless network with capacity 2Mbps. In Fig. 13, we show a connection from a wired network to a wireless network with up to 2 wireless hops. The one-way propagation delay on the wired link is 5 ms. The performance of - DCA, and the standard (with and without standard delayed ack option) shown in Fig. 14, represents averaged result of 1 simulation Table 3 IEEE 82.11b MAC overhead (with long preamble), L is packet size, R is capacity Data frame 192 ls +8L/R SIFS 1 ls DIFS 5 ls RTS 192 ls + 16/R CTS 192 ls + 112/R ACK 192 ls + 112/R Table 4 Delay window at receiver (delay-based, wireless bandwidth is at 2 Mbps and data packet size is 1 bytes) Minimum RTT (ms) Delay window limit RTT min < 18 Congestion window 18 6 RTT min RTT min > Internet 1 2 BS Fig. 13. Hybrid wired/wireless topology. 1 hop 2 hop Wireless Hop Count Fig. 14. Wired and wireless performance. running with different random seeds. Similar to Section 5.1, the variance of results is very small so we omit its error bar results. Since limits the cwnd, the performance of is limited. provides best performance in both cases of 1 and 2 hop paths. Therefore, is extensible to hybrid networks instead of working

16 J. Chen et al. / Ad Hoc Networks 6 (8) well only in an ad hoc network. It is an attractive point of as it becomes common for wireless networks to communicate with the outer world. For example, in the real life, one needs to connect a mobile device (e.g. laptop) to the Internet to download files from a remote server Friendliness hop 2 hop 5 hop (a) 1 hop 2 hop 5 hop (b) 1 hop 2 hop 5 hop (c) -SWITCH -SWITCH Fig. 15. Friendliness with -Newreno. (a). (b). (c) -SWITCH. An important issue about is its friendliness with since is currently more popular in all networks. In - DCA, cwnd increases slower than, so is not so aggressive as -STD- DA. The simulation results verify this point in Fig. 15b by running one and one over chain topologies with varying hop count. For each experiment, the average result of 1 simulation runs is shown. We also omit the variance of result since it is very small. For short paths, gets more throughput than. However, when path is long, becomes friendly to -STD- DA since the benefit from delaying acks in is not significant. Note that although is submissive to, its friendliness to is better than - DAA as shown in Fig. 15a. Since not only increases cwnd slower but also limits the cwnd to a small value, its friendliness to is the worst in all cases. One sidenote is that if the original coexists with or, the unfriendliness problem would be more severe since the original is more aggressive than. Fig. 15b shows that is not always friendly to, especially for short paths. However in reality, the friendliness on short paths are more important because the wireless path is usually not long, e.g., WLAN with one hop wireless path. Thus how to make friendly to the possible flows under different scenarios needs to be addressed. In the paper we propose an intelligent switch scheme to allow to match performance if they coexist. The approach is to let receiver detect if it is the only traffic in the network. If there is other traffic, switches to - STD-DA; otherwise if it is the only traffic in the network, it switches back to. The approach is imbed in the receiver of, without incurring any extra overhead. This simple end-toend approach consistent with our design goal for, and it works on receiver without standard delayed option as well since our

17 1114 J. Chen et al. / Ad Hoc Networks 6 (8) Instantaneous Throughput (Mbps) SWITCH DCA is disabled end if else if throughput > b Path Capacity then if no DCA DCA is enabled end if end if 6. Future work and discussion Time (second) Fig. 16. Responsiveness of -SWITCH. adaptive delayed ack scheme is triggered only when necessary. To detect the possible other flows, the receiver monitors the data throughput via receiver-side probe mechanism described above. If the throughput is much less than the path capacity, then there probably has other traffic on the path. We present our switch algorithm in Algorithm 1. a ¼ :4; b ¼ :6 in the simulation. Fig. 15c shows the performance of with switch (-SWITCH). Obviously, the friendliness of to is much improved compared with original - DCA, especially for short paths. The - SWITCH can share the path with quite fairly. To illustrate the transient behavior of - DCA-SWITCH coexisting with, Fig. 16 plots the typical instantaneous throughput in a simulation run. For this simulation set, the first -SWITCH flow runs from the beginning until 25 s and the second flow starts from 1 s to s. Fig. 16 shows that responds quickly to changing network conditions when new flow starts at 1 s as well as when flow stops at s. Using -SWITCH, a flow utilizes the available bandwidth when there are no competing flows and shares the bandwidth fairly when multiple flows compete for the bandwidth. Algorithm 1. Switch Algorithm if throughput < a Path Capacity then if DCA is enabled then In [23], ack congestion control is introduced to to combat bandwidth asymmetry on forward/backward path. Ack packet should be reduced since backward path has limited bandwidth. However, the ack congestion control is implemented at the gateway with a RED queue. It is not an endto-end approach since the receiver needs to get ECN signal from the gateway. In contrast, our scheme is end-to-end, and does not require any changes at intermediate nodes. Delayed ack inevitably triggers burst transportation at the sender. The burstiness increases the packet loss and potentially hurts performance. Since does not use large delay window except for short paths, the burstiness is limited. It is possible to further reduce the burstiness. For example, in [24], cwnd increase is limited by at most 2 packets for each delayed ack, or ack reconstruction in [23] to rate control the sender. How to further decrease the burstiness would be an interesting future work. Another issue is the effect of delayed ack for short flows. As we discussed before, using delayed ack could slow down the cwnd growth, thus delaying ack has potential negative impact on a short flow, as illustrated in [3]. We also found similar results, and we are investigating on improving short flow performance with delayed ack. One simple approach is to disable delaying ack when the cwnd is small, for example, the receiver would not delay acking if the cwnd is less than 4. It is also possible use the cwnd size to detect slow start phase or let sender notify the receiver if it is in slow start phase. If the sender is detected to be in slow start, the receiver temporarily disables the delayed ack scheme to make achieve good throughput at connection startup. After leaves the slow start, receiver begins delaying ack to better utilize the shared wireless medium. These enhancements are expected to be beneficial for short flows because most of short flows finish their

18 J. Chen et al. / Ad Hoc Networks 6 (8) transfer while they are still in the slow start phase [25]. 7. Conclusion, a dominant reliable transport protocol in wired networks, is a highly competitive candidate for providing reliable data transport in wireless and wired/wireless networks. This paper systematically examines the relationship of performance to delayed ack via analysis, and simulation and testbed experiments. We find that does not always get throughput gain by delaying unlimited acks. The maximal throughput is achieved at a certain delay window that balances decreasing ack flow and burst loss. Thus, we introduced two novel adaptive delayed ack mechanisms compatible with called, based on the path hop length and/or end-to-end delay, to maximize the performance for MANET and hybrid networks. Our proposed schemes are shown to have superior performance in static networks (wireless or hybrid wired/wireless networks). We also demonstrated better performance with our proposed mechanisms over different routing protocols in mobile ad hoc networks. It is well known that with untraditional delayed ack is less aggressive than -Newreno since the cwnd with untraditional delayed ack grows much slower. To combat such unfriendliness, we propose a receiver-side probe and detection to maximally increase the friendliness (when other flows exist) and efficiency (when no other flows exist). The detection is imbed in the and does not need to modify underlying layers. friendliness to - Newreno is significantly enhanced with this simple approach. References [1] M. Altman, Tcp congestion control, RFC 2581 (1999). [2] E. Altman, T. Jimenez, Novel delayed ack techniques for improving performance in multihop wireless networks, Personal Wireless Communications, September 3. [3] R. de Oliveira, T. Braun, A dynamic adaptive acknowledgment strategy for over multihop wireless networks, in: Infocom, 5. [4] T.D. Dyer, R.V. Boppana, A comparison of performance over three routing protocols for mobile ad hoc networks, in: Mobihoc, 1. [5] C.E. Perkins, E.M. Royer, Ad-hoc on-demand distance vector routing, in: Proceedings of IEEE WMCSA 99, February [6] D.B. Johnson, D.A. Maltz, in: in: T. Imielinski, H. Korth (Eds.), Dynamic Source Routing in Ad hoc Wireless Networks, Mobile Computing (chapter 5). [7] B. Karp, H.T. Kung, GPSR: greedy perimeter stateless routing for wireless networks, in: Proceedings of ACM MobiCom, August. [8] Z. Fu, P. Zerfos, H. Luo, S. Lu, L. Zhang, M. Gerla, The impact of multihop wireless channel on throughput and loss, in: IEEE INFOCOM 3, March 3. [9] K. Chandran, S. Raghunathan, S. Venkatesan, R. Prakash, A feedback-based scheme for improving performance in ad hoc wireless networks, IEEE Personal Communications Magazine 8 (1) (1). [1] G. Holland, N.H. Vaidya, Analysis of performance over mobile ad hoc networks, in: Proceedings of ACM MobiCom 99, August [11] K. Xu, M. Gerla, L. Qi, Y. Shu, Enhancing fairness in ad hoc wireless networks using neighborhood red, in: Mobicom, 3. [12] A.K. Singh, K. Kankipati, Tcp-ADA: Tcp with adaptive delayed acknowledgement for mobile ad hoc networks, in: WCNC, 4. [13] K. Nahm, A. Helmy, C.J. Kuo, Tcp over multihop networks: issues and performance enhancement, in: Mobihoc, 5. [14] K. Chen, Y. Xue, K. Nahrstedt, On setting s congestion window limit in mobile ad hoc networks, in: ICC, 3. [15] T.D. Dyer, R.V. Boppana, A comparison of performance over three routing protocols for mobile ad hoc networks, in: Proceedings of ACM MobiHoc 1, October 1. [16] The network simulator - ns2, Available from: [17] V. Jacobson, Tcp extensions for high performance, RFC 1323, [18] C. Bettstetter, H. Hartenstein, X. Prez-Costa, Stochastic properties of the random waypoint mobility model, ACM/ Kluwer Wireless Networks: Special Issue on Modeling and Analysis of Mobile Networks, September 4. [19] IEEE-82.11, Wireless LAN media access control (MAC) and physical layer (PHY) specifications, Available from: [2] M. Allman, Tcp congestion control with appropriate byte counting (abc), RFC 3465, 3. [21] L.L.M.G. Rohit Kapoor, Ling-Jyh Chen, M.Y. Sanadidi, Capprobe: a simple and accurate capacity estimation technique, in: ACM SIGCOMM, 4. [22] M. Kappes, An experimental performance analysis of mac multicast in 82.11b networks for voip traffic, in: SPECTS, 4. [23] H. Balakrishnan, V.N. Padmanabhan, R.H. Katz, The effects of asymmetry on performance, in: Mobicom, [24] M. Allman, On the generation and use of acknowledgments, ACM Computer Communication Review, [25] H. Balakrishnan, S. Seshan, M. Stemm, R. Katz, Tcp behavior of a busy internet server: Analysis and improvements, in: Infocom, 1998.

19 1116 J. Chen et al. / Ad Hoc Networks 6 (8) Jiwei Chen received the B.S. degree in Applied Mathematics and M.S. degree in Electrical Engineering from Nanjing University of Science and Technology, China. Currently, he is a Ph.D. candidate in Electrical Engineering of University of California, Los Angeles, where his research topics include protocol design, modeling and performance evaluation of transport, routing and MAC layers in ad hoc and heterogeneous networks. Mario Gerla received a graduate degree in engineering from the Politecnico di Milano in 1966, and the M.S. and Ph.D. degrees in engineering from UCLA in 197 and 1973, respectively. After working for Network Analysis Corporation from 1973 to 1976, he joined the Faculty of the Computer Science Department at UCLA where he is now Professor. His research interests cover the performance evaluation, design and control of distributed computer communication systems; high speed computer networks; wireless LANs, and; ad hoc wireless networks. He has worked on the design, implementation and testing of various wireless ad hoc network protocols (channel access, clustering, routing and transport) within the DARPA WAMIS, GloMo projects. Currently he is leading the ONR MINUTEMAN project at UCLA, and is designing a robust, scalable wireless ad hoc network architecture for unmanned intelligent agents in defense and homeland security scenarios. He is also conducting research on QoS routing, multicasting protocols and transport for the Next Generation Internet (see for recent publications). He became IEEE Fellow in 2. Yeng Zhong Lee received the M.S. Degree in Computer Science from University of California, Los Angeles in 1. He was working for HRL Laboratories, LLC (Formerly known as Hughes Research Laboratories) and Telcordia Technologies (Formerly known as Bell Communications Research) in 2 and 4, respectively. He is currently a Ph.D. candidate in Computer Science at University of California, Los Angeles. His research interests cover mobile ad hoc networks, mesh networks and wireless LANs (Bluetooth networks). He has been working on the DARPA, NSF-WHYNET, TSAT projects and most recently the ONR MINUTEMAN and ARO-MURI DAWN projects including the network protocols design and implementation for highly mobile, large scale ad-hoc networks. He has also carried out design and performance evaluation of QoS routing, multicasting protocols, transport and network security for wireless networks. (Yeng-Zhong s homepage: M. Yahya Medy Sanadidi was born in Egypt where he received his high school diploma from College Saint Marc D Alexandrie, and his B.Sc. from the University of Alexandria. He earned his M.Sc. from Penn State, and his Ph.D. in Computer Science from UCLA. He is currently an Adjunct Professor at the UCLA Computer Science Department, where he is co-principal Investigator on NSF sponsored research in high performance Internet protocols. At UCLA, he also teaches undergraduate and graduate courses on computer networks, queuing systems, and probability modeling and analysis. Dr. Sanadidi was a Manager and Senior Consulting Engineer at AT&T/Teradata, and previous to that, he held the position of Computer Scientist at Citicorp, and Assistant Professor at the Computer Science Department, University of Maryland, College Park, Madisson. Professor Sanadidi, who is a Senior Member of the IEEE and Member of the ACM, has co-authored over fourty conference and journal papers and has been awarded two best paper awards. He has served as reviewer of journal publications, member of technical program committees and organizing committees, keynote speaker, as well as chairman of professional conferences. Dr. Sanadidi has been awarded two patents, and has consulted for a number of industrial concerns. His current research interests are in path characteristics estimation and its applications in congestion control, adaptive multimedia streaming, and hybrid wire/ wireless networks.