![J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1099
performance. Then based on our findings, we pro- in turn causes severe packet interference with each
pose adaptive schemes that address the aforemen- other. When packet loss rate becomes high, the
tioned factors effectively. benefit of delaying acks, via reducing ack packets,
In standard TCP the receiver generates one ack disappears. Consequently, TCP performance
for each data packet, or more popularly, two in- degrades after reaching the peak at the optimal
order data packets with the standard delayed ack delay window size.
option [1]. This mechanism works well in wired net- Armed with the deep understanding of delayed
works. In multihop wireless networks, however, this ack impact on TCP performance over multihop
mechanism can be further improved due to: wireless links, we propose an adaptive scheme based
on the hop count of a TCP path. The basic idea
• Since the data and ack packets usually take the behind this scheme is, for a short path where inter-
same route (or spatially close routes), they inter- ference problem is minimal, we delay the ack as
fere with each other. The interference increases much as possible to maximally improve TCP
with the number of acks generated. throughput; while for a long path, we apply an
• Generating acks wastes scarce wireless resources. appropriate delay window size restriction to avoid
Though acks are essential to provide reliability, high packet loss to achieve optimal TCP perfor-
generating more acks than necessary is not desir- mance. Furthermore, we propose an end-to-end
able in wireless networks. Ideally, the receiver delay based scheme tailored for hybrid wired/wire-
should generate minimal number of acks less networks. These two schemes can be deployed
required for reliable data recovery. separately, but are also compatible with each other
to provide better performance in heterogenous net-
Recently, the delayed ack strategy has been stud- works. It is also worth noting that our proposed
ied to improve TCP performance [2,3]. However, schemes only reside in transport layers at the end
this field is not fully exploited and many issues hosts, thus no modification on intermediate nodes
remain unsolved. Some important questions is needed.
include: what is the relation between delayed acks However, TCP with untraditional delayed ack
and TCP performance, how to choose the proper (with delay window larger than 2) would cause
delay window (the number of in-order data packets friendliness to the standard TCP. Since TCP sender
to be waited for before generating an ack) for TCP, increases the congestion window (cwnd) by usually
and how to deal with the loss of acks in multihop counting the number of acks received in the current
wireless networks. In this paper we carry out a sys- popular implementations, the TCP with untradi-
tematic study to understand the effect of delayed tional delayed ack is slower in increasing the cwnd
acks on TCP over wireless links. We investigate than the standard TCP. If they coexist, the TCP
TCP performance and delayed ack related packet with untraditional delayed ack would not get fair
loss characteristics, under various wireless network share with the standard TCP. The unfriendliness
scenarios and flow patterns. problem was largely avoided in previous research
Our objective is to clearly identify the relation- and stays unsolved. In the paper, we address the
ship between TCP throughput and delayed ack over problem with a simple and intelligent end-to-end
multihop wireless links. We reveal several interest- receiver-side detection mechanism. If the receiver
ing findings, which are tremendously beneficial to detects other competing traffic, it switches to the
deeply understand TCP behavior in wireless multi- standard TCP, and vice versa. With this adaptive
hop networks, and to design enhanced TCP proto- switching scheme, the friendliness of the standard
cols. First, we found that TCP does not always TCP and our proposed adaptive delayed ack scheme
benefit from arbitrary delaying of acks. In fact, for is significantly improved.
a given network topology and flow pattern, there Another challenge for TCP with adaptive
exists an optimal delay window size at the receiver, delayed ack comes from mobile ad hoc networks.
at which TCP achieves maximum throughput. Sec- Although delayed acks may potentially improve
ond, since 802.11 does not guarantee collision free TCP throughput regardless of underlying routing
packet transmission, a packet is more likely to be protocols. Different routing mechanisms, however,
interfered with when going through a long path. If have great impact on TCP performance [4], and to
a large delay window is used over a long path, it what degree delayed acks can benefit TCP perfor-
causes large bursts of packets in transit, and this mance. The problem of more packet losses due to](https://image.slidesharecdn.com/tcp-delay1-121026005504-phpapp01/85/TCP-with-delayed-ack-for-wireless-networks-2-320.jpg)
![1100 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116
mobility could have more impact on TCP with than the transmission range, request-to-send (RTS)
untraditional delayed ack since fewer acks are gen- and clear-to-send (CTS) in 802.11 MAC cannot
erated. In this paper, we investigate the TCP-DCA ensure collision free transfer of packets. This causes
(Delayed Cumulative Ack) which addresses the hidden/exposed terminals leading to packet loss in
ack loss issue in mobile networks. Meanwhile, the ad hoc multihop wireless paths [8]. Many research
performance of various TCP schemes are presented efforts have been made to adapt TCP to the unique
over popular ad hoc routing protocols, namely, Ad- characteristic of wireless ad hoc network, e.g.
Hoc On-Demand Distance Vector Routing Transport layer ‘‘Fixed RTO’’ in [4], delayed ack
(AODV) [5], Dynamic Source Routing [6] and in [2,3], network layer support [9–11] and even
Greedy Perimeter Stateless Routing (GPSR) [7]. lower-layer assistance, for example, MAC support
The reason we include GPSR in our study is that in [8].
Geo-routing is a recently developed routing scheme Although a TCP ack packet is small, typically
promising to be scalable, and robust to node 40 bytes, the transmission of TCP ack packets
mobility. may require the same overhead as that of data pack-
While our work shares the common concept with ets in 802.11 MAC depending on the RTS thresh-
previous research in that the TCP receiver delays old. If interference from TCP acks could be
ack up to a certain number of data packets, we reduced, data packets would suffer less collisions
undergo a more complete and optimized study to resulting in higher throughput. Several approaches
understand the delayed ack impact and propose to delay acks have been proposed [2,3,12]. TCP-
more effective solutions suitable for various scenar- ADA (Adaptive Delayed Ack) [12] proposed to
ios. To the best of our knowledge, our proposed decrease the number of acks to improve TCP per-
schemes are the first existing mechanisms suitable formance. They claimed that maximum TCP
for both static/mobile wireless and hybrid networks, throughput is achieved when one ack acknowledges
and taking the ack loss and friendliness into account the full cwnd. However, the scheme did not address
as well. several important issues, such as packet loss and
The remainder of this paper is organized as fol- out-of-order packet reception. In fact, as we will
lows. Background work is provided in Section 2. show in this paper, TCP does not always benefit
A thorough study of delayed ack impact on TCP from delaying ack as much as possible.
performance is presented in Section 3 under various Altman and Jimenez [2] presented a basic delayed
network topologies and traffic patterns. We present ack scheme which was further improved in [3] where
our adaptive delayed ack (TCP-DCA) scheme in TCP-DAA (Dynamic Adaptive Acknowledgement)
Section 4. The issue of the loss of ack packets is dis- was proposed. In TCP-DAA, the receiver adjusts
cussed. Section 5 evaluates TCP-DCA on static and the delay window according to the channel condi-
mobile ad hoc networks, and hybrid wired/wireless tion (packet loss event). A TCP-DAA receiver
networks as well. Several enhancements are pre- delays acking until it receives a certain number of
sented to improve TCP-DCA’s efficiency and data packets, ranging from 2 to 4 packets. When
friendliness. Section 6 discusses a few issues to fur- there is no packet loss, the TCP-DAA receiver waits
ther improve TCP with untraditional delayed ack. for more data packets (up to 4) before generating an
We conclude the paper in Section 7. The impact of ack, but reduces the number to 2 in case of out-of-
different routing protocols on TCP is also consid- order packet arrival. However, since a TCP sender
ered in proper sections. automatically cuts the cwnd when packet loss
occurs, i.e. automatically adapting to the channel
2. Related work state at the sender side, the receiver side adaptation
provides little extra improvement. We will show
The standard TCP assumes that a packet loss is that in our adaptive delayed ack scheme, the recei-
invariably due to buffer overflow and reduces the ver does not respond dynamically to packet loss,
cwnd by half when packet loss happens. However, yet achieves better performance.
in ad hoc network, packet loss caused by buffer In [3], the ack time is set to be one average packet
overflow is rare. Instead, it is more likely due to inter-arrival time. That is, an ack is generated when
medium contention as shown in [8]. The fundamen- no data packet arrives after one average packet
tal problem resides in the limitations of IEEE 802.11 inter-arrival time since last unacknowledged data
MAC. Since the interference range is usually longer packet. In wireless networks, the inter-arrival time](https://image.slidesharecdn.com/tcp-delay1-121026005504-phpapp01/85/TCP-with-delayed-ack-for-wireless-networks-3-320.jpg)
![J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1101
is highly variable due to random MAC contention for further improvement. As it is a primarily recei-
and back-off, so it is very difficult to get accurate ver-side modification, it could be combined with
enough statistics in a complex large system. Another other sender-side schemes or low layer approaches.
impact of this timer implementation is that the For example, at sender-side, [15] studied the perfor-
receiver operates insensitively to the number of acks mance of TCP on three different routing protocols
(2–4) to be delayed because any unexpected extra and proposed a TCP modification called fixed
delayed data packet will trigger an ack. RTO, which essentially freezes the TCP RTO value
An important aspect of TCP with untraditional whenever there is a route loss. Recent work [9–11]
delayed ack is the delay window size selection. In are approaches at the network layer. Chandran
TCP-DAA [3], the receiver may delay up to four et al. [9] discussed a mechanism called TCP-Feed-
ack packets and this number is limited by the sender back, which uses route failure and re-establishment
cwnd, which is fixed at four packets. The similar notifications to provide feedback to TCP, and thus
delay window size of 3–4 packets is also picked in reduce the number of packet retransmissions and
[2] heuristically. There are issues in this scheme that TCP back-offs during route calculation, to improve
desires further clarification and improvement. First, throughput. Holland and Vaidya [10] evaluated an
although a small cwnd limit helps TCP operation in explicit link failure notification technique (ELFN)
wireless networks, it is not suitable for hybrid wired in the context of improving TCP performance over
and wireless network where a high bandwidth multi-hop ad-hoc networks. They studied the effect
delay product exists. Second, if the cwnd is not lim- of link failures due to mobility on throughput and
ited by four packets, the choice of a delay window showed through simulations that ELFN improves
size of 4 may no longer be suitable. In this paper, the performance of TCP. Xu et al. [11] proposed a
we address the above issues and provide more effec- Neighborhood RED (NRED) scheme, which
tive and general solutions that apply to various extended the RED concept to the distributed
environments. neighborhood queue, to address significant TCP
The delayed ack has impact on the TCP sender. unfairness in ad hoc wireless networks. These mech-
Since the receiver does not ack as frequently as in anisms are promising schemes to be integrated with
the standard TCP, the congestion window increas- our mechanism for more robustness against packet
ing rate for delayed ack is slowed down. Such slower loss and TCP fairness.
probing rate improves TCP performance in ad hoc
networks, as reported in [13]. The conservative win- 3. TCP throughput vs. delay window
dow increase, however, hinders efficient transmis-
sion in wired network where delay bandwidth In this section, we investigate the impact of the
product may be large. In this paper, we solve this receiver ack delay window size on TCP perfor-
problem and allow TCP-DCA to cope with ad hoc mance. The conclusion we draw is that, when the
and hybrid wired/wireless networks via adaptive path length is short, TCP achieves better perfor-
schemes. mance with a delay window as large as the entire
We also study the impact of cwnd limit at the sender cwnd. When the path length increases, how-
sender. A small sender cwnd can decrease interfer- ever, a large delay window triggers bursty transmis-
ence and maximize pipeline effect. [8] revealed that sions, which results in mutual data packet
there exists an optimal TCP cwnd size that maxi- contention and higher losses. In fact, for long paths,
mizes spatial channel reuse. Further increasing the there exists an optimal delay window size that
window size does not lead to better performance. achieves maximum TCP throughput. We verify the
On the contrary, it results in increased link layer delay ack impact using various network topologies,
contention and degraded TCP throughput. In [14], including cross and grid topologies with various
the optimal congestion window limit is determined flow patterns. Real testbed measurement results
based on the hop count for maximum pipeline effect are also presented to reinforce our conclusions.
and the interference/transmission range ratio. In the Firstly, we study a basic scheme with fixed delay
paper, we will show that in our scheme TCP-DCA, window limit to illustrate the impact of the delay
the sender’s cwnd is not limited, and yet, performs window size on TCP performance, named as TCP-
better than the case of the limited cwnd used in [3]. FDCA. We use static routing to investigate delayed
TCP-DCA resides at TCP layer, and it has poten- ack performance without any routing interference.
tial to be combined with other TCP enhancements Routing impact will be shown in Section 5.](https://image.slidesharecdn.com/tcp-delay1-121026005504-phpapp01/85/TCP-with-delayed-ack-for-wireless-networks-4-320.jpg)
![1102 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116
3.1. TCP with fixed delay window limit choose b as 0.8 and a as 1.5. The receiver also gen-
(TCP-FDCA) erates an ack within 500 ms of the arrival of an
unacknowledged data packet in accordance with
TCP-FDCA maintains a variable delay window RFC2581 [1].
‘‘w’’ representing the number of data packet arrivals One drawback to this approach is that the TCP-
before acking at the receiver. This delay window w is FDCA receiver needs to be informed of sender’s
bounded by a delay window limit, and also limited cwnd size to prevent delay window larger than cwnd
by the cwnd size to prevent the delay window from leading to unnecessary delaying at the receiver. To
exceeding the cwnd which results in delaying the address this problem, we propose two possible
ack, but for possibly non-existent packets. If data methods. One is to imbed the cwnd size in an option
packets arrive in order, the receiver generates one field of the packet header. The other is to reuse some
cumulative ack for every w data packets. If an out TCP header field. For example, the sender could
of order packet is received, or a packet that fills a reuse the advertised window field in data packet
gap in the sequence space of packets in the receiver header for ‘‘advertising’’ its cwnd size to the recei-
buffer (that is, recovery from earlier packet loss), the ver. This method is not unrealistic due to the fact
receiver acks immediately to inform the sender of that current TCP connection is mostly used for
the packet loss/recovery in a timely manner. The one direction, i.e. there is only data packets from
receiver also keeps a fall-back delay ack timer which the sender to the receiver and no backward data,
estimates the expected time to receive w data pack- so the advertised window field from the sender is
ets. At the TCP sender, when it receives a cumula- usually wasted. Even if two-way TCP data is
tive ack acknowledging w data packets, it updates included, it is still possible to reuse the advertised
the cwnd and sends out a burst of at least w packets, window field, for example every other packet with
provided such packets are ready for transmission in one bit in the reserved field to indicate whether
the sender buffer. cwnd or advertised window size is used. The lowed
To get the delay timer period, the receiver moni- frequency of cwnd notification is not expected to
tors the data packet inter-arrival interval and com- affect TCP-FDCA much since TCP-FDCA receiver
putes a smoothed interval through a low-pass does not need to know the exact cwnd so often, as
filter. The average inter-arrival interval is used to we will see later in the paper.
set the ack delay timer based on the current delay
window size. In TCP-FDCA, an accurate delay 3.2. Chain topology
timer is not needed and the timer is solely for fall
back purpose. In fact, our inter-arrival time compu- In order to understand TCP-FDCA, its perfor-
tation is rather simple and loose: the receiver sam- mance is shown over a chain topology in Ns2 [16].
ples the inter-arrival time of any consecutively The chain topology is general used to study TCP
arrived data packets, not necessarily in-order pack- performance since TCP typically runs over a single
ets. Therefore, the inter-arrival sample is an inflated path. Fig. 1 shows a chain example. The TCP con-
value compared with inter-arrival between in-order nection is sourced at the first node (node 1) and
packets. A TCP-FDCA receiver smoothes such packets travel over the chain to the end node (node
inflated inter-arrival samples through a low-pass 6). The interference range is 550 m and transmission
filter range is 250 m. We place the nodes at 200 m inter-
vals, so that nodes are 4 hops away can transmit
I iþ1 ¼ bI iavg þ ð1 À bÞI s ;
avg ð1Þ
simultaneously without interference. Notice that a
where I avg is the average of inflated packet inter-ar- node that is 3 hops away is a ‘‘hidden node’’. IEEE
rival time (I s ). I avg is used to set delay ack timer (T w ) 802.11 is the underlying MAC. Data rate on the
which defines the total time for receiving a whole de- wireless channel is 2 Mbps and one simulation run
lay window of in-order data packets lasts 300 s. Each data point represents an averaged
result of 5 simulation runs with different random
T w ¼ awI avg ; ð2Þ
seeds. The packet size is 1000 bytes and the TCP-
where w is the delay window size, a is a parameter to Newreno is adopted as it is the most popular TCP
tolerate high dynamic packet delay. Obviously, the protocol nowadays. The delay window limit applied
delay ack timer is rather inflated and quite robust at the TCP receiver ranges from 1 to the current
to high inter-arrival variation. In the paper, we cwnd size.](https://image.slidesharecdn.com/tcp-delay1-121026005504-phpapp01/85/TCP-with-delayed-ack-for-wireless-networks-5-320.jpg)
![J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1107
Table 1 mission timer period is set to 1 s. This ack retrans-
Delay window at TCP-DCA receiver mission timer is started after sending an ack, and
Path length (h) Delay window limit canceled after receiving a data packet.
h63 Congestion window Until now all previous results were obtained
3<h69 5 without ack retransmission. This is due to two facts:
h P 10 3 first, the results with and without ack retransmission
are almost the same for static networks. Mostly
4.2. Ack loss because the ack loss is rare, and the ack retransmis-
sion timer period is large (at least 1 s). Second, with-
The standard TCP generates more acks than out the ack retransmission timer, it is easy to
TCP-DCA, due to the cumulative property of the observe the sole impact of the delay window on
ack, one ack successfully received at the TCP sender TCP performance. In the following simulations,
can makes up for all the preceding lost acks. Thus, the ack retransmission timer is enabled unless
the ack loss in standard TCP is not serious. How- explicitly mentioned.
ever, the number of acks in TCP-DCA is much less
than standard TCP, and thus an ack loss has more 5. Performance evaluation
negative impact on TCP-DCA performance. In
order to be robust to ack loss, a retransmission This section presents the TCP-DCA performance
timer at the receiver is adopted to retransmit a over wireless and hybrid networks. First, we show
cumulative ack if necessary. TCP-DCA performance on static multihop wireless
A challenge here is to estimate RTT at the recei- networks and compare it with TCP-DAA in [3]. Sec-
ver which is needed for setting the ack retransmis- ond, we demonstrate TCP-DCA performance on
sion timeouts. If TCP includes two-way data, the mobile ad hoc network, and show TCP-DCA per-
receiver can get the accurate RTT measurement formance over several ad hoc routings. Third, we
since the sender acks the data immediately. If TCP extend the hop count based approach to an end-
only has one-way data, RTT measurement at the to-end delay based scheme to further improve
receiver may not be straightforward. If the TCP TCP-DCA in hybrid wired/wireless networks. Last,
timestamp option is used, the receiver could use it we propose a simple approach to address the friend-
to estimate RTT, though such an estimate may be liness problem of TCP-DCA to TCP-Newreno with
inflated if the sender does not send data packets standard delayed option.
immediately after receiving an ack [17]. However,
in TCP-DCA, the ack retransmission timer is only 5.1. Static ad hoc network
a coarse timer for predicting when to retransmit
acks, an accurate RTT measurement is not required In this section, we show TCP-DCA performance
and thus an inflated RTT can be tolerated. In the in static ad hoc networks. Meanwhile, we compare
paper, the timestamp option is used to estimate TCP-DCA with TCP-DAA and TCP-Newreno
the RTT for a TCP flow which has one-way data. (with and without standard delayed ack option
The receiver computes the ack retransmission [1]). More specially, we use the name ‘‘original
time based on this RTT measurement. We use the TCP’’ for TCP-Newreno without standard delayed
following formula to calculate the retransmission option, and TCP-STD-DA for TCP-Newreno with
timer period: standard delayed option. AODV routing is used
unless otherwise stated.
7 1 TCP-DAA [3] is an interesting extension of [2]
SRTTr ¼ SRTTr þ RTTr ;
kþ1 k kþ1
8 8 and has shown good performance in static ad hoc
3 1 networks. It is the most recent work and the major
RTTvar ¼ RTTvar þ jRTTr À SRTTr j;
kþ1 k kþ1 kþ1
4 4 differences between TCP-DAA and TCP-DCA are
r r var
RTOkþ1 ¼ SRTTkþ1 þ 4 Â RTTkþ1 ; listed in Table 2. For fair comparison, most simula-
tions scenarios presented in this subsection were
where RTTr is the RTT estimation at the receiver, shown in [3]. Each result is the average of five sim-
SRTTr is the smoothed RTT. RTTvar and RTOr ulation runs.
are RTT variance and retransmission timer period In Fig. 8, we compare TCP-DAA to TCP-DCA
at the receiver. The minimum value of the retrans- in the chain topology with different hop count and](https://image.slidesharecdn.com/tcp-delay1-121026005504-phpapp01/85/TCP-with-delayed-ack-for-wireless-networks-10-320.jpg)
![1108 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116
Table 2
Design difference
TCP-DAA TCP-DCA
Sender Duplicate acks threshold to trigger retransmission is 2; CW Puts CW into advertised window field
upper limit is 4; Retransmission (RTO) timer is increased fivefold 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 TCP-DCA is CWL and TCP-DAA achieve best performance
designed without cwnd limit, we include TCP-DCA and their difference is small. They provide better
with cwnd limit at 4 in this experiment, named TCP- performance, about 10–25%, than the standard
DCA-CWL, to show a fair comparison with TCP- TCP (with and without delayed ack option).
DAA. Note that in TCP-DAA, the cwnd limit is With static topology and steady traffic pairs, the
set to 4 by default. Fig. 8a shows the performance variance of results is very small (less than 5%) so we
of TCP-DCA over 3 hop chain. The aggregate omit its error bar results. Also note that TCP-DCA-
throughput of TCP-DCA decreases when the num- CWL does not bring significant benefit over TCP-
ber of flows increases. When there are few flows, DCA. Since the ack flow is decreased, the optimal
since each TCP-DCA generates as few acks as pos- cwnd size could actually increase (for details, please
sible, the advantage of TCP-DCA is obvious. When refer to [14]). Although it may be interesting to
more flows coexist, the cwnd of each flow becomes investigate the optimal cwnd for TCP-DCA, we
smaller since they compete for the bandwidth, thus do not pursue it because our ultimate goal to design
the benefit of delayed ack reduces because more TCP-DCA without cwnd limit. Moreover, this
acks are generated. If the cwnd is below the cwnd property makes TCP-DCA distinct since selecting
limit in TCP-DCA-CWL, they have similar perfor- an optimal cwnd is non-trivial, which needs to know
mance. The performance of TCP-STD-DA is better a lot of underlying information to accomplish. For
than original TCP, but worse than TCP-DCA and example, it requires the knowledge of transmis-
TCP-DAA. TCP-DCA and TCP-DCA-CWL pro- sion/interference ratio [14] at the physical layer
vides better performance than other TCP variants. which is usually hard to obtain. It could also poten-
Overall, TCP-DCA achieves best performance, up tially limit the application of TCP in a variety of
to 15% improvement over TCP-DAA, and 25% over wireless networks, such as that TCP with cwnd limit
TCP-STD-DA respectively. Fig. 8b–c show the per- in wired and wireless scenario does not perform well
formance of TCP-DCA on a 5 and 7 hop path. The (will be shown in Section 5.4). Taking this into
same trend persists and TCP-DCA outperforms all account, we do not choose cwnd limit in TCP-
other algorithms in most situations. DCA and thus in the following experiments, only
Note that the cwnd limit on TCP-DCA does not TCP-DCA (no cwnd limit) is adopted.
bring considerable benefit. For a short path in
Fig. 8a, the cwnd limit could considerably restrict 5.2. Mobile ad hoc network
the TCP performance as the number of flows is
small. For a long path in Fig. 8b–c, the cwnd limit We have demonstrated that applying delayed ack
only provides slightly more performance gain for 2 scheme improves TCP performance in a static net-
flows, and this advantage disappears for more flows. work, now we show that it can improve TCP perfor-
From Fig. 8, the performance of TCP-DCA without mance in a mobile network as well. Though our
cwnd limit is better than that of TCP-DCA-CWL in delayed ack scheme is not specifically designed for
most situations. It indicates that TCP-DCA can the TCP in mobile network, we show that our
achieve the distinguished performance without set- scheme works well in mobile scenarios.
ting cwnd limit. In Fig. 10, we show the performance of the origi-
We also give results for the grid topology. This nal TCP, TCP-STD-DA, TCP-DAA and TCP-
grid topology is shown in Fig. 9a and 6 flows run- DCA running over three routing schemes: GPSR,
ning on the grid. Fig. 9b compares the performance AODV and DSR. The experiment consists of 40
of original TCP, TCP-STD-DA, TCP-DAA, TCP- nodes randomly moving within a 1000 m · 1000 m
DCA-CWL and TCP-DCA, and we observe the area. Each node moves with the same speed without
similar results as before. TCP-DCA, TCP-DCA- pause. The Random Waypoint mobility model [18]](https://image.slidesharecdn.com/tcp-delay1-121026005504-phpapp01/85/TCP-with-delayed-ack-for-wireless-networks-11-320.jpg)
![1110 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116
900 800
TCP-DCA TCP-DCA
TCP TCP
800 TCP-STD-DA 700 TCP-STD-DA
TCP-DAA TCP-DAA
TCP-DCA-ACK-LOSS TCP-DCA-ACK-LOSS
700
600
Throughput (Kbps)
Throughput (Kbps)
600
500
500
400
400
300
300
200
200
100 100
0 0
GPSR AODV DSR GPSR AODV DSR
(a) 10 m/s (b) 20 m/s
Fig. 10. TCP performance over mobile network. (a) 10 m/s. (b) 20 m/s.
does not show consistent improvement across differ- acks are generated and more timeouts are likely to
ent routing protocols. It performs similarly as TCP- happen. Fig. 11 illustrates the idle time of TCP dur-
STD-DA over GPSR, but does not show better ing the whole simulation time (400 s) for different
performance than TCP (with or without standard TCP variants and routing schemes. As expected,
delayed option) over DSR. TCP-DCA (considering ack loss) has least idle time
In TCP-DCA, the receiver would probe the sen- over each routing scheme, which helps TCP-DCA
der by retransmitting acks in the event of high maintain better throughput in a mobile network.
packet loss (see Section 4.2). It prevents the sender Note that TCP-DCA-ACK-LOSS has similar idle
from entering long timeout due to the severe packet time with other TCP variants, which confirms that
loss event. In Fig. 10, we also show the result of retransmission scheme is effective in improving
TCP-DCA without ack retransmission (TCP- TCP performance in mobile networks.
DCA-ACK-LOSS) to confirm the effectiveness of
ack retransmission. When packet loss is severe 5.3. Lossy network
(e.g. TCP over DSR), ack retransmission consider-
ably boosts the TCP-DCA’s performance. On the In wireless networks, the wireless link in fact
other hand, the ack retransmission does not help could become unreliable and error-prone. After
much when packet loss is not severe (e.g. TCP over studying the performance of TCP-DCA under
GPSR). Therefore, TCP-DCA only incurs more ack packet losses in mobile networks, we investigate its
retransmission overhead when needed, and works performance under lossy wireless link. BER (bit
well with packet losses in mobile networks. error rate) is used as the loss metrics at the wireless
Comparing with static ad hoc networks in Sec- link and the error is assumed to be uniformly dis-
tion 5.1, mobility poses more challenges for TCP. tributed and independent on each link. It is worth
In a mobile network, the loss of packets (or acks) pointing out that when a packet is received with
may be caused by temporary route loss as well as bit error, the IEEE 802.11 MAC layer helps in
network congestion. Higher mobility causes more recovering packet loss by retransmissions until a
route loss and thus more packet losses. Since routes maximum retransmission limit is reached [19] or
are likely to be lost frequently in high node mobility retransmission is successful. Therefore, when BER
environments, and different routing algorithms have is low, the MAC retransmission scheme can easily
different capability to repair broken routes quickly, recover the packet and would not cause much prob-
thus different TCP performance is observed over lem for TCP. While the BER is high, the heavy
different routing schemes. Moreover, The frequent packet loss would knock down TCP performance.
route breakages could put TCP into long idle state Fig. 12 illustrates a scenario where a single flow
leading to poor TCP performance, especially for goes through 5 hop chain path under varying
TCP with untraditional delayed ack where fewer BER from very low BER to high BER. Each bar](https://image.slidesharecdn.com/tcp-delay1-121026005504-phpapp01/85/TCP-with-delayed-ack-for-wireless-networks-13-320.jpg)
![J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1111
350 TCP-DCA
350
TCP-DCA
TCP
TCP
TCP-STD-DA
TCP-STD-DA
300 TCP-DAA 300 TCP-DAA
TCP-DCA-ACK-LOSS
TCP-DCA-ACK-LOSS
250 250
Idle Time (second)
Idle Time (second)
200 200
150 150
100 100
50 50
0 0
GPSR AODV DSR GPSR AODV DSR
(a) 10 m/s (b) 20 m/s
Fig. 11. TCP idle time. (a) 10 m/s. (b) 20 m/s.
300
TCP-DCA
account the acknowledged data covered in the acks.
TCP
TCP-STD-DA Using delayed ack could potentially cause slow
TCP-DAA
250 cwnd increase and thus hurt TCP performance.
Therefore, the pure adaptive delay window based
on the hop count appears not adequate for this sce-
Throughput(Kbps)
200
nario. For example, in WLAN where 1 hop wireless
150 client can delay the cumulative ack for the whole
cwnd, the cwnd at the sender could increase slowly,
100 particularly if a large propagation delay is encoun-
tered in the wired part. Although slow increase for
50 cwnd is helpful for improving TCP performance in
ad hoc networks as shown in our previous results,
0 it is not desirable for the hybrid network. In
2e-6 6e-6 2e-5 6e-5
Bit Error Rate
RFC3465 [20] Allman proposed to increase the
cwnd based on the number of bytes acknowledged
Fig. 12. TCP performance under random loss. by the arriving acks rather than based on the num-
ber of acknowledgments that arrive at the sender in
is an averaged result of 20 random runs. Overall, the RFC2581 [1]. However, this approach potentially
performance of TCP-DCA is no worse than other causes large bursts if the delay window is large.
variants, and quite robust to wireless loss. TCP- An appropriate delay window limit needs to be
DAA also works very well unless the loss is too investigated in this scenario.
heavy. With the increasing loss rate, all TCP vari- Another challenge of the hybrid network is that
ants gracefully reduce their performance and the the hop count information from the packet header
variance of results increase. is no longer accurate for the wireless hops since it
includes the wired path. It is possible for a wireless
5.4. Hybrid (wired and wireless) network node to get hop count from the routing table to the
base station (or gateway) to the wired network,
In this section we study TCP-DCA in the wired however, this information is not guaranteed to be
and wireless ad hoc environment. Since wired net- available all the time.
work has abundant bandwidth, it demands a much To address the problems in hybrid networks, we
larger cwnd than the pure ad hoc network for effec- tailor and propose a novel receiver-side probe mech-
tive TCP performance. In current TCP implementa- anism based on CapProbe in [21], and extend the
tion, TCP sender increases the cwnd only based on previous hop-count based approach via the recei-
the number of acks received while not taking into ver-side probe in hybrid networks.](https://image.slidesharecdn.com/tcp-delay1-121026005504-phpapp01/85/TCP-with-delayed-ack-for-wireless-networks-14-320.jpg)
![1112 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116
The receiver-side probe works as follows: First Table 3
the receiver passively estimate the wireless band- IEEE 802.11b MAC overhead (with long preamble), L is packet
size, R is capacity
width and minimum RTT. Second it computes the
equivalent hop count based on the minimum RTT Data frame 192 ls + 8L/R
and wireless bandwidth, then the previous hop SIFS 10 ls
count approach can be applied afterwards. DIFS 50 ls
RTS 192 ls + 160/R
Now we address the problem to estimate the end- CTS 192 ls + 112/R
to-end hybrid path bandwidth at the TCP receiver. ACK 192 ls + 112/R
CapProbe [21] is a recently proposed bottleneck
capacity and path rate estimation technique shown
to be both fast and accurate. It combines the use
Table 4
of time interval measurements and end-to-end delay
Delay window at TCP-DCA receiver (delay-based, wireless
measurements to filter out bad packet pair probing bandwidth is at 2 Mbps and TCP data packet size is 1000 bytes)
samples. It is proved to be suitable for wired and
Minimum RTT (ms) Delay window limit
last-hop wireless networks. Inspired by CapProbe,
RTTmin < 18 Congestion window
we proposed a new approach at the TCP-DCA
18 6 RTTmin 6 60 5
receiver without triggering any probing traffic. The RTTmin > 60 3
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 Internet BS 1 2
packet pair arrives. Our approach is based on the
CapProbe, but it is different with CapProbe. First Fig. 13. Hybrid wired/wireless topology.
our approach is imbed at the TCP 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 1500
TCP-DCA
TCP
to work for wired, hybrid and purely wireless TCP-STD-DA
TCP-DAA
networks.
The other challenge is to estimate the minimum
Throughput (Kbps)
1000
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 TCP
500
data and ack packet sizes. For example, with wireless
bandwidth of 2 Mbps, data packet size of 1000 bytes
and ack size of 40 bytes, the minimum RTT for a 1
wireless hop path is about 6 ms considering various
0
overhead, such as header overhead at TCP (20 bytes), 1 hop 2 hop
IP (20 bytes) and MAC (34 bytes) and MAC layer Wireless Hop Count
fixed overhead shown in Table 3 (for details, please Fig. 14. Wired and wireless performance.
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 TCP connection from a running with different random seeds. Similar to Sec-
wired network to a wireless network with up to 2 tion 5.1, the variance of results is very small so we
wireless hops. The one-way propagation delay on omit its error bar results. Since TCP-DAA limits
the wired link is 50 ms. The performance of TCP- the cwnd, the performance of TCP-DAA is limited.
DCA, TCP-DAA and the standard TCP (with and TCP-DCA provides best performance in both cases
without standard delayed ack option) shown in of 1 and 2 hop paths. Therefore, TCP-DCA is
Fig. 14, represents averaged result of 10 simulation extensible to hybrid networks instead of working](https://image.slidesharecdn.com/tcp-delay1-121026005504-phpapp01/85/TCP-with-delayed-ack-for-wireless-networks-15-320.jpg)
![1114 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116
2
TCP-STD-DA
DCA is disabled
TCP-DCA-SWITCH
1.8 end if
else if TCP throughput > b · Path Capacity then
Instantaneous Throughput (Mbps)
1.6
if no DCA
1.4
DCA is enabled
1.2
end if
1 end if
0.8
0.6
0.4
6. Future work and discussion
0.2
In [23], ack congestion control is introduced to
0 TCP to combat bandwidth asymmetry on for-
50 100 150 200 250
Time (second) ward/backward path. Ack packet should be reduced
since backward path has limited bandwidth. How-
Fig. 16. Responsiveness of TCP-DCA-SWITCH.
ever, the ack congestion control is implemented at
the gateway with a RED queue. It is not an end-
to-end approach since the receiver needs to get
adaptive delayed ack scheme is triggered only when
ECN signal from the gateway. In contrast, our
necessary.
scheme is end-to-end, and does not require any
To detect the possible other flows, the receiver
changes at intermediate nodes.
monitors the TCP data throughput via receiver-side
Delayed ack inevitably triggers burst transporta-
probe mechanism described above. If the through-
tion at the sender. The burstiness increases the
put is much less than the path capacity, then there
packet loss and potentially hurts TCP performance.
probably has other traffic on the path. We present
Since TCP-DCA does not use large delay window
our switch algorithm in Algorithm 1. a ¼ 0:4; b ¼
except for short paths, the burstiness is limited. It
0:6 in the simulation.
is possible to further reduce the burstiness. For
Fig. 15c shows the performance of TCP-DCA
example, in [24], cwnd increase is limited by at most
with switch (TCP-DCA-SWITCH). Obviously, the
2 packets for each delayed ack, or ack reconstruc-
friendliness of TCP-DCA to TCP-STD-DA is
tion in [23] to rate control the sender. How to
much improved compared with original TCP-
further decrease the burstiness would be an interest-
DCA, especially for short paths. The TCP-DCA-
ing future work.
SWITCH can share the path with TCP-STD-DA
Another issue is the effect of delayed ack for
quite fairly.
short TCP flows. As we discussed before, using
To illustrate the transient behavior of TCP-
delayed ack could slow down the cwnd growth,
DCA-SWITCH coexisting with TCP-STD-DA,
thus delaying ack has potential negative impact
Fig. 16 plots the typical instantaneous throughput
on a short TCP flow, as illustrated in [3]. We also
in a simulation run. For this simulation set, the first
found similar results, and we are investigating on
TCP-DCA-SWITCH flow runs from the beginning
improving short flow performance with delayed
until 250 s and the second TCP-STD-DA flow starts
ack. One simple approach is to disable delaying
from 100 s to 200 s. Fig. 16 shows that TCP-DCA
ack when the cwnd is small, for example, the recei-
responds quickly to changing network conditions
ver would not delay acking if the cwnd is less than
when new flow starts at 100 s as well as when
4. It is also possible use the cwnd size to detect
flow stops at 200 s. Using TCP-DCA-SWITCH, a
slow start phase or let sender notify the receiver
TCP flow utilizes the available bandwidth when
if it is in slow start phase. If the sender is detected
there are no competing flows and shares the band-
to be in slow start, the receiver temporarily disables
width fairly when multiple flows compete for the
the delayed ack scheme to make TCP achieve good
bandwidth.
throughput at connection startup. After TCP leaves
Algorithm 1. Switch Algorithm the slow start, TCP receiver begins delaying ack to
better utilize the shared wireless medium. These
if TCP throughput < a · Path Capacity then enhancements are expected to be beneficial for
if DCA is enabled then short flows because most of short flows finish their](https://image.slidesharecdn.com/tcp-delay1-121026005504-phpapp01/85/TCP-with-delayed-ack-for-wireless-networks-17-320.jpg)
![J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1115
transfer while they are still in the slow start phase [5] C.E. Perkins, E.M. Royer, Ad-hoc on-demand distance
[25]. vector routing, in: Proceedings of IEEE WMCSA’99, Feb-
ruary 1999.
[6] D.B. Johnson, D.A. Maltz, in: in: T. Imielinski, H. Korth
(Eds.), Dynamic Source Routing in Ad hoc Wireless
7. Conclusion Networks, Mobile Computing (chapter 5).
[7] B. Karp, H.T. Kung, GPSR: greedy perimeter stateless
TCP, a dominant reliable transport protocol in routing for wireless networks, in: Proceedings of ACM
wired networks, is a highly competitive candidate MobiCom’00, August 2000.
for providing reliable data transport in wireless [8] Z. Fu, P. Zerfos, H. Luo, S. Lu, L. Zhang, M. Gerla, The
impact of multihop wireless channel on TCP throughput and
and wired/wireless networks. This paper systemati- loss, in: IEEE INFOCOM’03, March 2003.
cally examines the relationship of TCP performance [9] K. Chandran, S. Raghunathan, S. Venkatesan, R. Prakash,
to delayed ack via analysis, and simulation and test- A feedback-based scheme for improving TCP performance
bed experiments. We find that TCP does not always in ad hoc wireless networks, IEEE Personal Communica-
get throughput gain by delaying unlimited acks. The tions Magazine 8 (1) (2001).
[10] G. Holland, N.H. Vaidya, Analysis of TCP performance
maximal TCP throughput is achieved at a certain over mobile ad hoc networks, in: Proceedings of ACM
delay window that balances decreasing ack flow MobiCom’99, August 1999.
and burst loss. Thus, we introduced two novel adap- [11] K. Xu, M. Gerla, L. Qi, Y. Shu, Enhancing TCP fairness in
tive delayed ack mechanisms compatible with TCP ad hoc wireless networks using neighborhood red, in:
called TCP-DCA, based on the path hop length Mobicom, 2003.
[12] A.K. Singh, K. Kankipati, Tcp-ADA: Tcp with adaptive
and/or end-to-end delay, to maximize the TCP per- delayed acknowledgement for mobile ad hoc networks, in:
formance for MANET and hybrid networks. Our WCNC, 2004.
proposed schemes are shown to have superior per- [13] K. Nahm, A. Helmy, C.J. Kuo, Tcp over multihop 802.11
formance in static networks (wireless or hybrid networks: issues and performance enhancement, in: Mobi-
wired/wireless networks). We also demonstrated hoc, 2005.
[14] K. Chen, Y. Xue, K. Nahrstedt, On setting TCP’s congestion
better TCP performance with our proposed mecha- window limit in mobile ad hoc networks, in: ICC, 2003.
nisms over different routing protocols in mobile ad [15] T.D. Dyer, R.V. Boppana, A comparison of TCP perfor-
hoc networks. mance over three routing protocols for mobile ad hoc
It is well known that TCP with untraditional networks, in: Proceedings of ACM MobiHoc’01, October
delayed ack is less aggressive than TCP-Newreno 2001.
[16] The network simulator - ns2, Available from: http://
since the cwnd with untraditional delayed ack grows www.isi.edu/nsnam/ns/.
much slower. To combat such unfriendliness, we [17] V. Jacobson, Tcp extensions for high performance, RFC
propose a receiver-side probe and detection to 1323, 1992.
maximally increase the TCP-DCA friendliness [18] C. Bettstetter, H. Hartenstein, X. Prez-Costa, Stochastic
(when other flows exist) and efficiency (when no properties of the random waypoint mobility model, ACM/
Kluwer Wireless Networks: Special Issue on Modeling and
other flows exist). The detection is imbed in the Analysis of Mobile Networks, September 2004.
TCP-DCA and does not need to modify [19] IEEE-802.11, Wireless LAN media access control (MAC)
underlying layers. TCP-DCA friendliness to TCP- and physical layer (PHY) specifications, Available from:
Newreno is significantly enhanced with this simple http://standards.ieee.org/getieee802, 1999.
approach. [20] M. Allman, Tcp congestion control with appropriate byte
counting (abc), RFC 3465, 2003.
[21] L.L.M.G. Rohit Kapoor, Ling-Jyh Chen, M.Y. Sanadidi,
Capprobe: a simple and accurate capacity estimation tech-
References nique, in: ACM SIGCOMM, 2004.
[22] M. Kappes, An experimental performance analysis of mac
[1] M. Altman, Tcp congestion control, RFC 2581 (1999). multicast in 802.11b networks for voip traffic, in: SPECTS,
[2] E. Altman, T. Jimenez, Novel delayed ack techniques for 2004.
improving TCP performance in multihop wireless networks, [23] H. Balakrishnan, V.N. Padmanabhan, R.H. Katz, The
Personal Wireless Communications, September 2003. effects of asymmetry on TCP performance, in: Mobicom,
[3] R. de Oliveira, T. Braun, A dynamic adaptive acknowledg- 1997.
ment strategy for TCP over multihop wireless networks, in: [24] M. Allman, On the generation and use of TCP acknowledg-
Infocom, 2005. ments, ACM Computer Communication Review, 1998.
[4] T.D. Dyer, R.V. Boppana, A comparison of TCP perfor- [25] H. Balakrishnan, S. Seshan, M. Stemm, R. Katz, Tcp
mance over three routing protocols for mobile ad hoc behavior of a busy internet server: Analysis and improve-
networks, in: Mobihoc, 2001. ments, in: Infocom, 1998.](https://image.slidesharecdn.com/tcp-delay1-121026005504-phpapp01/85/TCP-with-delayed-ack-for-wireless-networks-18-320.jpg)
TCP with delayed ack for wireless networks
- 1. Available online at www.sciencedirect.com Ad Hoc Networks 6 (2008) 1098–1116 www.elsevier.com/locate/adhoc TCP with delayed ack for wireless networks Jiwei Chen *, Mario Gerla, Yeng Zhong Lee, M.Y. Sanadidi University of California, Los Angeles, CA 90095, United States Received 5 May 2006; received in revised form 25 October 2007; accepted 30 October 2007 Available online 6 November 2007 Abstract This paper studies the TCP performance with delayed ack in wireless networks (including ad hoc and WLANs) which use IEEE 802.11 MAC protocol as the underlying medium access control. Our analysis and simulations show that TCP 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 TCP 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 bur- sty 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 TCP 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 suit- able 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 TCP 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 detec- tion is proposed to improve friendliness between the standard TCP and our proposed TCP with adaptive delayed ack. Ó 2007 Elsevier B.V. All rights reserved. Keywords: TCP; Congestion control; Adaptive delayed ack; Wireless medium access/contention; Friendliness 1. Introduction networks, several other inherent factors attribute to the TCP performance deterioration, including The Transport Control Protocol (TCP) is the unpredictable channel errors, medium access con- most widely used reliable transport protocol over tention complicated by hidden/exposed terminal the Internet. TCP was originally designed for wired problems, and frequent route breakages caused by links where buffer overflows account for most node mobility. All these factors pose great chal- packet losses. However, in multihop ad hoc wireless lenges to the design of TCP protocols to provide efficient and reliable end-to-end communications. * Many researchers have made valiant effort to pro- Corresponding author. E-mail addresses: [email protected] (J. Chen), gerla@cs. pose various methods to make TCP survive in such ucla.edu (M. Gerla), [email protected] (Y.Z. Lee), medy@cs. challenging environments. In this paper, we focus ucla.edu (M.Y. Sanadidi). on studying the effect of delayed acks on TCP 1570-8705/$ - see front matter Ó 2007 Elsevier B.V. All rights reserved. doi:10.1016/j.adhoc.2007.10.004
- 2. J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1099 performance. Then based on our findings, we pro- in turn causes severe packet interference with each pose adaptive schemes that address the aforemen- other. When packet loss rate becomes high, the tioned factors effectively. benefit of delaying acks, via reducing ack packets, In standard TCP the receiver generates one ack disappears. Consequently, TCP performance for each data packet, or more popularly, two in- degrades after reaching the peak at the optimal order data packets with the standard delayed ack delay window size. option [1]. This mechanism works well in wired net- Armed with the deep understanding of delayed works. In multihop wireless networks, however, this ack impact on TCP performance over multihop mechanism can be further improved due to: wireless links, we propose an adaptive scheme based on the hop count of a TCP path. The basic idea • Since the data and ack packets usually take the behind this scheme is, for a short path where inter- same route (or spatially close routes), they inter- ference problem is minimal, we delay the ack as fere with each other. The interference increases much as possible to maximally improve TCP with the number of acks generated. throughput; while for a long path, we apply an • Generating acks wastes scarce wireless resources. appropriate delay window size restriction to avoid Though acks are essential to provide reliability, high packet loss to achieve optimal TCP perfor- generating more acks than necessary is not desir- mance. Furthermore, we propose an end-to-end able in wireless networks. Ideally, the receiver delay based scheme tailored for hybrid wired/wire- should generate minimal number of acks less networks. These two schemes can be deployed required for reliable data recovery. separately, but are also compatible with each other to provide better performance in heterogenous net- Recently, the delayed ack strategy has been stud- works. It is also worth noting that our proposed ied to improve TCP performance [2,3]. However, schemes only reside in transport layers at the end this field is not fully exploited and many issues hosts, thus no modification on intermediate nodes remain unsolved. Some important questions is needed. include: what is the relation between delayed acks However, TCP with untraditional delayed ack and TCP performance, how to choose the proper (with delay window larger than 2) would cause delay window (the number of in-order data packets friendliness to the standard TCP. Since TCP sender to be waited for before generating an ack) for TCP, increases the congestion window (cwnd) by usually and how to deal with the loss of acks in multihop counting the number of acks received in the current wireless networks. In this paper we carry out a sys- popular implementations, the TCP with untradi- tematic study to understand the effect of delayed tional delayed ack is slower in increasing the cwnd acks on TCP over wireless links. We investigate than the standard TCP. If they coexist, the TCP TCP performance and delayed ack related packet with untraditional delayed ack would not get fair loss characteristics, under various wireless network share with the standard TCP. The unfriendliness scenarios and flow patterns. problem was largely avoided in previous research Our objective is to clearly identify the relation- and stays unsolved. In the paper, we address the ship between TCP throughput and delayed ack over problem with a simple and intelligent end-to-end multihop wireless links. We reveal several interest- receiver-side detection mechanism. If the receiver ing findings, which are tremendously beneficial to detects other competing traffic, it switches to the deeply understand TCP behavior in wireless multi- standard TCP, and vice versa. With this adaptive hop networks, and to design enhanced TCP proto- switching scheme, the friendliness of the standard cols. First, we found that TCP does not always TCP and our proposed adaptive delayed ack scheme benefit from arbitrary delaying of acks. In fact, for is significantly improved. a given network topology and flow pattern, there Another challenge for TCP with adaptive exists an optimal delay window size at the receiver, delayed ack comes from mobile ad hoc networks. at which TCP achieves maximum throughput. Sec- Although delayed acks may potentially improve ond, since 802.11 does not guarantee collision free TCP throughput regardless of underlying routing packet transmission, a packet is more likely to be protocols. Different routing mechanisms, however, interfered with when going through a long path. If have great impact on TCP performance [4], and to a large delay window is used over a long path, it what degree delayed acks can benefit TCP perfor- causes large bursts of packets in transit, and this mance. The problem of more packet losses due to
- 3. 1100 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 mobility could have more impact on TCP with than the transmission range, request-to-send (RTS) untraditional delayed ack since fewer acks are gen- and clear-to-send (CTS) in 802.11 MAC cannot erated. In this paper, we investigate the TCP-DCA ensure collision free transfer of packets. This causes (Delayed Cumulative Ack) which addresses the hidden/exposed terminals leading to packet loss in ack loss issue in mobile networks. Meanwhile, the ad hoc multihop wireless paths [8]. Many research performance of various TCP schemes are presented efforts have been made to adapt TCP to the unique over popular ad hoc routing protocols, namely, Ad- characteristic of wireless ad hoc network, e.g. Hoc On-Demand Distance Vector Routing Transport layer ‘‘Fixed RTO’’ in [4], delayed ack (AODV) [5], Dynamic Source Routing [6] and in [2,3], network layer support [9–11] and even Greedy Perimeter Stateless Routing (GPSR) [7]. lower-layer assistance, for example, MAC support The reason we include GPSR in our study is that in [8]. Geo-routing is a recently developed routing scheme Although a TCP ack packet is small, typically promising to be scalable, and robust to node 40 bytes, the transmission of TCP ack packets mobility. may require the same overhead as that of data pack- While our work shares the common concept with ets in 802.11 MAC depending on the RTS thresh- previous research in that the TCP receiver delays old. If interference from TCP acks could be ack up to a certain number of data packets, we reduced, data packets would suffer less collisions undergo a more complete and optimized study to resulting in higher throughput. Several approaches understand the delayed ack impact and propose to delay acks have been proposed [2,3,12]. TCP- more effective solutions suitable for various scenar- ADA (Adaptive Delayed Ack) [12] proposed to ios. To the best of our knowledge, our proposed decrease the number of acks to improve TCP per- schemes are the first existing mechanisms suitable formance. They claimed that maximum TCP for both static/mobile wireless and hybrid networks, throughput is achieved when one ack acknowledges and taking the ack loss and friendliness into account the full cwnd. However, the scheme did not address as well. several important issues, such as packet loss and The remainder of this paper is organized as fol- out-of-order packet reception. In fact, as we will lows. Background work is provided in Section 2. show in this paper, TCP does not always benefit A thorough study of delayed ack impact on TCP from delaying ack as much as possible. performance is presented in Section 3 under various Altman and Jimenez [2] presented a basic delayed network topologies and traffic patterns. We present ack scheme which was further improved in [3] where our adaptive delayed ack (TCP-DCA) scheme in TCP-DAA (Dynamic Adaptive Acknowledgement) Section 4. The issue of the loss of ack packets is dis- was proposed. In TCP-DAA, the receiver adjusts cussed. Section 5 evaluates TCP-DCA on static and the delay window according to the channel condi- mobile ad hoc networks, and hybrid wired/wireless tion (packet loss event). A TCP-DAA receiver networks as well. Several enhancements are pre- delays acking until it receives a certain number of sented to improve TCP-DCA’s efficiency and data packets, ranging from 2 to 4 packets. When friendliness. Section 6 discusses a few issues to fur- there is no packet loss, the TCP-DAA receiver waits ther improve TCP with untraditional delayed ack. for more data packets (up to 4) before generating an We conclude the paper in Section 7. The impact of ack, but reduces the number to 2 in case of out-of- different routing protocols on TCP is also consid- order packet arrival. However, since a TCP sender ered in proper sections. automatically cuts the cwnd when packet loss occurs, i.e. automatically adapting to the channel 2. Related work state at the sender side, the receiver side adaptation provides little extra improvement. We will show The standard TCP assumes that a packet loss is that in our adaptive delayed ack scheme, the recei- invariably due to buffer overflow and reduces the ver does not respond dynamically to packet loss, cwnd by half when packet loss happens. However, yet achieves better performance. in ad hoc network, packet loss caused by buffer In [3], the ack time is set to be one average packet overflow is rare. Instead, it is more likely due to inter-arrival time. That is, an ack is generated when medium contention as shown in [8]. The fundamen- no data packet arrives after one average packet tal problem resides in the limitations of IEEE 802.11 inter-arrival time since last unacknowledged data MAC. Since the interference range is usually longer packet. In wireless networks, the inter-arrival time
- 4. J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1101 is highly variable due to random MAC contention for further improvement. As it is a primarily recei- and back-off, so it is very difficult to get accurate ver-side modification, it could be combined with enough statistics in a complex large system. Another other sender-side schemes or low layer approaches. impact of this timer implementation is that the For example, at sender-side, [15] studied the perfor- receiver operates insensitively to the number of acks mance of TCP on three different routing protocols (2–4) to be delayed because any unexpected extra and proposed a TCP modification called fixed delayed data packet will trigger an ack. RTO, which essentially freezes the TCP RTO value An important aspect of TCP with untraditional whenever there is a route loss. Recent work [9–11] delayed ack is the delay window size selection. In are approaches at the network layer. Chandran TCP-DAA [3], the receiver may delay up to four et al. [9] discussed a mechanism called TCP-Feed- ack packets and this number is limited by the sender back, which uses route failure and re-establishment cwnd, which is fixed at four packets. The similar notifications to provide feedback to TCP, and thus delay window size of 3–4 packets is also picked in reduce the number of packet retransmissions and [2] heuristically. There are issues in this scheme that TCP back-offs during route calculation, to improve desires further clarification and improvement. First, throughput. Holland and Vaidya [10] evaluated an although a small cwnd limit helps TCP operation in explicit link failure notification technique (ELFN) wireless networks, it is not suitable for hybrid wired in the context of improving TCP performance over and wireless network where a high bandwidth multi-hop ad-hoc networks. They studied the effect delay product exists. Second, if the cwnd is not lim- of link failures due to mobility on throughput and ited by four packets, the choice of a delay window showed through simulations that ELFN improves size of 4 may no longer be suitable. In this paper, the performance of TCP. Xu et al. [11] proposed a we address the above issues and provide more effec- Neighborhood RED (NRED) scheme, which tive and general solutions that apply to various extended the RED concept to the distributed environments. neighborhood queue, to address significant TCP The delayed ack has impact on the TCP sender. unfairness in ad hoc wireless networks. These mech- Since the receiver does not ack as frequently as in anisms are promising schemes to be integrated with the standard TCP, the congestion window increas- our mechanism for more robustness against packet ing rate for delayed ack is slowed down. Such slower loss and TCP fairness. probing rate improves TCP performance in ad hoc networks, as reported in [13]. The conservative win- 3. TCP throughput vs. delay window dow increase, however, hinders efficient transmis- sion in wired network where delay bandwidth In this section, we investigate the impact of the product may be large. In this paper, we solve this receiver ack delay window size on TCP perfor- problem and allow TCP-DCA to cope with ad hoc mance. The conclusion we draw is that, when the and hybrid wired/wireless networks via adaptive path length is short, TCP achieves better perfor- schemes. mance with a delay window as large as the entire We also study the impact of cwnd limit at the sender cwnd. When the path length increases, how- sender. A small sender cwnd can decrease interfer- ever, a large delay window triggers bursty transmis- ence and maximize pipeline effect. [8] revealed that sions, which results in mutual data packet there exists an optimal TCP cwnd size that maxi- contention and higher losses. In fact, for long paths, mizes spatial channel reuse. Further increasing the there exists an optimal delay window size that window size does not lead to better performance. achieves maximum TCP throughput. We verify the On the contrary, it results in increased link layer delay ack impact using various network topologies, contention and degraded TCP throughput. In [14], including cross and grid topologies with various the optimal congestion window limit is determined flow patterns. Real testbed measurement results based on the hop count for maximum pipeline effect are also presented to reinforce our conclusions. and the interference/transmission range ratio. In the Firstly, we study a basic scheme with fixed delay paper, we will show that in our scheme TCP-DCA, window limit to illustrate the impact of the delay the sender’s cwnd is not limited, and yet, performs window size on TCP performance, named as TCP- better than the case of the limited cwnd used in [3]. FDCA. We use static routing to investigate delayed TCP-DCA resides at TCP layer, and it has poten- ack performance without any routing interference. tial to be combined with other TCP enhancements Routing impact will be shown in Section 5.
- 5. 1102 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 3.1. TCP with fixed delay window limit choose b as 0.8 and a as 1.5. The receiver also gen- (TCP-FDCA) erates an ack within 500 ms of the arrival of an unacknowledged data packet in accordance with TCP-FDCA maintains a variable delay window RFC2581 [1]. ‘‘w’’ representing the number of data packet arrivals One drawback to this approach is that the TCP- before acking at the receiver. This delay window w is FDCA receiver needs to be informed of sender’s bounded by a delay window limit, and also limited cwnd size to prevent delay window larger than cwnd by the cwnd size to prevent the delay window from leading to unnecessary delaying at the receiver. To exceeding the cwnd which results in delaying the address this problem, we propose two possible ack, but for possibly non-existent packets. If data methods. One is to imbed the cwnd size in an option packets arrive in order, the receiver generates one field of the packet header. The other is to reuse some cumulative ack for every w data packets. If an out TCP header field. For example, the sender could of order packet is received, or a packet that fills a reuse the advertised window field in data packet gap in the sequence space of packets in the receiver header for ‘‘advertising’’ its cwnd size to the recei- buffer (that is, recovery from earlier packet loss), the ver. This method is not unrealistic due to the fact receiver acks immediately to inform the sender of that current TCP connection is mostly used for the packet loss/recovery in a timely manner. The one direction, i.e. there is only data packets from receiver also keeps a fall-back delay ack timer which the sender to the receiver and no backward data, estimates the expected time to receive w data pack- so the advertised window field from the sender is ets. At the TCP sender, when it receives a cumula- usually wasted. Even if two-way TCP data is tive ack acknowledging w data packets, it updates included, it is still possible to reuse the advertised the cwnd and sends out a burst of at least w packets, window field, for example every other packet with provided such packets are ready for transmission in one bit in the reserved field to indicate whether the sender buffer. cwnd or advertised window size is used. The lowed To get the delay timer period, the receiver moni- frequency of cwnd notification is not expected to tors the data packet inter-arrival interval and com- affect TCP-FDCA much since TCP-FDCA receiver putes a smoothed interval through a low-pass does not need to know the exact cwnd so often, as filter. The average inter-arrival interval is used to we will see later in the paper. set the ack delay timer based on the current delay window size. In TCP-FDCA, an accurate delay 3.2. Chain topology timer is not needed and the timer is solely for fall back purpose. In fact, our inter-arrival time compu- In order to understand TCP-FDCA, its perfor- tation is rather simple and loose: the receiver sam- mance is shown over a chain topology in Ns2 [16]. ples the inter-arrival time of any consecutively The chain topology is general used to study TCP arrived data packets, not necessarily in-order pack- performance since TCP typically runs over a single ets. Therefore, the inter-arrival sample is an inflated path. Fig. 1 shows a chain example. The TCP con- value compared with inter-arrival between in-order nection is sourced at the first node (node 1) and packets. A TCP-FDCA receiver smoothes such packets travel over the chain to the end node (node inflated inter-arrival samples through a low-pass 6). The interference range is 550 m and transmission filter range is 250 m. We place the nodes at 200 m inter- vals, so that nodes are 4 hops away can transmit I iþ1 ¼ bI iavg þ ð1 À bÞI s ; avg ð1Þ simultaneously without interference. Notice that a where I avg is the average of inflated packet inter-ar- node that is 3 hops away is a ‘‘hidden node’’. IEEE rival time (I s ). I avg is used to set delay ack timer (T w ) 802.11 is the underlying MAC. Data rate on the which defines the total time for receiving a whole de- wireless channel is 2 Mbps and one simulation run lay window of in-order data packets lasts 300 s. Each data point represents an averaged result of 5 simulation runs with different random T w ¼ awI avg ; ð2Þ seeds. The packet size is 1000 bytes and the TCP- where w is the delay window size, a is a parameter to Newreno is adopted as it is the most popular TCP tolerate high dynamic packet delay. Obviously, the protocol nowadays. The delay window limit applied delay ack timer is rather inflated and quite robust at the TCP receiver ranges from 1 to the current to high inter-arrival variation. In the paper, we cwnd size.
- 6. J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1103 1600 1hop 2hop 3hop 1400 Throughput (Kbps) 1200 1 2 3 4 5 6 1000 800 600 Fig. 1. Chain topology. The solid-line circle denotes a node’s 400 valid transmission range. The dotted-line circle denotes a node’s 1 2 3 4 5 6 7 8 9 10 ...CW Delay Window Limit (packets) interference range. Node 4’s transmission will interfere with node 1’s transmissions to node 2. (a) Path Length h ≤ 3 350 4hop 5hop 6hop 3.2.1. Single TCP flow 7hop 8hop TCP performance over wireless multihop inevita- 9hop bly depends on the path length (hop count). The Throughput (Kbps) 300 longer the path is, the lower the throughput would be. We present TCP-FDCA throughput as a func- tion of delay window limits on a path with different lengths in Fig. 2. Fig. 2a shows that when a sender 250 and a receiver are within 3 hops, TCP-FDCA gets steady performance gain by increasing the delay window size up to the entire cwnd size. For the 200 one hop case, the fastest throughput increase can 1 2 3 4 5 6 7 8 9 10 ...CW be seen when the delay window is small, say less Delay Window Limit (packets) than 5. The increasing trend becomes slower for (b) Path Length 10 > h >3 delay windows larger than 5 and the throughput 240 approaches the limit when the delay window reaches 10hop 12hop the cwnd size (denoted by CW). The same trend is 230 14hop 16hop also observed when the path length is 2 and 3. 220 18hop 20hop The reason for this performance gain is that Throughput (Kbps) 802.11 MAC provides nearly collision free packet 210 transmissions for short paths. Since the interference 200 range is larger than 2 times the transmission range, 190 when the sender and receiver are within 2 hops, every node along the path can sense all other nodes 180 transmissions. In this case, no hidden nodes exist, 170 thus packet loss is minimal. When the hop length is 3, the TCP receiver is the only node hidden from 160 1 2 3 4 5 6 7 8 9 10 ...CW the TCP sender. If the receiver acks only after Delay Window Limit (packets) receiving all packets in a cwnd, no data packet (c) Path Length 20 ≥ h ≥ 10 can be interfered. The maximal performance gain is achieved when the receiver acks after receiving Fig. 2. TCP-FDCA throughput vs. delay window on chain all packets in a cwnd, and attains up to 25% topology. (a) Path length h 6 3. (b) Path length 10 > h > 3. (c) Path length 20 P h P 10. throughput gain relative to the TCP with delay win- dow at 1. Since the interference range is larger than the solve the hidden/exposed terminal problem. As the transmission range, RTS/CTS cannot completely path becomes longer than 3 hops, packet collision
- 7. 1104 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 is unavoidable. For example, node 1 and node 4 are ver are used for 802.11b devices and the devices are interfering nodes in Fig. 1, so simultaneous packet set to ad-hoc mode. The topology of the testbed is a transmissions on them will be interfered. The hidden chain and the route is statically configured. There is terminal results in packet loss and this problem one source laptop and one selected destination lap- becomes more severe for longer paths because a top among other 5 laptops depending on the num- packet has more chances to be interfered with. ber of hops in our experiments. We use ‘‘Iperf’’ to Another problem is that when the sender receives generate FTP traffic with packet size of 1000 bytes. a cumulative ack, it immediately sends a burst of From the experiment results shown in Fig. 3, we packets. The larger the delay window, the larger confirm that the TCP-FDCA throughput vs. recei- the burst and the more packet loss due to increased ver delay window follows the same trends as in interfering. (Note that the packet burst size is the simulation results shown in Fig. 1. directly related with the receiver delay window since Fig. 4 compares TCP-FDCA throughput for 1 it is at least equal to the size of the delay window.) hop and 5 hops paths from simulation and actual When the packet loss becomes high, the TCP- testbed, when delay window limit is equal to 1 and FDCA throughput gain from delaying ack is lost cwnd respectively. We note that the average differ- due to the increased packet losses. Fig. 2(b) shows ence between the testbed TCP throughput and the this tradeoff of TCP-FDCA performance gain with simulation results is below 10%. Such difference is delay window size for paths larger than 3 hops. When the hop count is 4 or 5, we do observe unsuc- cessful packet transmissions caused by interference, 1600 however, since a TCP sender is able to recover 1hop 2hop packet loss rather rapidly due to the small round 3hop 1400 trip time (RTT), TCP-FDCA maintains perfor- mance gain by delaying ack for more data packets. 1200 Throughput (Kbps) After peak performance at certain delay window, the TCP-FDCA performance gets worse due to 1000 more packet losses. Similar trend also exists for paths longer than 5 hops, where TCP-FDCA 800 achieves throughput gain only when the delay win- dow size is small; for large delay window size, delay- 600 ing ack cannot maintain throughput gain because of excessive data packet losses. Further, now that RTT 400 is larger, TCP-FDCA spends more time detecting 1 2 3 4 5 ...CW Delay Window Size (packets) packet loss and recovering lost packets by entering fast retransmit/recovery in which only one packet (a) 1 to 3 hop is recovered per RTT. Therefore, for long paths, 370 large delay window is not preferred. 4hop 5hop 360 We also show TCP-FDCA performance over a 350 very long path h P 10 in Fig. 2(c) which has similar results with Fig. 2(b). TCP-FDCA only gets perfor- 340 Throughput (Kbps) mance gain for small delay window size. For large 330 delay window size, TCP-FDCA gets poor 320 throughput. 310 300 3.2.2. Real testbed verification We investigate the effectiveness of our TCP- 290 FDCA in an actual ad-hoc network testbed. Our 280 testbed consists of six laptops equipped with Ori- 270 noco 802.11b PCMICA cards with channel rate of 1 2 3 4 5 ...CW Delay Window Size (packets) 2Mbps. The laptops run Mandrake Linux distribu- (b) 4 and 5 hop tion 7 with kernel version 2.4.3. Linux PCMCIA package version 3.2.0 and Orinoco wavelan2-cs dri- Fig. 3. Real testbed measurement. (a) 1 to 3 hop. (b) 4 and 5 hop.
- 8. J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1105 1500 d 3.3. More complex topologies and flow patterns 1 dCW We expand our study to scenarios of more com- plex topologies and flow patterns, including cross Total Throughput (Kbps) 1000 and grid topologies. For all cases, we observe the similar results to what is described above, i.e., when the path is short, TCP-FDCA achieves optimal throughput by delaying acks as much as possible, 500 up to the entire sender’s cwnd. On the other hand, when the path becomes longer, a large delay win- dow hinders performance gain and deteriorates TCP throughput gradually. We show that there 0 exists a certain delay window size, at which TCP- Simulation Experiment FDCA achieves optimal throughput performance. (a) 1 hop The following provides a summary of results on 300 multiple flows and various topologies. The perfor- d 1 mance over random topologies will be shown in Sec- d 250 CW tion 5.2. Total Throughput (Kbps) 200 3.3.1. Multiple TCP flows Fig. 5 exhibits result for 5 concurrent TCP- FDCA flows on a chain topology with hop count 150 varying from 3 to 7, and has similar results to Fig. 2. 100 3.3.2. Cross and grid topology Fig. 6 shows examples of cross and grid topolo- 50 gies with a 4 hop path for each TCP-FDCA flow, and the results with different delay window size are 0 Simulation Experiment shown in Fig. 7. Additionally the results from vary- (b) 5 hop ing path lengths are shown. Similar trend is observed again. We also run extensive simulations on cross Fig. 4. Simulation vs. experiment. (a) 1 hop. (b) 5 hop. and grid topologies with multiple overlapped flows instead of one flow. The results, not included here, mainly caused by parameter setting in the testbed also confirm the same trends discussed above. measurements. In our testbed experiments, RTS/ CTS control packets are adopted for unicast data 500 packets to overcome the hidden terminal problem 3hop 5hop only if the packet size is above the minimum thresh- 7hop 450 old 256 bytes. Since the size of TCP ack packet Aggregate Throughput (Kbps) (40 bytes) is below the minimum RTS/CTS threshold 400 in Linux settings so that no RTS/CTS handshake is performed when transmitting acks in testbed experi- 350 ments. Compared with simulation results where RTS/CTS is always performed regardless of packet 300 size, such ‘‘zero’’ MAC overhead in testbed measure- ments has impact on the performance: the through- 250 put gain derived from TCP-FDCA in the testbed measurements is lower than the corresponding results 200 in the simulations because of the lack of RTS/CTS 1 2 3 4 5 6 7 8 9 10 ...CW reduction for acks in simulations. Other than this Delay Window Limit (packets) slight difference, the results are similar in simulation Fig. 5. TCP-FDCA throughput vs. delay window on chain and in testbed measurements. topology (5 flows).
- 9. 1106 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 8 Flow 1 7 TCP flow 1 0 1 2 3 4 Flow 2 TCP flow 2 6 5 Flow 3 (a) Cross To pology (b) Gri d Topology Fig. 6. More complex topologies. Left: cross topology with 4 hop path on each direction. 200 m distance between two adjacent nodes. Right: 5 · 5 grid topology, 200 m distance between horizontal and vertical adjacent nodes. (a) Cross topology. (b) Grid topology. 340 340 4hop 5x5 6hop 5x7 8hop 5x9 320 320 Aggregate Throughput (Kbps) Aggregate Throughput (Kbps) 300 300 280 280 260 260 240 240 220 220 200 200 1 2 3 4 5 6 7 8 9 10 ...CW 1 2 3 4 5 6 7 8 9 10 ...CW Delay Window Limit (packets) Delay Window Limit (packets) (a) Cross Topology (b) Grid Topology Fig. 7. TCP-FDCA throughput vs. delay window on complex topologies. (a) Cross topology. (b) Grid topology. 4. TCP with adaptive delayed cumulative ack vation, we dynamically determine the delay window (TCP-DCA) size based on the hop count of a TCP connection and call it as TCP-DCA. For a short path (h 6 3), In this section we propose TCP-DCA with adap- TCP-DCA could delay ack for a whole cwnd to tive delayed window according to the underlying get best performance; otherwise a small delay win- path information to optimize TCP performance. dow for long paths. Since TCP still gets slight ben- The issue of ack loss is also discussed and a simple efit by delaying more for moderate long paths receiver retransmission scheme is proposed. (3 < h 6 9), therefore the delay window could be slightly larger for such paths. TCP-DCA receiver 4.1. TCP-DCA with adaptive delayed ack can get the path length information from the TTL field in TCP packet header or from the routing Up to now we have presented the tradeoff layer at the receiver node Table 1 lists the delay win- between TCP-FDCA throughput and delay window dow size applied in TCP-DCA according to the path size on different path lengths. Inspired by this obser- length.
- 10. J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1107 Table 1 mission timer period is set to 1 s. This ack retrans- Delay window at TCP-DCA receiver mission timer is started after sending an ack, and Path length (h) Delay window limit canceled after receiving a data packet. h63 Congestion window Until now all previous results were obtained 3<h69 5 without ack retransmission. This is due to two facts: h P 10 3 first, the results with and without ack retransmission are almost the same for static networks. Mostly 4.2. Ack loss because the ack loss is rare, and the ack retransmis- sion timer period is large (at least 1 s). Second, with- The standard TCP generates more acks than out the ack retransmission timer, it is easy to TCP-DCA, due to the cumulative property of the observe the sole impact of the delay window on ack, one ack successfully received at the TCP sender TCP performance. In the following simulations, can makes up for all the preceding lost acks. Thus, the ack retransmission timer is enabled unless the ack loss in standard TCP is not serious. How- explicitly mentioned. ever, the number of acks in TCP-DCA is much less than standard TCP, and thus an ack loss has more 5. Performance evaluation negative impact on TCP-DCA performance. In order to be robust to ack loss, a retransmission This section presents the TCP-DCA performance timer at the receiver is adopted to retransmit a over wireless and hybrid networks. First, we show cumulative ack if necessary. TCP-DCA performance on static multihop wireless A challenge here is to estimate RTT at the recei- networks and compare it with TCP-DAA in [3]. Sec- ver which is needed for setting the ack retransmis- ond, we demonstrate TCP-DCA performance on sion timeouts. If TCP includes two-way data, the mobile ad hoc network, and show TCP-DCA per- receiver can get the accurate RTT measurement formance over several ad hoc routings. Third, we since the sender acks the data immediately. If TCP extend the hop count based approach to an end- only has one-way data, RTT measurement at the to-end delay based scheme to further improve receiver may not be straightforward. If the TCP TCP-DCA in hybrid wired/wireless networks. Last, timestamp option is used, the receiver could use it we propose a simple approach to address the friend- to estimate RTT, though such an estimate may be liness problem of TCP-DCA to TCP-Newreno with inflated if the sender does not send data packets standard delayed option. immediately after receiving an ack [17]. However, in TCP-DCA, the ack retransmission timer is only 5.1. Static ad hoc network a coarse timer for predicting when to retransmit acks, an accurate RTT measurement is not required In this section, we show TCP-DCA performance and thus an inflated RTT can be tolerated. In the in static ad hoc networks. Meanwhile, we compare paper, the timestamp option is used to estimate TCP-DCA with TCP-DAA and TCP-Newreno the RTT for a TCP flow which has one-way data. (with and without standard delayed ack option The receiver computes the ack retransmission [1]). More specially, we use the name ‘‘original time based on this RTT measurement. We use the TCP’’ for TCP-Newreno without standard delayed following formula to calculate the retransmission option, and TCP-STD-DA for TCP-Newreno with timer period: standard delayed option. AODV routing is used unless otherwise stated. 7 1 TCP-DAA [3] is an interesting extension of [2] SRTTr ¼ SRTTr þ RTTr ; kþ1 k kþ1 8 8 and has shown good performance in static ad hoc 3 1 networks. It is the most recent work and the major RTTvar ¼ RTTvar þ jRTTr À SRTTr j; kþ1 k kþ1 kþ1 4 4 differences between TCP-DAA and TCP-DCA are r r var RTOkþ1 ¼ SRTTkþ1 þ 4 Â RTTkþ1 ; listed in Table 2. For fair comparison, most simula- tions scenarios presented in this subsection were where RTTr is the RTT estimation at the receiver, shown in [3]. Each result is the average of five sim- SRTTr is the smoothed RTT. RTTvar and RTOr ulation runs. are RTT variance and retransmission timer period In Fig. 8, we compare TCP-DAA to TCP-DCA at the receiver. The minimum value of the retrans- in the chain topology with different hop count and
- 11. 1108 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 Table 2 Design difference TCP-DAA TCP-DCA Sender Duplicate acks threshold to trigger retransmission is 2; CW Puts CW into advertised window field upper limit is 4; Retransmission (RTO) timer is increased fivefold 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 TCP-DCA is CWL and TCP-DAA achieve best performance designed without cwnd limit, we include TCP-DCA and their difference is small. They provide better with cwnd limit at 4 in this experiment, named TCP- performance, about 10–25%, than the standard DCA-CWL, to show a fair comparison with TCP- TCP (with and without delayed ack option). DAA. Note that in TCP-DAA, the cwnd limit is With static topology and steady traffic pairs, the set to 4 by default. Fig. 8a shows the performance variance of results is very small (less than 5%) so we of TCP-DCA over 3 hop chain. The aggregate omit its error bar results. Also note that TCP-DCA- throughput of TCP-DCA decreases when the num- CWL does not bring significant benefit over TCP- ber of flows increases. When there are few flows, DCA. Since the ack flow is decreased, the optimal since each TCP-DCA generates as few acks as pos- cwnd size could actually increase (for details, please sible, the advantage of TCP-DCA is obvious. When refer to [14]). Although it may be interesting to more flows coexist, the cwnd of each flow becomes investigate the optimal cwnd for TCP-DCA, we smaller since they compete for the bandwidth, thus do not pursue it because our ultimate goal to design the benefit of delayed ack reduces because more TCP-DCA without cwnd limit. Moreover, this acks are generated. If the cwnd is below the cwnd property makes TCP-DCA distinct since selecting limit in TCP-DCA-CWL, they have similar perfor- an optimal cwnd is non-trivial, which needs to know mance. The performance of TCP-STD-DA is better a lot of underlying information to accomplish. For than original TCP, but worse than TCP-DCA and example, it requires the knowledge of transmis- TCP-DAA. TCP-DCA and TCP-DCA-CWL pro- sion/interference ratio [14] at the physical layer vides better performance than other TCP variants. which is usually hard to obtain. It could also poten- Overall, TCP-DCA achieves best performance, up tially limit the application of TCP in a variety of to 15% improvement over TCP-DAA, and 25% over wireless networks, such as that TCP with cwnd limit TCP-STD-DA respectively. Fig. 8b–c show the per- in wired and wireless scenario does not perform well formance of TCP-DCA on a 5 and 7 hop path. The (will be shown in Section 5.4). Taking this into same trend persists and TCP-DCA outperforms all account, we do not choose cwnd limit in TCP- other algorithms in most situations. DCA and thus in the following experiments, only Note that the cwnd limit on TCP-DCA does not TCP-DCA (no cwnd limit) is adopted. bring considerable benefit. For a short path in Fig. 8a, the cwnd limit could considerably restrict 5.2. Mobile ad hoc network the TCP performance as the number of flows is small. For a long path in Fig. 8b–c, the cwnd limit We have demonstrated that applying delayed ack only provides slightly more performance gain for 2 scheme improves TCP performance in a static net- flows, and this advantage disappears for more flows. work, now we show that it can improve TCP perfor- From Fig. 8, the performance of TCP-DCA without mance in a mobile network as well. Though our cwnd limit is better than that of TCP-DCA-CWL in delayed ack scheme is not specifically designed for most situations. It indicates that TCP-DCA can the TCP in mobile network, we show that our achieve the distinguished performance without set- scheme works well in mobile scenarios. ting cwnd limit. In Fig. 10, we show the performance of the origi- We also give results for the grid topology. This nal TCP, TCP-STD-DA, TCP-DAA and TCP- grid topology is shown in Fig. 9a and 6 flows run- DCA running over three routing schemes: GPSR, ning on the grid. Fig. 9b compares the performance AODV and DSR. The experiment consists of 40 of original TCP, TCP-STD-DA, TCP-DAA, TCP- nodes randomly moving within a 1000 m · 1000 m DCA-CWL and TCP-DCA, and we observe the area. Each node moves with the same speed without similar results as before. TCP-DCA, TCP-DCA- pause. The Random Waypoint mobility model [18]
- 12. J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1109 550 TCP-DCA Flow 4 Flow 5 Flow 6 TCP-DCA-CWL TCP-DAA TCP TCP-STD-DA Flow 1 Aggregate Throughput (Kbps) 500 Flow 2 450 400 2 4 6 8 10 12 14 16 18 20 Flow 3 Number of Flows (a) 3 Hop Path (a) Grid Topology 300 TCP-DCA TCP-DCA-CWL TCP-DAA 290 TCP TCP- TCP-STD-DA DCA Aggregate Throughput (Kbps) 280 TCP- 270 DCA- CWL 260 TCP- 250 DAA TCP- 240 STD- DA 230 TCP 220 2 4 6 8 10 12 14 16 18 20 Number of Flows 0 100 200 300 400 500 600 700 (b) 5 Hop Path Aggregate Throughput (Kbps) 280 (b) Aggregate Throughput for TCP variants over TCP-DCA TCP-DCA-CWL Grid Topology TCP-DAA TCP 260 TCP-STD-DA Fig. 9. Performance comparison on grid topology. (a) Grid topology. (b) Aggregate throughput for TCP variants over grid Aggregate Throughput (Kbps) 240 topology. 220 but with different randomly generated mobility traces. We run the set of experiments with 40 differ- 200 ent mobility traces and plot the 95% confidence interval as error bars on the figures. 180 We illustrate results on speed 10 m/s and 20 m/s in Fig. 10 with one TCP flow, for other speeds and 160 2 4 6 8 10 12 14 16 18 20 multiple flows, the results are similar. From Fig. 10 Number of Flows we observe that the delayed cumulative ack contrib- (c) 7 Hop Path utes to the performance gain compared with other Fig. 8. Aggregate throughput for TCP variants over chain TCP variants. TCP-DCA produces at least 10% per- topology. (a) 3 hop path. (b) 5 hop path. (c) 7 hop path. formance gain over other TCP variants regardless of routing protocols and mobility speed, especially more gain for TCP over DSR. TCP-STD-DA is is used. All results displayed are calculated as the unsurprisingly consistently better than the original average of 40 runs with the same traffic pattern, TCP without delayed ack. However, TCP-DAA
- 13. 1110 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 900 800 TCP-DCA TCP-DCA TCP TCP 800 TCP-STD-DA 700 TCP-STD-DA TCP-DAA TCP-DAA TCP-DCA-ACK-LOSS TCP-DCA-ACK-LOSS 700 600 Throughput (Kbps) Throughput (Kbps) 600 500 500 400 400 300 300 200 200 100 100 0 0 GPSR AODV DSR GPSR AODV DSR (a) 10 m/s (b) 20 m/s Fig. 10. TCP performance over mobile network. (a) 10 m/s. (b) 20 m/s. does not show consistent improvement across differ- acks are generated and more timeouts are likely to ent routing protocols. It performs similarly as TCP- happen. Fig. 11 illustrates the idle time of TCP dur- STD-DA over GPSR, but does not show better ing the whole simulation time (400 s) for different performance than TCP (with or without standard TCP variants and routing schemes. As expected, delayed option) over DSR. TCP-DCA (considering ack loss) has least idle time In TCP-DCA, the receiver would probe the sen- over each routing scheme, which helps TCP-DCA der by retransmitting acks in the event of high maintain better throughput in a mobile network. packet loss (see Section 4.2). It prevents the sender Note that TCP-DCA-ACK-LOSS has similar idle from entering long timeout due to the severe packet time with other TCP variants, which confirms that loss event. In Fig. 10, we also show the result of retransmission scheme is effective in improving TCP-DCA without ack retransmission (TCP- TCP performance in mobile networks. DCA-ACK-LOSS) to confirm the effectiveness of ack retransmission. When packet loss is severe 5.3. Lossy network (e.g. TCP over DSR), ack retransmission consider- ably boosts the TCP-DCA’s performance. On the In wireless networks, the wireless link in fact other hand, the ack retransmission does not help could become unreliable and error-prone. After much when packet loss is not severe (e.g. TCP over studying the performance of TCP-DCA under GPSR). Therefore, TCP-DCA only incurs more ack packet losses in mobile networks, we investigate its retransmission overhead when needed, and works performance under lossy wireless link. BER (bit well with packet losses in mobile networks. error rate) is used as the loss metrics at the wireless Comparing with static ad hoc networks in Sec- link and the error is assumed to be uniformly dis- tion 5.1, mobility poses more challenges for TCP. tributed and independent on each link. It is worth In a mobile network, the loss of packets (or acks) pointing out that when a packet is received with may be caused by temporary route loss as well as bit error, the IEEE 802.11 MAC layer helps in network congestion. Higher mobility causes more recovering packet loss by retransmissions until a route loss and thus more packet losses. Since routes maximum retransmission limit is reached [19] or are likely to be lost frequently in high node mobility retransmission is successful. Therefore, when BER environments, and different routing algorithms have is low, the MAC retransmission scheme can easily different capability to repair broken routes quickly, recover the packet and would not cause much prob- thus different TCP performance is observed over lem for TCP. While the BER is high, the heavy different routing schemes. Moreover, The frequent packet loss would knock down TCP performance. route breakages could put TCP into long idle state Fig. 12 illustrates a scenario where a single flow leading to poor TCP performance, especially for goes through 5 hop chain path under varying TCP with untraditional delayed ack where fewer BER from very low BER to high BER. Each bar
- 14. J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1111 350 TCP-DCA 350 TCP-DCA TCP TCP TCP-STD-DA TCP-STD-DA 300 TCP-DAA 300 TCP-DAA TCP-DCA-ACK-LOSS TCP-DCA-ACK-LOSS 250 250 Idle Time (second) Idle Time (second) 200 200 150 150 100 100 50 50 0 0 GPSR AODV DSR GPSR AODV DSR (a) 10 m/s (b) 20 m/s Fig. 11. TCP idle time. (a) 10 m/s. (b) 20 m/s. 300 TCP-DCA account the acknowledged data covered in the acks. TCP TCP-STD-DA Using delayed ack could potentially cause slow TCP-DAA 250 cwnd increase and thus hurt TCP performance. Therefore, the pure adaptive delay window based on the hop count appears not adequate for this sce- Throughput(Kbps) 200 nario. For example, in WLAN where 1 hop wireless 150 client can delay the cumulative ack for the whole cwnd, the cwnd at the sender could increase slowly, 100 particularly if a large propagation delay is encoun- tered in the wired part. Although slow increase for 50 cwnd is helpful for improving TCP performance in ad hoc networks as shown in our previous results, 0 it is not desirable for the hybrid network. In 2e-6 6e-6 2e-5 6e-5 Bit Error Rate RFC3465 [20] Allman proposed to increase the cwnd based on the number of bytes acknowledged Fig. 12. TCP performance under random loss. by the arriving acks rather than based on the num- ber of acknowledgments that arrive at the sender in is an averaged result of 20 random runs. Overall, the RFC2581 [1]. However, this approach potentially performance of TCP-DCA is no worse than other causes large bursts if the delay window is large. variants, and quite robust to wireless loss. TCP- An appropriate delay window limit needs to be DAA also works very well unless the loss is too investigated in this scenario. heavy. With the increasing loss rate, all TCP vari- Another challenge of the hybrid network is that ants gracefully reduce their performance and the the hop count information from the packet header variance of results increase. is no longer accurate for the wireless hops since it includes the wired path. It is possible for a wireless 5.4. Hybrid (wired and wireless) network node to get hop count from the routing table to the base station (or gateway) to the wired network, In this section we study TCP-DCA in the wired however, this information is not guaranteed to be and wireless ad hoc environment. Since wired net- available all the time. work has abundant bandwidth, it demands a much To address the problems in hybrid networks, we larger cwnd than the pure ad hoc network for effec- tailor and propose a novel receiver-side probe mech- tive TCP performance. In current TCP implementa- anism based on CapProbe in [21], and extend the tion, TCP sender increases the cwnd only based on previous hop-count based approach via the recei- the number of acks received while not taking into ver-side probe in hybrid networks.
- 15. 1112 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 The receiver-side probe works as follows: First Table 3 the receiver passively estimate the wireless band- IEEE 802.11b MAC overhead (with long preamble), L is packet size, R is capacity width and minimum RTT. Second it computes the equivalent hop count based on the minimum RTT Data frame 192 ls + 8L/R and wireless bandwidth, then the previous hop SIFS 10 ls count approach can be applied afterwards. DIFS 50 ls RTS 192 ls + 160/R Now we address the problem to estimate the end- CTS 192 ls + 112/R to-end hybrid path bandwidth at the TCP receiver. ACK 192 ls + 112/R CapProbe [21] is a recently proposed bottleneck capacity and path rate estimation technique shown to be both fast and accurate. It combines the use Table 4 of time interval measurements and end-to-end delay Delay window at TCP-DCA receiver (delay-based, wireless measurements to filter out bad packet pair probing bandwidth is at 2 Mbps and TCP data packet size is 1000 bytes) samples. It is proved to be suitable for wired and Minimum RTT (ms) Delay window limit last-hop wireless networks. Inspired by CapProbe, RTTmin < 18 Congestion window we proposed a new approach at the TCP-DCA 18 6 RTTmin 6 60 5 receiver without triggering any probing traffic. The RTTmin > 60 3 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 Internet BS 1 2 packet pair arrives. Our approach is based on the CapProbe, but it is different with CapProbe. First Fig. 13. Hybrid wired/wireless topology. our approach is imbed at the TCP 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 1500 TCP-DCA TCP to work for wired, hybrid and purely wireless TCP-STD-DA TCP-DAA networks. The other challenge is to estimate the minimum Throughput (Kbps) 1000 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 TCP 500 data and ack packet sizes. For example, with wireless bandwidth of 2 Mbps, data packet size of 1000 bytes and ack size of 40 bytes, the minimum RTT for a 1 wireless hop path is about 6 ms considering various 0 overhead, such as header overhead at TCP (20 bytes), 1 hop 2 hop IP (20 bytes) and MAC (34 bytes) and MAC layer Wireless Hop Count fixed overhead shown in Table 3 (for details, please Fig. 14. Wired and wireless performance. 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 TCP connection from a running with different random seeds. Similar to Sec- wired network to a wireless network with up to 2 tion 5.1, the variance of results is very small so we wireless hops. The one-way propagation delay on omit its error bar results. Since TCP-DAA limits the wired link is 50 ms. The performance of TCP- the cwnd, the performance of TCP-DAA is limited. DCA, TCP-DAA and the standard TCP (with and TCP-DCA provides best performance in both cases without standard delayed ack option) shown in of 1 and 2 hop paths. Therefore, TCP-DCA is Fig. 14, represents averaged result of 10 simulation extensible to hybrid networks instead of working
- 16. J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1113 well only in an ad hoc network. It is an attractive 1400 TCP-DAA point of TCP-DCA as it becomes common for wire- TCP-STD-DA 1200 less networks to communicate with the outer world. For example, in the real life, one needs to connect a 1000 mobile device (e.g. laptop) to the Internet to down- Throughput (Kbps) load files from a remote server. 800 600 5.5. Friendliness 400 An important issue about TCP-DCA is its friend- liness with TCP-STD-DA since TCP-STD-DA is 200 currently more popular in all networks. In TCP- DCA, cwnd increases slower than TCP-STD-DA, 0 1 hop 2 hop 5 hop so TCP-DCA is not so aggressive as TCP-STD- (a) TCP-DAA DA. The simulation results verify this point in Fig. 15b by running one TCP-STD-DA and one 1400 TCP-DCA TCP-DCA over chain topologies with varying hop TCP-STD-DA count. For each experiment, the average result of 1200 10 simulation runs is shown. We also omit the var- 1000 iance of result since it is very small. Throughput (Kbps) For short paths, TCP-STD-DA gets more 800 throughput than TCP-DCA. However, when path is long, TCP-DCA becomes friendly to TCP-STD- 600 DA since the benefit from delaying acks in 400 TCP-DCA is not significant. Note that although TCP-DCA is submissive to TCP-STD-DA, its 200 friendliness to TCP-STD-DA is better than TCP- DAA as shown in Fig. 15a. Since TCP-DAA not 0 1 hop 2 hop 5 hop only increases cwnd slower but also limits the cwnd (b) TCP-DCA to a small value, its friendliness to TCP-STD-DA is the worst in all cases. One sidenote is that if the ori- 1400 TCP-DCA-SWITCH ginal TCP coexists with TCP-DCA or TCP-DAA, TCP-STD-DA the unfriendliness problem would be more severe 1200 since the original TCP is more aggressive than 1000 TCP-STD-DA. Throughput (Kbps) Fig. 15b shows that TCP-DCA is not always 800 friendly to TCP-STD-DA, especially for short paths. However in reality, the friendliness on short 600 paths are more important because the wireless path 400 is usually not long, e.g., WLAN with one hop wire- less path. Thus how to make TCP-DCA friendly to 200 the possible TCP-STD-DA flows under different scenarios needs to be addressed. In the paper we 0 1 hop 2 hop 5 hop propose an intelligent switch scheme to allow (c) TCP-DCA-SWITCH TCP-DCA to match TCP-STD-DA performance if they coexist. The approach is to let TCP-DCA recei- Fig. 15. Friendliness with TCP-Newreno. (a) TCP-DAA. (b) ver detect if it is the only traffic in the network. If TCP-DCA. (c) TCP-DCA-SWITCH. there is other traffic, TCP-DCA switches to TCP- STD-DA; otherwise if it is the only traffic in the net- work, it switches back to TCP-DCA. The approach end approach consistent with our design goal is imbed in the receiver of TCP-DCA, without for TCP-DCA, and it works on TCP receiver with- incurring any extra overhead. This simple end-to- out standard delayed option as well since our
- 17. 1114 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 2 TCP-STD-DA DCA is disabled TCP-DCA-SWITCH 1.8 end if else if TCP throughput > b · Path Capacity then Instantaneous Throughput (Mbps) 1.6 if no DCA 1.4 DCA is enabled 1.2 end if 1 end if 0.8 0.6 0.4 6. Future work and discussion 0.2 In [23], ack congestion control is introduced to 0 TCP to combat bandwidth asymmetry on for- 50 100 150 200 250 Time (second) ward/backward path. Ack packet should be reduced since backward path has limited bandwidth. How- Fig. 16. Responsiveness of TCP-DCA-SWITCH. ever, the ack congestion control is implemented at the gateway with a RED queue. It is not an end- to-end approach since the receiver needs to get adaptive delayed ack scheme is triggered only when ECN signal from the gateway. In contrast, our necessary. scheme is end-to-end, and does not require any To detect the possible other flows, the receiver changes at intermediate nodes. monitors the TCP data throughput via receiver-side Delayed ack inevitably triggers burst transporta- probe mechanism described above. If the through- tion at the sender. The burstiness increases the put is much less than the path capacity, then there packet loss and potentially hurts TCP performance. probably has other traffic on the path. We present Since TCP-DCA does not use large delay window our switch algorithm in Algorithm 1. a ¼ 0:4; b ¼ except for short paths, the burstiness is limited. It 0:6 in the simulation. is possible to further reduce the burstiness. For Fig. 15c shows the performance of TCP-DCA example, in [24], cwnd increase is limited by at most with switch (TCP-DCA-SWITCH). Obviously, the 2 packets for each delayed ack, or ack reconstruc- friendliness of TCP-DCA to TCP-STD-DA is tion in [23] to rate control the sender. How to much improved compared with original TCP- further decrease the burstiness would be an interest- DCA, especially for short paths. The TCP-DCA- ing future work. SWITCH can share the path with TCP-STD-DA Another issue is the effect of delayed ack for quite fairly. short TCP flows. As we discussed before, using To illustrate the transient behavior of TCP- delayed ack could slow down the cwnd growth, DCA-SWITCH coexisting with TCP-STD-DA, thus delaying ack has potential negative impact Fig. 16 plots the typical instantaneous throughput on a short TCP flow, as illustrated in [3]. We also in a simulation run. For this simulation set, the first found similar results, and we are investigating on TCP-DCA-SWITCH flow runs from the beginning improving short flow performance with delayed until 250 s and the second TCP-STD-DA flow starts ack. One simple approach is to disable delaying from 100 s to 200 s. Fig. 16 shows that TCP-DCA ack when the cwnd is small, for example, the recei- responds quickly to changing network conditions ver would not delay acking if the cwnd is less than when new flow starts at 100 s as well as when 4. It is also possible use the cwnd size to detect flow stops at 200 s. Using TCP-DCA-SWITCH, a slow start phase or let sender notify the receiver TCP flow utilizes the available bandwidth when if it is in slow start phase. If the sender is detected there are no competing flows and shares the band- to be in slow start, the receiver temporarily disables width fairly when multiple flows compete for the the delayed ack scheme to make TCP achieve good bandwidth. throughput at connection startup. After TCP leaves Algorithm 1. Switch Algorithm the slow start, TCP receiver begins delaying ack to better utilize the shared wireless medium. These if TCP throughput < a · Path Capacity then enhancements are expected to be beneficial for if DCA is enabled then short flows because most of short flows finish their
- 18. J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 1115 transfer while they are still in the slow start phase [5] C.E. Perkins, E.M. Royer, Ad-hoc on-demand distance [25]. vector routing, in: Proceedings of IEEE WMCSA’99, Feb- ruary 1999. [6] D.B. Johnson, D.A. Maltz, in: in: T. Imielinski, H. Korth (Eds.), Dynamic Source Routing in Ad hoc Wireless 7. Conclusion Networks, Mobile Computing (chapter 5). [7] B. Karp, H.T. Kung, GPSR: greedy perimeter stateless TCP, a dominant reliable transport protocol in routing for wireless networks, in: Proceedings of ACM wired networks, is a highly competitive candidate MobiCom’00, August 2000. for providing reliable data transport in wireless [8] Z. Fu, P. Zerfos, H. Luo, S. Lu, L. Zhang, M. Gerla, The impact of multihop wireless channel on TCP throughput and and wired/wireless networks. This paper systemati- loss, in: IEEE INFOCOM’03, March 2003. cally examines the relationship of TCP performance [9] K. Chandran, S. Raghunathan, S. Venkatesan, R. Prakash, to delayed ack via analysis, and simulation and test- A feedback-based scheme for improving TCP performance bed experiments. We find that TCP does not always in ad hoc wireless networks, IEEE Personal Communica- get throughput gain by delaying unlimited acks. The tions Magazine 8 (1) (2001). [10] G. Holland, N.H. Vaidya, Analysis of TCP performance maximal TCP throughput is achieved at a certain over mobile ad hoc networks, in: Proceedings of ACM delay window that balances decreasing ack flow MobiCom’99, August 1999. and burst loss. Thus, we introduced two novel adap- [11] K. Xu, M. Gerla, L. Qi, Y. Shu, Enhancing TCP fairness in tive delayed ack mechanisms compatible with TCP ad hoc wireless networks using neighborhood red, in: called TCP-DCA, based on the path hop length Mobicom, 2003. [12] A.K. Singh, K. Kankipati, Tcp-ADA: Tcp with adaptive and/or end-to-end delay, to maximize the TCP per- delayed acknowledgement for mobile ad hoc networks, in: formance for MANET and hybrid networks. Our WCNC, 2004. proposed schemes are shown to have superior per- [13] K. Nahm, A. Helmy, C.J. Kuo, Tcp over multihop 802.11 formance in static networks (wireless or hybrid networks: issues and performance enhancement, in: Mobi- wired/wireless networks). We also demonstrated hoc, 2005. [14] K. Chen, Y. Xue, K. Nahrstedt, On setting TCP’s congestion better TCP performance with our proposed mecha- window limit in mobile ad hoc networks, in: ICC, 2003. nisms over different routing protocols in mobile ad [15] T.D. Dyer, R.V. Boppana, A comparison of TCP perfor- hoc networks. mance over three routing protocols for mobile ad hoc It is well known that TCP with untraditional networks, in: Proceedings of ACM MobiHoc’01, October delayed ack is less aggressive than TCP-Newreno 2001. [16] The network simulator - ns2, Available from: http:// since the cwnd with untraditional delayed ack grows www.isi.edu/nsnam/ns/. much slower. To combat such unfriendliness, we [17] V. Jacobson, Tcp extensions for high performance, RFC propose a receiver-side probe and detection to 1323, 1992. maximally increase the TCP-DCA friendliness [18] C. Bettstetter, H. Hartenstein, X. Prez-Costa, Stochastic (when other flows exist) and efficiency (when no properties of the random waypoint mobility model, ACM/ Kluwer Wireless Networks: Special Issue on Modeling and other flows exist). The detection is imbed in the Analysis of Mobile Networks, September 2004. TCP-DCA and does not need to modify [19] IEEE-802.11, Wireless LAN media access control (MAC) underlying layers. TCP-DCA friendliness to TCP- and physical layer (PHY) specifications, Available from: Newreno is significantly enhanced with this simple http://standards.ieee.org/getieee802, 1999. approach. [20] M. Allman, Tcp congestion control with appropriate byte counting (abc), RFC 3465, 2003. [21] L.L.M.G. Rohit Kapoor, Ling-Jyh Chen, M.Y. Sanadidi, Capprobe: a simple and accurate capacity estimation tech- References nique, in: ACM SIGCOMM, 2004. [22] M. Kappes, An experimental performance analysis of mac [1] M. Altman, Tcp congestion control, RFC 2581 (1999). multicast in 802.11b networks for voip traffic, in: SPECTS, [2] E. Altman, T. Jimenez, Novel delayed ack techniques for 2004. improving TCP performance in multihop wireless networks, [23] H. Balakrishnan, V.N. Padmanabhan, R.H. Katz, The Personal Wireless Communications, September 2003. effects of asymmetry on TCP performance, in: Mobicom, [3] R. de Oliveira, T. Braun, A dynamic adaptive acknowledg- 1997. ment strategy for TCP over multihop wireless networks, in: [24] M. Allman, On the generation and use of TCP acknowledg- Infocom, 2005. ments, ACM Computer Communication Review, 1998. [4] T.D. Dyer, R.V. Boppana, A comparison of TCP perfor- [25] H. Balakrishnan, S. Seshan, M. Stemm, R. Katz, Tcp mance over three routing protocols for mobile ad hoc behavior of a busy internet server: Analysis and improve- networks, in: Mobihoc, 2001. ments, in: Infocom, 1998.
- 19. 1116 J. Chen et al. / Ad Hoc Networks 6 (2008) 1098–1116 Jiwei Chen received the B.S. degree in works). He has been working on the DARPA, NSF-WHYNET, Applied Mathematics and M.S. degree in TSAT projects and most recently the ONR MINUTEMAN and Electrical Engineering from Nanjing ARO-MURI DAWN projects including the network protocols University of Science and Technology, design and implementation for highly mobile, large scale ad-hoc China. Currently, he is a Ph.D. candi- networks. He has also carried out design and performance eval- date in Electrical Engineering of Uni- uation of QoS routing, multicasting protocols, TCP transport versity of California, Los Angeles, where and network security for wireless networks. (Yeng-Zhong’s his research topics include protocol homepage: www.cs.ucla.edu/~yenglee). design, modeling and performance eval- uation of transport, routing and MAC M. Yahya ‘‘Medy’’ Sanadidi was born in layers in ad hoc and heterogeneous Egypt where he received his high school networks. diploma from College Saint Marc D’Alexandrie, and his B.Sc. from the Mario Gerla received a graduate degree University of Alexandria. He earned his in engineering from the Politecnico di M.Sc. from Penn State, and his Ph.D. in Milano in 1966, and the M.S. and Ph.D. Computer Science from UCLA. He is degrees in engineering from UCLA in currently an Adjunct Professor at the 1970 and 1973, respectively. After UCLA Computer Science Department, working for Network Analysis Corpo- where he is co-Principal Investigator on ration from 1973 to 1976, he joined the NSF sponsored research in high perfor- Faculty of the Computer Science mance Internet protocols. At UCLA, he also teaches under- Department at UCLA where he is now graduate and graduate courses on computer networks, queuing Professor. His research interests cover systems, and probability modeling and analysis. Dr. Sanadidi was the performance evaluation, design and a Manager and Senior Consulting Engineer at AT&T/Teradata, control of distributed computer communication systems; high and previous to that, he held the position of Computer Scientist speed computer networks; wireless LANs, and; ad hoc wireless at Citicorp, and Assistant Professor at the Computer Science networks. He has worked on the design, implementation and Department, University of Maryland, College Park, Madisson. testing of various wireless ad hoc network protocols (channel Professor Sanadidi, who is a Senior Member of the IEEE and access, clustering, routing and transport) within the DARPA Member of the ACM, has co-authored over fourty conference WAMIS, GloMo projects. Currently he is leading the ONR and journal papers and has been awarded two best paper awards. MINUTEMAN project at UCLA, and is designing a robust, He has served as reviewer of journal publications, member of scalable wireless ad hoc network architecture for unmanned technical program committees and organizing committees, key- intelligent agents in defense and homeland security scenarios. He note speaker, as well as chairman of professional conferences. Dr. is also conducting research on QoS routing, multicasting proto- Sanadidi has been awarded two patents, and has consulted for a cols and TCP transport for the Next Generation Internet (see number of industrial concerns. His current research interests are www.cs.ucla.edu/NRL for recent publications). He became IEEE in path characteristics estimation and its applications in conges- Fellow in 2002. tion control, adaptive multimedia streaming, and hybrid wire/ wireless networks. Yeng Zhong Lee received the M.S. Degree in Computer Science from Uni- versity of California, Los Angeles in 2001. He was working for HRL Labo- ratories, LLC (Formerly known as Hughes Research Laboratories) and Telcordia Technologies (Formerly known as Bell Communications Research) in 2002 and 2004, respectively. He is currently a Ph.D. candidate in Computer Science at University of Cali- fornia, Los Angeles. His research interests cover mobile ad hoc networks, mesh networks and wireless LANs (Bluetooth net-