Performance Evaluation for TCP in Tactical Mobile Ad Hoc Networks
Jonas Karlsson
1
, Velizar G. Dimitrov
2
, Andreas Kassler
1
, Anna Brunstrom
1
, Jan Nilsson
3
and Anders Hansson
3
1
Department of Computer Science, Karlstad University, Universitetsgatan 2, Karlstad, Sweden
2
Department of Communication Networks, Technical University of Sofia, Boulevard Kliment Ohridski 8, Sofia, Bulgaria
3
Swedish Defence Research Agency (FOI), Box 1165, SE - 581 11 Link¨oping, Sweden
Keywords:
MANET, TCP, Tactical Networks, TDMA.
Abstract:
Tactical networks are used in military and rescue operations to provide timely and accurate information to
operating teams. Tactical networks have traditionally used long distance narrow band radio links. However,
although these links provide robust real-time communication the limited bandwidth makes them less suited
for high data-rate applications. To support high-data rate TCP applications such as providing digital images
and maps, emerging tactical networks use shorter range but higher data-rate wide band radio links and multi-
hop. Due to the requirement of cheap up-front cost, most MANET research has focused on Carrier Sense
Multiple Access (CSMA) networks. However, in tactical networks, where bounded delays are important,
Time Division Multiple Access (TDMA) can give better possibility to support the Quality of Service needed
for real-time communication. The purpose of this paper is to assess and compare the throughput of three
state-of-the-art TCP versions and two routing protocols over TDMA based MANETs.
1 INTRODUCTION
During a disaster or a military campaign the regular
wired infrastructure is at best semi-functional. There-
fore, relying on an infrastructure-based system is in
most cases not possible. Prime examples of this are
disaster areas, caused by e.g. earthquakes, tsunamis
and nuclear disasters, which can render an infrastruc-
ture unusable. Tactical networks are also relatively
small networks where nodes follow one or several
group leaders (Li et al., 2012). A tactical network
should also be resilient to node failures and operate
with little or no backbone infrastructure.
One network type that promises to do this is Mo-
bile Ad Hoc Networks (MANETs). A tactical net-
work can be seen as a multi-hop MANET where node
mobility and traffic follows a special pattern. In a
MANET, a node functions both as a host and as a
router and nodes can automatically organize them-
selves and form a network. However, due to limited
transmission range, intermediate nodes need to for-
ward packets to form a connected network topology.
This creates congestion and multiple points of failure.
Node mobility also leads to a highly dynamic network
topology, which is prone to frequent changes and er-
rors. This network dynamicity poses several challen-
ges on routing protocols. Two state-of-the art rout-
ing protocols for MANETs are AODV and OLSR.
AODV is a reactive routing protocol that determines
routes only when needed (Perkins et al., 2003). OLSR
is a table driven proactive protocol that regularly
exchanges topology information with its neighbors
(Clausen and Jacquet, 2009).
Although voice, positioning and short messages
are the main applications for tactical networks it is
also important to support standard applications used
on todays Internet, to e.g. provide maps or other con-
tent. The obvious solution to support standard appli-
cations would be to use standard TCP/IP. However,
TCP was designed for data delivery over wired net-
works. In a MANET, which has a substantially higher
packet loss rate and jitter compared to a wired net-
work, the performance of TCP dramatically degrades.
Furthermore, due to node mobility, routes break and
merge more frequently. This leads to a higher degree
of routing induced losses and reordered packets, com-
pared with a wired network.
MANET research has mainly focused on TCP in-
teractions with CSMA MAC based systems due to the
availability of cheap IEEE 802.11 radio cards, and the
requirement of low up front cost. However, in a tac-
tical network using a TDMA scheme has several ad-
277
Karlsson J., G. Dimitrov V., Kassler A., Brunstrom A., Nilsson J. and Hansson A..
Performance Evaluation for TCP in Tactical Mobile Ad Hoc Networks.
DOI: 10.5220/0004067502770282
In Proceedings of the International Conference on Signal Processing and Multimedia Applications and Wireless Information Networks and Systems
(WINSYS-2012), pages 277-282
ISBN: 978-989-8565-25-9
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
vantages, e.g. support of bounded delays and a stable
network under heavy traffic loads. The novelty of this
paper includes a TDMA based MANET evaluation
of three well established TCP versions: TCP ELFN
(Holland and Vaidya, 2002), TCP New Reno (Floyd
et al., 2004) and TCP Westwood+ (Grieco and Mas-
colo, 2002)(Grieco and Mascolo, 2004), using both
reactive and proactive routing protocols and multiple
mobility scenarios and node densities.
The rest of the paper is structured as follows. Sec-
tion 2 describes the background and motivation for
this paper. Section 3 describes the implementation
and the motivation for the chosen topologies. Finally
Sections 4 and 5, respectively describe the experiment
results, our conclusions and future work.
2 BACKGROUND
MANETs are wireless networks that do not require
any infrastructure for their operation. Intermediate
nodes must therefore participate in the route discov-
ery and packet forwarding to other nodes. As nodes
are mobile, the network topology and link state is con-
stantly changing (Macker et al., 1999). This makes
routing a challenging task and routing protocols need
to quickly respond to topology changes.
Routing protocols in MANETs are commonly di-
vided into two categories based on how and when
routes are discovered and maintained: reactive and
proactive routing protocols. Reactive or on-demand
routing protocols establish the route to a destination
only when it is needed (Abolhasan et al., 2004). This
ensures a low routing overhead when there is a low
amount of source-destination pairs. This can lead to
an increase of routing overhead in high traffic condi-
tions. A proactive routing protocol on the other hand
maintains a consistent and updated view of the net-
work by periodically propagating route updates in the
network (Clausen and Jacquet, 2009). This is costly
when there is limited amount of traffic (Abolhasan
et al., 2004). The routing overhead of a proactive ap-
proach should in an ideal case not be influenced by the
traffic volume, i.e. the overhead should be low when
the traffic volume is large. However, studies in CSMA
networks have shown that the neighbor sensing mech-
anism of proactive routing protocols can be sensitive
to both traffic and mobility, due to collisions and ra-
dio environment (Voorhaen and Blondia, 2006). De-
pending on if information is spread regularly or trig-
gered by a link loss, with high mobility, table driven
proactive routing protocols can suffer from old route
information or high overhead. In this paper we have
chosen two state-of-art routing protocols to represent
these two categories, reactive AODV and proactive
OLSR.
Another important consideration in MANETs is
the Media Access Control (MAC) layer. A TDMA
channel access method allows several stations to share
the same medium by dividing the time into small time
slots. While mobility and radio environment are the
same for TDMA as for CSMA based networks, there
are significant differences where packets are dropped
in the two channel access methods. In static TDMA
based networks there should be no collisions; neigh-
bor sensing for proactive protocols should therefore
be more stable than in CSMA networks. Therefore
in a static network, no packets are dropped due to
collisions. However with node mobility, the TDMA
slot assignment could temporally be incorrect creat-
ing collisions and packet loss.
By design TDMA have a number of other advan-
tages over contention-basedapproaches, such as: fair-
ness, bounded delays and asymptotic behavior under
heavy traffic loads. The main drawback of TDMA
scheduling is that it requires clock synchronization
between the nodes which increases the overhead and
makes nodes more complex and expensiveto build. In
addition in multi-hop environments, an interference
free slot assignment becomes computational expen-
sive.
The critical component of a TDMA protocol in
MANETs is how to assign different time slots to any
two conflicting nodes in a distributed way. One solu-
tion that we evaluate in this paper is DRAND (Rhee
et al., 2006). DRAND was introduced to minimize
collisions and promote bounded delay for the purpose
of real-time communication (e.g. voice communica-
tion). One of the benefits of DRAND is that it can
re-compute the TDMA schedule without involving
global changes, i.e. DRAND performance is scalable
to partial topology changes.
Because of its inheritably probability based na-
ture, DRAND requires time to adapt with topology
changes and is expected to perform worse when the
number of topology changes increases. The imple-
mentation of DRAND that we used is the imple-
mentation done by the Computer Science department
of the North Carolina State University, USA (Net-
working Research Lab, 2012). The choice of using
DRAND is not for optimality to the evaluated scenar-
ios. The choice should rather be seen as one viable
way of implementing a TDMA scheme that works
both for stationary and mobile networks.
In the scenarios targeted in this paper, communi-
cations are often among groups which tend to coordi-
nate their movements, i.e. a rescue team. We have
therefore focused on using Reference Point Group
WINSYS2012-InternationalConferenceonWirelessInformationNetworksandSystems
278
Mobility Model (RPGM) that try to capture the move-
ments when nodes are influenced by a group leader
(Camp et al., 2002)(Aschenbruck et al., 2008).
In a MANET there will due to the node mobility
always be a certain amount of route breaks. This leads
to packet loss or jitter, as re-routing and link layer re-
transmission causes delay variations. The main prob-
lem for TCP in MANETs is its inability to distinguish
congestion loss from other losses and that the proper
response can be orthogonal. A congestion loss re-
quires the sender to reduce the sending rate to not
overwhelm the network. However, a loss due to a
lossy wireless channel, on the other hand, requires
the sender to quickly retransmit without necessarily
reducing the congestion window.
A great amount of research has been invested in
dealing with TCP performance issues in MANETs,
e.g. (Wang and Zhang, 2002) (Chen et al., 2003).
Most proposals are based on the idea of changing the
functionality and/or the behavior of TCP to adapt it
to the network environment. In this paper we instead
focus on three TCP variants that represent three differ-
ent design choices: TCP New Reno, TCP Westwood+
and TCP ELFN. While TCP New Reno (Floyd et al.,
2004) is one of the most common and well established
TCP variants, TCP Westwood+ and TCP ELFN have
been designed for specific environments.
TCP Westwood+ (Grieco and Mascolo, 2002) is
a sender-side-only modification of TCP New Reno
that is intended to better handle large bandwidth de-
lay product paths with potential packet loss due to
transmission or other errors. This property makes
TCP Westwood+ very attractive to use in wireless sys-
tems. TCP Westwood+ relies on monitoring the ACK
stream for information to help set the congestion con-
trol parameters, i.e. slow start threshold (ssthresh)
and congestion window (cwnd), better. Due to mo-
bility, the ACK stream will have larger fluctuations
in a MANET than in a static network. These none
congestion related fluctuations can reduce TCP West-
wood+’s ability to correctly determine the available
throughput.
TCP ELFN (Explicit Link Failure Notification)
(Holland and Vaidya, 2002) is a cross layer approach
to inform the TCP layer about route failures. The pro-
posal uses an ELFN message, which is transported by
or piggy backed on routing messages to the sender
upon a route break. The ELFN message contains
the sender, receiver addresses and ports, as well as
the TCP packet sequence number. On receiving the
ELFN message, the source responds by disabling
its retransmission timers and enters a ”frozen state.
During the ”frozen” period, the TCP sender probes
the network to check if the route is restored. When
the route is restored, i.e. the sender starts to see ac-
knowledgments of the probe packets; the TCP sender
leaves the ”frozen state and resumes its state as be-
fore the freeze event. The probe interval is a crucial
parameter as it determines how quickly TCP ELFN
will detect that a new route is established. However, a
too small probing interval will introduce unnecessary
overhead. In (Holland and Vaidya, 2002) the authors
propose to use a probe interval of 2 sec. Upon route
restoration, the proposals use the values of RTO and
cwnd from prior to the route failure. In the original
paper the authors used DSR (Johnson et al., 2007),
which is a reactive source routing protocol. In this
paper we are evaluating TCP ELFN with both proac-
tive and reactive routing protocols.
The authors of (Liao et al., 2002) propose a
novel reactive QoS routing protocol for TDMA based
MANETs. With the proposed protocol, the authors
show that it is possible to search for and establish
routes in a MANET supporting a given bandwidth
constraint. The bandwidth requirement is realized by
letting the route reply reserve time slots as it traverses
links on the reverse path. Our work differs in that we
evaluate TCP and different routing protocols over a
TDMA based MANET.
3 IMPLEMENTATION
The original ELFN was designed and tested with
DSR (Holland and Vaidya, 2002). In short, when-
ever DSR detects a route break, it generates a route er-
ror message which traverses back to the traffic source.
With the ELFN modification, these messages are in-
tercepted and processed at the TCP layer. For the pur-
pose of the current study and to be independent of the
underlying routing protocol we have implemented the
routing error message by using an additional ICMP
message. The main advantage of using an ICMP mes-
sage as opposed to piggy backing the information on a
routing message is that the routing protocol is decou-
pled from the TCP layer. This makes it possible to
implement a TCP layer that can operate with different
routing protocols in the same framework, e.g. IEEE
802.11s (Hiertz et al., 2010). The main drawback of
our proposal is the extra overhead of the ICMP mes-
sages (in the case of AODV). As with all approaches
that do an indirect coupling between protocol layers
there is a possibility for race conditions. In our case
this race condition is between the routing layer de-
tecting a path loss and a TCP timeout. That is, if TCP
times out due to a path loss before the routing layer
have detected and transmitted the ICMP message to
the sender, TCP ELFN might freeze in an undesirable
PerformanceEvaluationforTCPinTacticalMobileAdHocNetworks
279
state.
AODV was modified in a similar way as in (Ro-
manowicz, 2008) to work together with TCP ELFN.
When a link loss (route break) is detected AODV gen-
erates a route error message. In our modification, the
route error message is preceded by an ICMP message
which is sent back to the traffic source. Once the
TCP sender recives the ICMP message it will freeze
its timers and start the probing phase as described in
(Holland and Vaidya, 2002). When AODV has estab-
lished a new route and the TCP source have received
an acknowledgementfrom one of the probing packets,
it restores the timers. A link break is detected when
three or more HELLO messages are lost.
When using the pro-active OLSR, the ELFN op-
eration becomes much more peculiar. With OLSR,
a node is notified about any topology change (route
break) regardless of whether it is participating in a
data forwarding process or not. The implementation
has been changed to reflect this by removing the ex-
plicit ICMP message from the node that detects the
route break. Instead, whenever the OLSR agent at the
traffic source tries to forward packets to a destination
that it does not have a route for, it sends an ELFN
message to itself. In the simulations we set the OLSR
neighbor link discovery (HELLO) interval to 1 sec-
ond and the topology control (TC) message interval
to 5 seconds.
The ELFN probe packet interval was set to two
seconds for both routing protocols. We further dis-
abled link layer feedback for both ADOV and OLSR
as initial experiments showed that link layer feedback
caused performance degradation. Preliminary analy-
sis shows that in certain situations the links became
unstable, triggering unnecessary route error messages
and route flapping. This is, however, outside of the
scope of the current paper and has been left for future
work.
The topology we used was a square of 1000 me-
ters by 1000 meters without any obstacles or hetero-
geneity. Radio communication distance was 150 me-
ters and carrier sensing distance was 300 meters. The
physical layer speed was set to 1 Mbps. The reason
for selecting 1 Mbps is that it can be considered as a
challenging data rate when mobile tactical networks
need to support TCP services. Moreover, today most
mobile tactical radios rarely support more than a few
Mbps. For the mobility model we used the RPGM
mobility model with two node densities, consisting of
3 and 16 groups of 6 nodes thus forming scenarios
with 18 and 96 nodes. Group radius was set to 250
meters to ensure that internal group communication
was forwarded by at least one intermediate node. The
node speed was random in the range of 1,5 - 5,0 m/s.
The simulation time was 360 seconds and traffic flows
started at 60 sec, giving time for the routing protocols
to converge. There where two independent TCP flows
simulating large file transfers, between two different
node pairs in all scenarios. Each experiment was
performed both with and without background traffic.
When using background traffic, 10% of the nodes in
a scenario where each generating a 20 Kbit/s UDP
flow. The total background traffic volume varied be-
tween 20 Kbit/sec (18 nodes scenario) and 180 Kbit/s
(96 nodes scenario). This setup will produce a low
amount of background traffic when the node density
is low and a high amount of background traffic when
the node density is high. However, when the node
density is high there is more possibility for concur-
rent non competing traffic.
4 RESULTS AND ANALYSIS
In this section we will describe the results from our
experiments. The results show the aggregate aver-
age TCP throughput measured in kilobits per seconds
(Kbit/s). To compare TCP performance, we used the
ns 2.26 simulator (McCanne et al., 2012) with TCP-
New Reno, -ELFN and -Westwood+. The routing
protocols are slightly modified versions of OLSR and
AODV as described in the previous section.
The results shown in Figure 1 are from a sce-
nario with 18 nodes in 3 groups and with no back-
ground traffic. The network density is low which
causes temporal route breaks and network divisions
between nodes and groups. However, due to the
low node density, the impact of the choice of rout-
ing layer is limited. The low congestion in this sce-
nario also gives the possibility to achieve a high TCP
throughput. However, both TCP New Reno and TCP
Westwood+ cannot utilize this possibility as the route
breaks limit the congestion window. Therefore both
of the approaches have similar performance. TCP
ELFN on the other hand freezes its congestion win-
dow during the route breaks and can therefore main-
tain a high congestion window.
Figure 2 refers to a high node density scenario
with 96 nodes divided in 16 groups. In this sce-
nario the connectivity is high, but the mobility com-
bined with the amount of nodes increases the stress
on both routing layer and TDMA scheduling. Due
to this the overall throughput is less. This leads to a
smaller average congestion window for all TCP ver-
sions and consequently TCP ELFN has less benefits
of freezing the congestion window. What also can
be seen is that in this scenario the routing layer starts
to impact the results. Due to the increased amount
WINSYS2012-InternationalConferenceonWirelessInformationNetworksandSystems
280
Ϭ
ϭϬϬ
ϮϬϬ
ϯϬϬ
ϰϬϬ
ϱϬϬ
ϲϬϬ
ϳϬϬ
Ks K>^Z
<ďŝƚͬƐ
ϯ'ƌŽƵƉƐϭϴEŽĚĞƐ
EĞǁZĞŶŽ
ƐƚǁŽŽĚн
>&E
Figure 1: Low node density with no background traffic.
of topology changes, a proactive protocol as OLSR
experiences a higher amount of stale routes and the
possible formation of routing loops. In this scenario
the amount of TC messages for OLSR was, as ex-
pected with a higher node density, ten fold higher then
when using 18 nodes. TCP New Reno also performs
better with AODV compared to using OLSR. On the
other hand as the route lookup is more variable with
AODV (OLSR have a route or not), TCP Westwood+
havemore difficultiesto determine the availableband-
width.
Ϭ
ϱϬ
ϭϬϬ
ϭϱϬ
ϮϬϬ
ϮϱϬ
ϯϬϬ
ϯϱϬ
Ks K>^Z
<ďŝƚͬƐ
ϭϲ'ƌŽƵƉƐ ϵϲEŽĚĞƐ
EĞǁZĞŶŽ
ƐƚǁŽŽĚн
>&E
Figure 2: High node density with no background traffic.
Figure 3 shows results for an 18 node density sce-
nario with 20 Kbit/s of background traffic. With low
congestion, TCP ELFN outperforms both TCP West-
wood+ and TCP New Reno. In this scenario the TCP
throughput for all TCP versions is reduced as com-
pared to Figure 1, reflecting the lower available band-
width due to the background traffic. Since we are now
sending trafficbetween more node pairs and as AODV
only detects route errors on actively used routes there
is also a higher possibility for route errors. This is dif-
ferent from when using OLSR where all route errors
are detected, whether it is on an active route or not.
The amount of route error messages when us-
ing AODV also increased, with all TCP variants, by
around 50% compared to when we had no back-
ground traffic.
Ϭ
ϭϬϬ
ϮϬϬ
ϯϬϬ
ϰϬϬ
ϱϬϬ
ϲϬϬ
Ks K>^Z
<ďŝƚͬƐ
ϯ'ƌŽƵƉƐϭϴEŽĚĞƐtŝƚŚĂĐŬŐƌŽƵŶĚ dƌĂĨĨŝĐ
EĞǁZĞŶŽ
ƐƚǁŽŽĚн
>&E
Figure 3: Low node density with low amount of background
traffic.
The amount of congestion in the scenario with 180
Kbit/s of background traffic and 96 nodes reduced the
performance of all TCP proposals. In this scenario the
network is overloaded with traffic and therefore none
of the TCP proposals has a throughput higher than 50
Kbit/s, due to space constraints we have omitted the
figure.
5 CONCLUSIONS
In this paper we have investigated the TCP perfor-
mance of three TCP versions in a TDMA based
MANET using both OLSR and AODV as routing
protocols. We used reference point group mobility
model to simulate a typical tactical network scenario.
The results show that TCP ELFN achieves the overall
highest throughput. The gains are, however, reduced
in low throughput scenarios as the benefit of freezing
the congestion window is less. The performance of
TCP ELFN was similar regardless of which routing
protocol was used.
Future work will further investigate the perfor-
mance of TCP ELFN in high density networks and
the possibility to enhance the performance when us-
ing OLSR.
ACKNOWLEDGEMENTS
This research was supported by grant 2008-539 from
the Knowledge Foundation, Sweden, and the Euro-
pean Regional Development Fund through the Inter-
reg IVB North Sea Region Project E-CLIC.
REFERENCES
Abolhasan, M., Wysocki, T., and Dutkiewicz, E. (2004).
A review of routing protocols for mobile ad hoc net-
PerformanceEvaluationforTCPinTacticalMobileAdHocNetworks
281
works. Ad hoc networks, 2(1):1–22.
Aschenbruck, N., Gerhards-Padilla, E., and Martini, P.
(2008). A survey on mobility models for perfor-
mance analysis in tactical mobile networks. Journal
of Telecommunications and Information Technology,
2:54–61.
Camp, T., Boleng, J., and Davies, V. (2002). A survey of
mobility models for ad hoc network research. Wireless
communications and mobile computing, 2(5):483–
502.
Chen, K., Xue, Y., and Nahrstedt, K. (2003). On setting
TCP’s congestion window limit in mobile ad hoc net-
works. In Communications, 2003. ICC’03. IEEE In-
ternational Conference on, volume 2, pages 1080
1084. IEEE.
Clausen, T. and Jacquet, P. (2009). RFC: 2626 Optimized
Link State Routing Protocol (OLSR).
Floyd, S., Henderson, T., and Gurtov, A. (2004). RFC 3782-
The NewReno modification to TCP’s fast recovery al-
gorithm.
Grieco, L. and Mascolo, S. (2002). TCP Westwood and
Easy RED to Improve Fairness in High-Speed Net-
works. In Carle, G. and Zitterbart, M., editors, Pro-
tocols for High Speed Networks, volume 2334 of
Lecture Notes in Computer Science, pages 130–146.
Springer Berlin / Heidelberg.
Grieco, L. and Mascolo, S. (2004). Performance evaluation
and comparison of Westwood+, New Reno, and Vegas
TCP congestion control. ACM SIGCOMM Computer
Communication Review, 34(2):25–38.
Hiertz, G., Denteneer, D., Max, S., Taori, R., Cardona, J.,
Berlemann, L., and Walke, B. (2010). IEEE 802.11
s: The WLAN mesh standard. Wireless Communica-
tions, IEEE, 17(1):104–111.
Holland, G. and Vaidya, N. (2002). Analysis of TCP Per-
formance over Mobile Ad Hoc Networks. Wireless
Networks, 8:275–288. 10.1023/A:1013798127590.
Johnson, D., Hu, Y., and Maltz, D. (2007). RFC 4728:
The dynamic source routing protocol for mobile ad
hoc networks for IPv4.
Li, L., Vigneron, P., Brown, C., Shi, M., and Kunz, T.
(2012). On designing bandwidth constrained mobile
tactical networks for complex terrains. Communica-
tions Magazine, IEEE, 50(2):188–194.
Liao, W., Tseng, Y., and Shih, K. (2002). A TDMA-based
bandwidth reservation protocol for QoS routing in a
wireless mobile ad hoc network. In Communications,
2002. ICC 2002. IEEE International Conference on,
volume 5, pages 3186–3190. IEEE.
Macker, J. et al. (1999). RFC 2501: Mobile Ad hoc Net-
working (MANET): Routing Protocol Performance
Issues and Evaluation Considerations.
McCanne, S., Floyd, S., Fall, K., Varadhan, K., et al. (2012).
Network simulator NS-2. The Vint project, available
for download at http://www.isi.edu/nsnam/ns.
Networking Research Lab (2012). Networking Research
Lab. Online, http://research.csc.ncsu.edu/netsrv/.
Perkins, C. E., Belding-Royer, E. M., and Das, S. R.
(2003). RFC 3561: Ad hoc On-Demand Distance Vec-
tor (AODV) Routing. Published Online.
Rhee, I., Warrier, A., Min, J., and Xu, L. (2006). DRAND:
distributed randomized TDMA scheduling for wire-
less ad-hoc networks. In Proceedings of the 7th ACM
international symposium on Mobile ad hoc network-
ing and computing, MobiHoc’06, pages 190–201,
New York, NY, USA. ACM.
Romanowicz, E. (2008). TCP with Explicit Link
Failure Notification. Published Online, http://
www.ewaromanowicz.com/academics/.
Voorhaen, M. and Blondia, C. (2006). Analyzing the impact
of neighbor sensing on the performance of the OLSR
protocol. In Modeling and Optimization in Mobile, Ad
Hoc and Wireless Networks, 2006 4th International
Symposium on, pages 1–6. IEEE.
Wang, F. and Zhang, Y. (2002). Improving TCP perfor-
mance over mobile ad-hoc networks with out-of-order
detection and response. In Proceedings of the 3rd
ACM international symposium on Mobile ad hoc net-
working & computing, pages 217–225. ACM.
WINSYS2012-InternationalConferenceonWirelessInformationNetworksandSystems
282