This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts

for publication in the IEEE INFOCOM 2009 proceedings.

Minimizing End-to-End Delay: A Novel Routing

Metric for Multi-Radio Wireless Mesh Networks
Hongkun Li, Yu Cheng, Chi Zhou Weihua Zhuang
Department of Electrical and Department of Electrical and
Computer Engineering Computer Engineering
Illinois Institute of Technology University of Waterloo
{hli55, cheng, zhou}@iit.edu [email protected]

Abstract—This paper studies how to select a path with the X 1 Y

minimum cost in terms of expected end-to-end delay (EED) in a 0.1
multi-radio wireless mesh network. Different from the previous 1 M = 10 M = 10
efforts, the new EED metric takes the queuing delay into account, D
since the end-to-end delay consists of not only the transmission S
delay over the wireless links but also the queuing delay in the
buffer. In addition to minimizing the end-to-end delay, the EED 0.5 M=2 M=2 M=3 1
metric implies the concept of load balancing. We develop EED-
based routing protocols for both single-channel and multi-channel 0.1 1
wireless mesh networks. In particular for the multi-radio multi- A B C
channel case, we develop a generic iterative approach to calculate
a multi-radio achievable bandwidth (MRAB) for a path, taking Fig. 1. The impact of queuing delay on path selection.
the impacts of inter/intra-flow interference and space/channel
diversity into account. The MRAB is then integrated with EED
to form the metric of weighted end-to-end delay (WEED). As being served at the MAC layer [13], [15], while in many cases
a byproduct of MRAB, a channel diversity coefficient can be
the queuing delay takes a significant portion of the total delay
defined to quantitatively represent the channel diversity along a
given path. Both numerical analysis and simulation studies are over a hop. The delay through a node, which has many packets
presented to validate the performance of the routing protocol in queue but short transmission time, could be larger than
based on the EED/WEED metric, with comparison to some well- through the one, which has less packets in the queue but longer
known routing metrics. transmission delay.
I. I NTRODUCTION We here use an example, as illustrated in Fig. 1 to emphasize
the impact of network-layer queuing delay on routing. The
Routing in wireless mesh networks has been a hot re-
number annotating each link is the success probability for
search area in recent years, with the objective to achieve
a transmission over the link, denoted as psuc , which means
as high throughput as possible over the network. The main
on average it takes 1/psuc transmission trails to successfully
methodology adopted by most of the existing work is selecting
deliver a packet. The number M denotes the number of packets
path based on interference-aware or load-balancing routing
in the network-layer queue, waiting to be served by the MAC
metrics to reduce network-wide channel contentions. It has
layer. Suppose that the bandwidth of each link is 11Mbit/s and
been revealed that the capacity of a single-radio multi-hop
the packet length is 1100bytes; it gives a transmission time of
wireless network can not scale up with the network size,
0.8ms. If the queue delay is not included, routing based on the
due to the co-channel interference [1]–[3]. The multi-radio
expected transmission time (ETT) would prefer the path S-X-
multi-channel connection has been widely considered as an
Y-D (9.6ms) over the path S-A-B-C-D (11.2ms). Nevertheless,
efficient approach to increase the wireless network capacity
the path S-A-B-C-D would be the better one with the queuing
[8]. Design of efficient routing schemes for multi-radio multi-
delay taken into account. In this case, the end-to-end delay
channel wireless mesh network is much more challenging
over S-X-Y-D is 97.6ms, but only 24 ms over S-A-B-C-D. In
compared to the single-channel case.
this example, the delay values ignore the backoff overhead,
Many popular multimedia applications, e.g., voice over IP,
which will be considered in our routing metric design.
IPTV, and on-line gaming, have strict delay requirement. In
The newly proposed routing metric of end-to-end delay
this paper, we aim at designing a routing metric to minimize
(EED) in fact exploits the cross-layer design: each node needs
the end-to-end delay, considering not only the transmission
to not only monitor the transmission failure probability at
delay at the medium access control (MAC) layer, but also
the MAC layer to estimate the MAC transmission delay, but
the queuing delay at the network layer. Most of the previous
also count the number of packets waiting in the network-
studies focus only on the transmission delay of the packet
layer buffer to estimate the queuing delay. The EED metric
This work was supported in part by NSF grant CNS-0832093. also implies the concept of load-balancing. The path with

Authorized licensed use limited to: Technische Informationsbibliothek (TIB). Downloaded on January 04,2024 at 14:55:41 UTC from IEEE Xplore. Restrictions apply.
978-1-4244-3513-5/09/$25.00 ©2009 IEEE 46
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE INFOCOM 2009 proceedings.

minimum EED normally passes through the links with less count (ETX) [21] to describe the channel contention level
packets in the queue, and thus balances the traffic off those experienced by a wireless link, which works well in a ho-
congested links. Moreover, counting the number of packets in mogeneous single-radio environment. However, ETX is not
the buffer is a convenient implementation; most of the existing capable of describing the complex scenarios in a multi-radio
load-balancing routing schemes require the traffic information wireless mesh network, normally involving inter-/intra-flow
available, which is usually not easy to obtain in practice [16]. interferences and different rate/intererence/topology profiles
In addition to the transmission delay and queuing delay over different channels. mETX and ENT [6] are proposed to
at each hop, the end-to-end delay over a multi-hop wireless enhance ETX by considering the variable link reliability. The
network is particularly impacted by the interferences among ETOP metric enhances ETX by incorporating the impact of
different hops, which can be classified into inter-flow and intra- link positions [5].
flow interference [23]. In this paper, we further propose a path A bandwidth-aware routing with QoS requirement is pro-
metric called multi-radio achievable bandwidth (MRAB) to posed in [26]. The link metric of expected transmission time
accurately capture the impacts of inter/intra-flow interferences (ETT) and the associated path metric of weighted cumulative
and space/channel diversity along a path. We consider a ETT (WCETT) are proposed in [13] for multi-channel mesh
practical scenario that an end-to-end path may consist of networks, which try to enhance the ETX by counting the
both multi-radio hops and single-radio hops, where different heterogeneous channel rate and intra-flow interference, but
channels do not interfere with each other but interferences exist the inter-flow interference is still not considered. Furthermore,
within the same channel. We particularly develop a sub-path when calculating the intra-flow interference, WCETT always
based iterative approach to model the complex interactions takes all links into account and overlooks the situation that two
among inter-flow interference, intra-flow interference, and links far away enough can transmit packets simultaneously.
simultaneous transmission due to space and channel diversity. The metric of interference and channel switching (MIC)
The MRAB is then integrated with EED to form the metric [15] incorporates both inter-flow and intra-flow interference,
of weighted end-to-end delay (WEED). As a byproduct of whereas it only contains the number of interfering nodes
MRAB, a channel diversity coefficient can be defined to rather than the total amount of interference on these nodes
quantitatively represent the channel diversity along a given for the inter-flow interference. In [25], we propose a metric
path. We evaluate the performance of the WEED based routing of multi-hop effective bandwidth (MHEB) to compute the
protocol via numerical analysis and ns2 simulations, with usable bandwidth when both inter- and intra-flow interferences
comparison to some popular metrics, under both single and are present. However, the MHEB metric just uses a simple
multiple channel cases. It is confirmed that the EED/WEED weighted average to combine the inter- and intra- flow in-
metric consistently yields better performance. terferences. In this paper, the MRAB is based on MHEB, but
The reminder of this paper is organized as follows: Sec- use a more accurate approach to capture the complex interplay
tion II reviews more related work. Section III derives the between the two types of interferences.
routing metric of EED. Section IV presents the algorithm to
III. E ND - TO -E ND D ELAY M ETRIC
compute the MRAB, which captures the interaction between
the inter- and intra-flow interferences. The MRAB metric is The end-to-end delay over a path is the summation of delays
integrated with the EED metric to form the WEED metric experienced by all the hops along the path. For convenience,
for routing over the multi-radio mesh networks. The routing we also use EED to denote the delay metric at each link.
protocol is described in Section V. Section VI presents some The meaning of EED will be clear in the context. In order
numerical analysis and simulation studies to validate the to compute the EED metric over a wireless channel, each
routing performance based on the EED/WEED metric, with node needs to monitor the number of packets buffered at
comparison to some well-known routing metrics. Section VII the network layer waiting for MAC layer service, as well
gives the concluding remarks. as measuring the transmission failure probability at the MAC
layer. The transmission failure probability is the probability
II. R ELATED WORK that a MAC-layer transmission fails due to either collisions or
The routing metric plays a critical role in a routing protocol. bad channel quality. While counting the number of packets in
The studies in [8], [16], [17] design routing metrics for load- the queue is straightforward, we will discuss how to measure
balancing in a multi-hop wireless network. The routing metrics the transmission failure probability over a link in Section V.
however require the real-time traffic information. To exploit The EED over a link i, say between node ni and ni+1 , consists
the space diversity, the link conflict graph is normally applied of the queuing delay and transmission delay as
to model the interference among different hops [18], and
EEDi = E [queuing delay + transmission delay] . (1)
the interference clique transmission time is proposed as a
routing metric in [20]. However, the conflict graph based The transmission delay can also be interpreted as the packet
approaches normally induce large computation overhead in service time, which is defined as the period from the instant
searching for the maximal independent sets or cliques, and that a packet begins to be serviced by the MAC layer to the
are not suitable for dynamic distributed routing protocols. instant that it is either successfully transmitted or dropped after
De Couto et al. propose the metric of expected transmission a predefined number of retransmissions.

Authorized licensed use limited to: Technische Informationsbibliothek (TIB). Downloaded on January 04,2024 at 14:55:41 UTC from IEEE Xplore. Restrictions apply.
47
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE INFOCOM 2009 proceedings.

Suppose that the 802.11 distributed coordination function continuously go through the path hop-by-hop. However, in
(DCF) MAC protocol is used, each transmission or retrans- a multi-hop wireless network, if two links working over the
mission includes protocol overhead due to the binary backoff same channel are located close, when one link is in trans-
mechanism [28]. Let pi denote the transmission failure prob- mission, the MAC protocol will freeze the other link. Such
ability over link i, and assume it is stable through all the channel freezing can be due to either intra-flow transmissions
retransmissions of the packet. Also, let Ti denote the packet or inter-flow transmissions, which results in extra delays in
service time over link i, and K the maximum number of addition to the basic EED as shown in (6). In the following
retransmissions. The average transmission delay section, we will discuss how to extend the basic EED with the
K+1
 k   co-channel interferences taken into account.
k−1 I{k<K+1} L
E[Ti ] = pi (1 − pi ) E[Wj ] +
j=1
B IV. ACHIEVABLE BANDWIDTH OVER A MULTI - RADIO
k=1
(2) MULTI - CHANNEL PATH

where Wj denotes the contention window at the jth backoff In this section, we will develop an algorithm to compute the
stage, and L, B denote the packet length and link bandwidth, achievable bandwidth along a multi-radio multi-channel path,
respectively. According to the 802.11 standard [7], [28], termed as multi-radio achievable bandwidth, by capturing the
Wj = 2j−1 Wmin if ignoring the constraint of backoff stage, complex interplay between inter-flow and intra-flow interfer-
W −1
and E[Wj ] = j2 . In (2), the indicator I(A) is equal to 1 if ences. The end-to-end delay over a multi-radio multi-channel
A is true, which is incurred to include the case that a packet is path can be described more accurately by incorporating the
dropped when the retransmission limit is reached. In addition, MRAB metric into the EED computation to form a new metric
the MAC overhead due to acknowledgement is incorporated WEED. A side-effect benefit of MRAB analysis is that a
into the packet length L for convenience. After applying some channel diversity coefficient can be defined to quantify the
manipulations over (2), we can get resource consumption along a multi-radio multi-channel path.
 
L 1 − pK i
E[Ti ] = + E[backoff time] (3) A. Multi-Radio System
B 1 − pi
with We consider a wireless mesh network, where each node is
K+1 k
equipped with one or more radio interfaces. The interfaces
 
E[backoff time] = pik−1 (1 − pi )I{k<K+1} E[Wj ] assigned with different channels, located either in the same
k=1 j=1
node or in different nodes, can be active simultaneously.
  Thus, the network throughput could be significantly improved
Wmin 1 − (2pi )K+1 1 − pK
i
= − . (4) compared with single-radio system [8]. The interfaces work-
2(1 − 2pi ) 2(1 − pi ) ing on different channels would form distinct interference
If there are Mi packets in the queue when a new packet topologies. Channel assignment [10] plays an critical role
reaches node ni , the EED metric can be defined as in determining the interference topology, and then impacts
on system performance. The channel assignment itself is a
EEDi = (Mi + 1)E[Ti ] (5)
challenging research topic, which is out of the scope of this
which means the total delay passing through the hop equals paper. We assume that the channel assignment for each node
the MAC service time of those packets queuing ahead of the is given. All the nodes are stationary, and any one can be used
new packet plus the MAC service time of the new packet as a router. We define the transmission range of a node as one
itself. Note that the EED value in (5) implies the memoryless hop, while the interference range is r (≥ 2) hops. We consider
property of the packet service time, as the head-of-line packet that the WMN operates over the IEEE 802.11 based MAC, and
may only need to finish a residue packet service time when assume that the routing control information exchanged among
the new packet comes in. It is well-known that only an neighbor nodes is error free.
exponentially distributed services time has the memoryless We adopt the physical model presented in [18] to describe
property. It has been demonstrated in [30] that the MAC packet the interference among different hops. Such an interference
service time over a 802.11 DCF can indeed be approximated model indicates that a transmission from node u to v is
by an exponential random variable. successful if the signal to interference and noise ratio (SINR)
Consider an end-to-end path including H hops, the EED at receiver v is above the pre-determined threshold γ, i.e.,
metric for the path is defined as
Pv (u)
H
 ≥γ (7)
N+ k∈v Pv (k)
EED = EEDi . (6)
i=1 where N denotes the background noise, Pv (u) the received
We would like to emphasize that the EED given in (6) does power at node v from node u, v the set of nodes located in
not consider the effect of co-channel interference in the multi- interference range of v, and Pv (k) the interference power from
hop wireless networks, which assumes that all the packets can an interfering node k.

Authorized licensed use limited to: Technische Informationsbibliothek (TIB). Downloaded on January 04,2024 at 14:55:41 UTC from IEEE Xplore. Restrictions apply.
48
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE INFOCOM 2009 proceedings.

B. Multi-Radio Achievable Bandwidth

1) Inter-flow interference: we first compute the achievable
bandwidth under the inter-flow interference (ABITF). Every
node can monitor the received power to infer the magnitude A 1 B 2 C 1 D 3 E

of the inter-flow interference around its neighborhood. Based

on the interference model (7), the SINR threshold implicitly
denotes the maximum interference a node could tolerate to
process a successful communication. We define the interfer-
ence degree ratio IDRi (uv) for link i between u and v as: Fig. 2. Illustration of a multi-radio multi-channel path.

k∈v Pv (k)
IDRi (uv) = . (8) where L denotes the packet length. It can then be obtained
Pmax
that
The ratio reflects the utilization of the channel assigned to Bi Bj
link i. Pmax is the maximum tolerable interference power at ABIRF (ij) = . (11)
Bi + Bj
receiver and can be calculated by (7). k∈v Pv (k) is sum of
undesired powers at node v from other transmissions. Note Extending the ABIRF (ij) result to the whole sub-path can be
that if there is no interference, the IDR is 0, implying that iteratively implemented. In each iteration, consider those links
entire bandwidth of this channel is available for link i. On the having been processed as one virtual link, with the bandwidth
contrary, an IDR of 1 will indicate that the channel has been set as the ABIRF value already obtained; and then apply the
fully occupied by other links, and no additional bandwidth computation of (11) over the virtual link and the next-hop
can be allocated for link i until the ratio becomes less than 1. link. Note that the impact of inter-flow interference on path
Based on this definition, we evaluate the ABITF at link i as capacity can be conveniently integrated with the intra-flow
(1 − IDRi )Bi interference impact by using the ABITF value given by (9)
ABIT Fi = (9) as link capacity in ABIRF computation, instead of using the
ET Xi
physical bandwidth.
where Bi denotes the physical bandwidth of link i, and ET Xi 3) Multi-radio achievable bandwidth: The multi-radio
[21] denotes the expected transmission attempts to achieve a multi-channel connection makes the capacity analysis of a
successful transmission over link i. The value (1−IDRi )∗Bi sub-path more complicated. When two links work on dif-
indicates the available bandwidth for a transmission under the ferent channels through different radio interfaces, they could
inter-flow interference. The physical meaning of (9) can be send/receive packets simultaneously without interference. It is
interpreted as: given the transmission failure probability pi , possible that the two end-hops of a sub-path are co-channel
a successful transmission needs ET Xi attempts in average; links, while other hops in the middle may work on different
the bandwidth is effectively used for only one of the ET Xi channels. The iterative procedure discussed above to compute
transmission attempts. the ABIRF for a co-channel sub-path can also be extended
2) Intra-flow interference: There exists intra-flow interfer- to the multi-channel sub-path, with consideration that the
ence if two links belonging to the same path work on the achievable bandwidth over two links is min(Bi , Bj ), when
same channel and are located within each other’s interference link i and j work over different channels. Specifically, the
range, i.e. within r (≥ 2) hops. We define a new concept of iterative steps to compute the ABIRF for a given multi-radio
sub-path spanning r + 2 hops, based on the observation that a sub-path is as follows:
link will potentially interfere with another link at most r + 2
Step 1 : For the first link of the sub-path, set ABIRF equal
hops away. A sub-path with r = 2 is illustrated in Fig.2. In
to the ABITF associated with the channel over which
general, a H-hop path contains H − r − 1 sub-paths.
the link works.
Under the impact of intra-flow interference, a sub-path
Step 2 : Go to the next link, say link i, and check whether
is equivalent to a virtual link. The reason is that a new
the channel of the next link is different from those
packet can enter a sub-path only after the previous one leaves.
used by the previous links or not. If different, go to
The achievable bandwidth over a sub-path can be iteratively
step 3; otherwise go to step 4.
obtained from the achievable bandwidth over two interfering
Step 3 : In this case, set
links. For example, consider two neighboring co-channel links
i and j along a path. Links i and j have bandwidth Bi ABIRF = min(ABIRF, ABIT Fi ) (12)
and Bj , respectively. Since the two links can not be active
simultaneously, the equivalent achievable bandwidth under the and then go to step 5.
intra-flow interference (ABIRF) over links i and j, denoted as Step 4 : In this case, set
ABIRF (ij), satisfies ABIRF × ABIT Fi
ABIRF = (13)
L L L ABIRF + ABIT Fi
= + (10)
ABIRF (ij) Bi Bj and then go to step 5.

Authorized licensed use limited to: Technische Informationsbibliothek (TIB). Downloaded on January 04,2024 at 14:55:41 UTC from IEEE Xplore. Restrictions apply.
49
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE INFOCOM 2009 proceedings.

Step 5 : Check whether all the (r + 2) hops of the sub-path the MRAB value guarantees that MRAB will not increase
have been considered or not. If not, go to step 2; when one more sub-path is generated. Hence, MNRAB P ·L
part
otherwise, terminate the iteration. achieves a larger or equal value too in the the second case,
For a H-hop path including multiple sub-paths, let when one hop is involved. Since both parts constituting the
ABIRFj denote the achievable bandwidth over the jth sub- WEED metric becomes larger or maintains equal when the
path. The multi-radio achievable bandwidth can be computed path goes longer, WEED is monotonic. The proof applies to
as: both left-monotonicity and right-monotonicity.

M RAB = min (ABIRFj ) (14) D. Channel Diversity Coefficient

A challenging issue being widely studied in the area of
for j = 1, 2, ...H − r − 1. If H − r − 1 ≤ 0, we set j = 1,
multi-channel wireless networks is how to quantify the chan-
which means the path is so short that there is only one sub-
nel diversity for a given path. Intuitively, an ideal quantity
path along the whole path. The computation in (14) exploits
describing the channel diversity incorporates the impacts from
the bottleneck concept, but is applied at the sub-path level
various aspects, including the number of hops, the number of
instead of the link level.
channels, and the interference relationship among the links.
C. WEED Metric Our development efforts in the above have demonstrated that
1) Weighted end-to-end delay: Over a multi-radio multi- the MRAB metric indeed takes all of the factors into account.
channel path, the MRAB metric is integrated with the EED Therefore, we define a channel diversity coefficient (CDC)
metric to form a weighted end-to-end delay metric as: based on the MRAB as
M RAB
H
 NP · L CDC = (16)
W EED = α × EEDi + (1 − α) × (15) Bs
i=1
M RAB where Bs denotes the achievable bandwidth of the same path
where 0 ≤ α ≤ 1 is tunable weight factor. In the WEED if all links of this path work over the same channel, defined
metric, the first part represents the accumulation of the delivery as the single-channel path capacity. For the convenience of
delay due to hop-by-hop transmissions in a store-and-forward comparison, we adopt the smallest ABIT Fi value as the
manner; the second part represents the extra delay due to single-channel path capacity Bs . Thus, CDC is always larger
the interference nature of a multi-hop wireless network. In than or equal to 1, and a higher CDC indicates a better channel
particular, the NP induced in (15) denotes the total number diversity.
of packets queued in the buffers along the path, because the V. ROUTING P ROTOCOL D ESIGN
interference effect described by the MRAB applies to all the
packets that are being served by the path. A. Basic DSR Implementation
We would like to emphasize that the WEED metric contains We implement the proposed EED/WEED based routing
not only the end-to-end delay information regarding a single by modifying the dynamic source routing (DSR) protocol
packet transmission, but also the transmission delay for a block [12]. We select DSR due to the following reasons: (i) DSR
of packets due to the bottleneck bandwidth MRAB. Therefore, is one of the most popular protocols in multi-hop wireless
selecting a shortest path based on the WEED metric tends to networks, and the implementation codes are publicly available.
minimize both the short-term and the long-term delay. (ii) We have noticed that some well-known routing metrics,
2) Monotonicity analysis: It has been indicated in [11] such as ETX, ETT and WCETT, are implemented based on
that monotonicity is one of the necessary properties of a DSR; a common implementation will significantly facilitate
routing metric to result in a consistent and loop-free routing the performance comparison among different routing metrics.
implementation. For example, the well-known WCETT metric (iii) The EED/WEED metric does not have the property of
[13] is monotonic. We here prove that WEED also has the isotonicity [11], so the source routing approach (adopted by
property of monotonicity. DSR) is preferred to guarantee the optimality, consistency, and
H
Consider a given path, it is obvious that the i=1 EEDi loop-freeness in routing.
part will becomes larger when one more hop is attached to With basic DSR, a node attempts to find a route to a given
the path. For the MRAB part, one more hop may lead to two destination by initiating a route request (RREQ) message.
possible cases. In the first case, the number of sub-paths does Every RREQ has a unique broadcast ID to prevent routing
not change; the new hop just makes the (H − r − 1)th sub- loops and redundant flooding. An intermediate node will
path one hop longer. According to the iterative computation further broadcast a RREQ only when the broadcast ID appears
given in Section IV-B , the ABIRF of a sub-path is less for the first time; also the node will insert its address in the
than or equal to the previous value when one more hop is source route field of the RREQ message. Once the destination
included. Thus, the MNRAB P ·L
part achieves a larger/equal value receives the RREQ, it will reverse the hop sequence of the
when the path goes longer. In the second case, a new sub- received path and insert the reversed path into the source route
path is incurred by the new hop, while the existing sub-paths field of a route reply (RREP) message, which is then unicasted
will not change. The “minimization” operation in obtaining back to the source node. The source node will determine a

Authorized licensed use limited to: Technische Informationsbibliothek (TIB). Downloaded on January 04,2024 at 14:55:41 UTC from IEEE Xplore. Restrictions apply.
50
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE INFOCOM 2009 proceedings.

shortest path to a destination based on the path information it 3.069

receives from the RREP messages. With DSR, all the RREP
messages and data packets carry a complete path between a 2.728

Total throughput (Mbps)

certain source/destination pair in their source route fields. All
2.387
the nodes can overhear such path information and store it in
their route caches for later use. In cases that a node finds that 2.046

a packet could not be successfully delivered over a link after

1.705
a maximum number of retransmission, the node will return
a route error (RRER) to the source node of the path; every 1.364
intermediate node receiving the RRER message will mark this ETX
ETT
link invalid. 1.024
EED

B. EED/WEED Based Routing 0.682

5 10 15 20 25 30
Flow rate (pkts/second)
1) EED link metric: The EED metric itself can be used
as a routing metric, especially for single-channel multi-hop (a) Total network throughput
networks. We consider that a directional link is defined by an
upstream end and a downstream end, and the communication 2.5

between two neighbor nodes is through two directional links.

End−to−end delay (seconds)

To obtain the EED link metric, a downstream node needs to 2

monitor the transmission failure probability over the link and

knows the number of packets in the upstream node’s buffer. In 1.5
our routing protocol, each node periodically broadcasts probe
packets to its downstream neighbors at a predetermined rate
λ, with the number of packets M in its buffer carried in
1

each probe. Each downstream node maintains a neighbor list.

When the downstream node receives a probe packet, it will 0.5
ETX

update the value M for the corresponding upstream node in ETT

EED
its neighbor list. Moreover, a downstream node will count the 0
5 10 15 20 25 30
number of probes received from each upstream node during a Flow rate (pkts/second)
period T ; use Vi to denote the number of probes received from
(b) Average end-to-end delay
the upstream node associated with link i. The transmission
Vi
failure probability over link i can be estimated as pi = λT . Fig. 3. The routing performance in grid topology versus flow rate.
The pi value will also be stored into the neighbor list. After a
downstream node finishes processing a received probe packet, the message. To calculate the MRAB of each sub-path, two
the probe will be discarded. extra fields need to be added to each record in the neighbor
A new field, called link metric is established in the RREQ list: the channel ID indicating the channel associated with
message to store the metric value of each link. Once a node a link, and the IDR value computed according to (8) based
receives a RREQ, it first checks its own neighbor list to get the on power monitoring at the downstream node. When a node
values of Mi and pi , and then computes the EED according to forwards a RREQ message, the channel ID, the IDR values,
(5). The metric EEDi will be inserted into the RREQ message and EED all are attached as link metrics. Once the destination
if the node needs to continuously forward it. In addition, every receives the RREQ, it will return a RREP to the source node,
node collects the EED metrics carried in the RREQ or RREP which duplicates all the routing information retrieved from
messages and store them in a link cache. The cached link the RREQ. Based on the RREP messages, the source can
metrics can be used to establish a network topology. If a node finally compute a shortest path regarding the WEED metric,
has a packet for a destination with the path information not according to (15).
established yet, the node can apply the Dijkstra’s algorithm to
compute an EED based shortest path based on the link cache VI. P ERFORMANCE E VALUATION
information.
2) WEED over a multi-radio path: To implement the A. EED-Based Routing in a Single-Channel Network
WEED based routing, each radio interface in the network is The EED metric by itself can be used as an efficient routing
uniquely identified with a separate IP address for each inter- metric, since it effectively captures not only the queuing delay
face. By considering each interface as the entity involved in the at the network layer but also the retransmission delay at the
routing, the operations of probing, maintaining neighbor list MAC layer. We use NS2 simulation results for a single-channel
and link cache, estimating the transmission failure probability wireless mesh network to demonstrate the performance, with
are similar to the node based case. There is an additional field comparison with the two well-know metrics ETX and ETT.
in all control messages to denote which interface is processing The MAC protocol used for simulation is 802.11b.

Authorized licensed use limited to: Technische Informationsbibliothek (TIB). Downloaded on January 04,2024 at 14:55:41 UTC from IEEE Xplore. Restrictions apply.
51
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE INFOCOM 2009 proceedings.

2 2.7

2.65
1.8
2.6

Total throughput (Mbps)

Total throughput (Mbps)

1.6
2.55
1.4
2.5

1.2 2.45

2.4
1
2.35
0.8
ETX 2.3
0.6 ETT
EED 2.25

0.4 2.2
5 10 15 20 25 30 20 40 60 80 100 120 140 160
Flow rate (pkts/second) Update interval (seconds)

(a) Total network throughput (a) Total network throughput

4 1.65

3.5 1.6

End−to−end delay (seconds)

End−to−end delay (seconds)

3 1.55

2.5 1.5

2 1.45

1.5 1.4

1 1.35
ETX
0.5 ETT 1.3
EED

0 1.25
5 10 15 20 25 30 20 40 60 80 100 120 140 160
Flow rate (pkts/second) Update interval (seconds)

(b) Average end-to-end delay (b) Average end-to-end delay

Fig. 4. The routing performance in random topology versus flow rate. Fig. 5. The impact of EED update interval on routing performance.

We specifically consider two network topologies. The first transmission failure probability for computing the link metric,
is a grid topology over a 1400m×1400m area. The area is di- while the transmission failure probability is directly related to
vided into 200m×200m square cells, with each cell containing the MAC throughput [7]. In most of the cases, ETT has slightly
one network node in its center. Four flows are deployed at the higher throughput, which is due to the larger computation
1st, 3rd, 5th, and 7th rows of the grid, respectively, with the overhead with EED and implementation overhead due to path
source/destination nodes of each flow located at both ends of change incurred by the random queue length behavior. In
the row correspondingly. The other topology randomly places addition, we observe in the grid topology that the throughput
40 nodes in a 1000m×1000m area with necessary adjustment curves under all the three metrics become flat when the flow
to maintain the connectivity. We also run 4 flows over the rate exceeds a certain level, which indicates that the network
random topology. For both topologies, the transmission range is approaching its maximum capacity.
is 250m, and the interference range is 550m. All flows are We next examine the impact of EED update interval on
constant-bit-rate (CBR) flow, with the packet size of 512 bytes. the routing performance. The grid topology is simulated, with
Fig. 3 and Fig. 4 present the total network throughput the queue size at each node being 20 packets and flow rate
and end-to-end delay, under ETX, ETT, and EED metrics, 25 packets/second. Fig. 5(a) and 5(b) present the network
respectively, versus the flow rate. The queue size at each throughput and the end-to-end delay versus the EED updating
node is 20 packets, and the link metric update interval in our interval, respectively. From the figure, it can be seen that
EED implementation is 50 seconds. In the two figures, it is both inappropriate small and large intervals result in low
explicitly demonstrated that EED metric can result in much throughput and large delay. On one hand, an inappropriate
better end-to-end delay performance than ETX and ETT, under small update interval induces over-frequent link metric update
both the grid and random topologies. Regarding the network and results in a large bandwidth overhead. On the other hand,
throughout, ETT and EED has the similar performance, while an inappropriate large update interval will not timely respond
outperforming the ETX. The reason that ETT and EED have to a congested link and result in unnecessary packet loss due
similar throughput performance is that both of them exploit the to a full buffer. Fig. 5 demonstrates that an optimal update

Authorized licensed use limited to: Technische Informationsbibliothek (TIB). Downloaded on January 04,2024 at 14:55:41 UTC from IEEE Xplore. Restrictions apply.
52
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE INFOCOM 2009 proceedings.

S 1 5 2 8 3 D
2.8 I
0.2 0.2 0.25

2.7
1 2 1
II S 3 2 D
Total throughput (Mbps)

0.1 0.3 0.35

2.6

1 2 3 1
2.5 III S 2 5 4 D
0.2 0.15 0.25 0.4
2.4
1 2 1 3
IV S 2 4 6 D
2.3 0.1 0.2 0.3 0.2

2.2
ETX Fig. 7. WEED and WCETT Examples.
ETT
EED
TABLE I
2.1
5 10 15 20 25 30 35 40 45 50 PARAMETERS USED IN NUMERICAL ANALYSIS
Queue size (pkts)

(a) Total network throughput Parameters Values

Packet length 600 bytes
3.5
Link bandwidth over channel 1 8 Mbps
Link bandwidth over channel 2 12 Mbps
3 Link bandwidth over channel 3 6 Mbps
End−to−end delay (seconds)

Minimum contention window 0.02 ms

2.5
Maximum retransmission times 5

1.5 package. Thus, we resort to numerical analysis here to obtain

the results.
1
The path selection example is shown in Fig.7, where the
0.5
ETX interference range r = 1 hop. There are four candidate
ETT
EED
paths with three possible channels. The numbers within each
0
5 10 15 20 25 30 35 40 45 50
circle indicate the number of packets waiting in the buffer
Queue size (pkts) of that node. The channel assignment is marked above the
(b) Average end-to-end delay
link, and the transmission failure probability below. Other
necessary parameters for WEED computation are illustrated
Fig. 6. The impact of queue size on routing performance.
in Table I. For the convenience of analysis, we ignore the
power monitoring part for inter-flow interference analysis,
interval exists that can lead to largest throughput and smallest
while the intra-flow interference analysis here is sufficient
end-to-end delay.
for demonstrating the capability of WEED in quantifying the
We further evaluate the effect of queue size on the rout-
channel diversity.
ing performance, with results illustrated in Fig. 6. The grid
We compare the path selection based on WEED and
topology is considered with the EED update interval of 50
WCETT, with results shown in Table II and III. WCETT
seconds and the flow rate of 20 packets/second. Regarding
prefers path I over II, and IV over III, while WEED prefers
the throughput as shown in Fig. 6(a), we can see that ETT
the other way. By investigating the configuration of Path I and
performance better for small buffer sizes, while EED better
II, it is obvious that WEED makes a different decision from
for large ones. The reason is that when the buffer size is small,
WCETT’s because it takes the number of packets in the buffers
in most of the cases all the buffers are full, where the EED
into consideration. The advantage of WEED over WCETT can
could not exploit more benefit compared to the ETT. The extra
be demonstrated by comparing path III with path IV. Since the
computation overhead and route updating overhead, however,
two interfering links over channel 1 are located further away
will lead to a smaller network throughput. When the buffer is
in path III than in path IV, it is intuitively clear that path
large, EED can select a path with more buffer space, which
III suffers a less amount of intra-flow interference. Moreover,
will lead to less tail-dropping of the packets and thus a higher
both path III and path IV consist of the same set of links (two
throughput. Regarding the delay as shown in Fig. 6(a), EED
channel-1 links, one channel-2, one channel-3), so path III will
achieves the smallest end-to-end delay in all the cases.
achieve a higher path capacity than path IV due to the less
B. Metric Comparison in Multi-channel Environment interference over it. Thus, WCETT makes a wrong decision,
In this part, we use an example to illustrate the effectiveness while WEED makes a correct one. The reason for WCETT’s
of WEED in quantifying the capacity of a multi-radio multi- wrong decision is that it deems that co-channel links along a
channel path. Since there is no much reference on NS2 path always interfere with each other, disregard the distance
simulation of DSR-based routing over multi-radio wireless among them. With r = 1, it is obvious that the two channel-1
networks, we are still developing such a NS2 simulation links in path III will not interfere with each other, in fact.

Authorized licensed use limited to: Technische Informationsbibliothek (TIB). Downloaded on January 04,2024 at 14:55:41 UTC from IEEE Xplore. Restrictions apply.
53
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE INFOCOM 2009 proceedings.

TABLE II
N UMERICAL R ESULTS FOR WCETT [13] [6] C. Koksal and H. Balakrishnan, “Quality-Aware Routing Metrics
for Time-Varying Wireless Mesh Networks,” IEEE J. Select. Areas
H Commun., vol. 24, no.11, pp. 1984–1994, November.2006.
Path i=1 ET Ti max(Xj ) WCETT (β = 0.5) [7] Y. Cheng, X. Ling, W. Song, L.X. Cai, W. Zhuang, and X. Shen “A
I 2.3881 1.0667 1.7274 Cross-Layer Approach for WLAN Voice Capacity Planning,” IEEE J.
II 2.1612 1.5898 1.8755 Select. Areas Commun., vol. 25, no. 4, pp. 678–688, May 2007.
III 3.2873 1.75 2.5187 [8] A. Raniwala, T.-c. Chiueh, “Architecture and Algorithms for an IEEE
802.11-Based Multi-Channel Wireless Mesh Network,” in Proc. IEEE
IV 3.0238 1.5238 2.2738 INFOCOM, 2005, pp. 2223–2234.
[9] A.Cerpa, J. L. Wong, M. Potkonjak and D. Estrin, “Temporal Properties
of Low Power Wireless Links: Modeling and Implications on Multi-Hop
TABLE III
Routing,” in Proc. ACM MobiHoc, 2005, pp. 414–425.
N UMERICAL R ESULTS FOR WEED
[10] K. N. Ramachandran, E. M. Belding, K. C. Almeroth and M. M.
H
Buddhikot, “Interference-Aware Channel Assignment in Multi-Radio
Path i=1 EEDi MRAB WEED (α = 0.5) CDC Wireless Mesh Networks,” in Proc. IEEE INFOCOM, 2006, pp.1-12.
I 16.6991 6Mbit/s 13.5496 3.0 [11] Y. Yang, J. Wang, “Design Guidelines for Routing Metrics in Multihop
II 5.8638 4Mbit/s 5.9319 2.0 Wireless Networks,” in Proc. IEEE INFOCOM, 2008, pp.1615 - 1623.
[12] D. B. Johnson, D. A. Maltz and Y. Hu, “The Dynamic Source Routing
III 13.8753 6Mbit/s 11.33765 4.0
Protocol for Mobile Ad Hoc Networks (DSR),” in IETF, INTERNET-
IV 13.6927 4Mbit/s 14.04635 2.6667 DRAFT, 2003, April.
[13] R. Draves, J. Padhye, and B. Zill, “Routing in Multi-Radio, Multi-Hop
Wireless Mesh Networks,” in ACM MOBICOM, 2004, pp. 114–128.
[14] R. Draves, J. Padhye, and B. Zill, ”Comparison of Routing Metrics for
Table III illustrates that the MRAB and CDC value can Static Multi-Hop Wireless Networks,” in ACM SIGCOMM, 2004, pp.
133–144.
quantitatively demonstrate the capacity difference between [15] Y. Yang, J. Wang, and R. Kravets, “Designing Routing Metrics for
two multi-radio multi-channel paths. Moreover, the MRAB Mesh Networks,” in Proc. IEEE Workshop on Wireless Mesh Networks
and CDC values also reveals the interesting insight that the (WiMesh), 2005.
[16] J. So, N. H. Vaidya, “Load-Balancing Routing in Multichannel Hybrid
relationship between the channel diversity and the length of Wireless Networks With Single Network Interface,” in IEEE Trans. Veh.
the path in terms of hop count is not monotonic. For example, Technol., vol. 56, no. 1, pp. 342–348, Jan, 2007.
path III achieves better channel diversity than path I, being [17] T. Liu, W. Liao, “Capacity-Aware Routing in Multi-Channel Multi-Rate
Wireless Mesh Networks,” in Proc. IEEE ICC, 2006, pp. 1971–1976.
one-hop longer; but path I has better channel diversity than [18] K. Jain, J. Padhye, V. N. Padmanabhan, and L. Qiu, “Impact of
path IV, being one-hop shorter. Interference on Multi-hop Wireless Network Performance,” in ACM
MOBICOM,, 2003, pp. 66–80.
VII. C ONCLUSION [19] X. Li, H. Chen, Y. Shu and X.Chu, “Energy Efficient Routing With
Unreliable Links in Wireless Networks,” in Proc. IEEE International
In this paper, we aims at designing link/path metrics that Conference on Mobile Adhoc and Sensor Systems (MASS), 2006, pp.
can lead to path selection with the minimum end-to-end delay, 160–169.
while a high network throughput can also be achieved. The [20] H. Zhai, Y. Fang ”Impact of Routing Metrics on Path Capacity in Multi-
rate and Multi-hop Wireless Ad Hoc Networks,” in Proc. IEEE ICNP,
paper has key contributions in two aspects: 1) Based on the 2007, pp. 86–95.
concept of network/MAC cross-layer design, both the queuing [21] D. S. J. De Couto, D. Aguayo, J. Bicket, and R. Morris, “A High-
delay in network layer and transmission delay in the network Throughput Path Metric for Multi-Hop Wireless Routing,” in ACM
MOBICOM, 2003, pp. 134–142.
layer are included in the EED link metric computation; 2) [22] K. Kim, K. G. Shin, “On accurate measurement of link quality in multi-
A generic iterative approach is developed to compute the hop wireless mesh networks,” in ACM MOBICOM, 2006, pp. 38–49.
achievable bandwidth over a multi-radio multi-channel path, [23] H. Zhai, J. Wang and Y. Fang, “Distributed packet scheduling for
multihop flows in ad hoc networks,” in Proc. IEEE WCNC, 2004, pp.
which captures the complex interaction among hop count, 1081–1086.
channel assignment, and inter/intra flow interferences to form [24] Y. Xiao, K. Thulasiraman and G. Xue, “QoS routing in communication
the WEED path metric. A side benefit of our EED/WEED link networks: approximation algorithms based on the primal simplex method
of linear programming,” in IEEE Trans. Comput., vol. 55, no. 7, pp.
metric computation is a quantitative channel diversity coeffi- 815–829, July. 2006.
cient. We demonstrate the performance of EED/WEED based [25] H. Li, Y. Cheng, C. Zhou, “Multi-hop effective bandwidth based routing
routing via extensive numerical analysis and NS2 simulation in multi-radio wireless mesh networks,” in Proc. IEEE Globecom, 2008.
[26] J. Tang, G. Xue, and W. Zhang, “Interference-Aware Topology Control
results. and QoS Routing in Multi-Channel Wireless Mesh Networks,” in ACM
MobiHoc, 2005, pp. 68–77.
R EFERENCES [27] The Network Simulator - NS2, http://www.isi.edu/nsnam/ns/.
[1] P. Gupta and P. R. Kumar, “The cpacity of wireless networks,” IEEE [28] “Wireless Lan Medium Access Control (MAC) and Physical Layer
Trans. Inform. Theory, vol. 46, no. 2, pp. 388–404, Mar. 2000. (PHY) specifications,” ANSI/IEEE Std 802.11: 1999 (E) Part 11,
[2] M. Gastpar and M. Vetterli, “On the capacity of wireless networks: the ISO/IEC 8802-11, 1999.
relaycase,” in Proc. IEEE INFOCOM, 2002, pp. 1577–1586. [29] A. Raniwala, T.-c. Chiueh, ”Centralized Channel Assignment and
[3] A. E. Gamal, J. Mammen, B. Prabhakar, and D. Shah, “Throughput- Routing Algorithms for Multi-Channel Wireless Mesh Networks,” in
delay trade-off in wireless networks,” in Proc. IEEE INFOCOM, 2004, ACM SIGMOBILE Mobile Computing and Communications Review,
pp. 464–475. 2004, pp. 50–65.
[4] I. F. Akyildiz, X. Wang, and W. Wang, “Wireless mesh networks: a [30] A. Abdrabou and W. Zhuang, “Service time approximation in IEEE
survey,” Computer Networks., 2005, pp. 523–530. 802.11 single-hop ad hoc networks,” IEEE Trans. Wireless Commun.,
[5] G. Jakllari, S. Eidenbenz, N. Hengartner, S. V. Krishnamurthy, and vol. 7, no. 1, pp. 305-313, Jan. 2008.
M.Faloutsos, ”Link Positions Matter: A Noncommutative Routing
Metric for Wireless Mesh Network,” in Proc. IEEE INFOCOM, 2008,
pp.744–752.

Authorized licensed use limited to: Technische Informationsbibliothek (TIB). Downloaded on January 04,2024 at 14:55:41 UTC from IEEE Xplore. Restrictions apply.
54