A High-Throughput Path Metric for
Multi-Hop Wireless Routing
Douglas S. J. De Couto Daniel Aguayo John Bicket Robert Morris
M.I.T. Computer Science and Artificial Intelligence Laboratory
{
decouto, aguayo, jbicket, rtm}@csail.mit.edu
http://www.pdos.lcs.mit.edu/grid
Abstract
This paper presents the expected transmission count metric (ETX), which
finds high-throughput paths on multi-hop wireless networks. ETX mini-
mizes the expected total number of packet transmissions (including retrans-
missions) required to successfully deliver a packet to the ultimate destina-
tion. The ETX metric incorporates the effects of link loss ratios, asymmetry
in the loss ratios between the two directions of each link, and interference
among the successive links of a path. In contrast, the minimum hop-count
metric chooses arbitrarily among the different paths of the same minimum
length, regardless of the often large differences in throughput among those
paths, and ignoring the possibility that a longer path might offer higher
throughput.
This paper describes the design and implementation of ETX as a metric
for the DSDV and DSR routing protocols, as well as modifications to DSDV
and DSR which allow them to use ETX. Measurements taken from a 29-
node 802.11b test-bed demonstrate the poor performance of minimum hop-
count, illustrate the causes of that poor performance, and confirm that ETX
improves performance. For long paths the throughput improvement is often
a factor of two or more, suggesting that ETX will become more useful as
networks grow larger and paths become longer.
Categories and Subject Descriptors
C.2.1 [Computer-Communication Networks]: Network Archi-
tecture and Design—Wireless communication; C.2.2 [Computer-
Communication Networks]: Network Protocols—Routing proto-
cols
General Terms
Design, Experimentation, Measurement, Performance
Keywords
Multi-hop wireless networks, Ad hoc networks, Rooftop networks,
Wireless routing, Route metrics, 802.11b, DSR, DSDV, ETX
This research was supported by grants from NTT Corporation un-
der the NTT-MIT collaboration, and by MIT’s Project Oxygen.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
MobiCom ’03, September 14–19, 2003, San Diego, California, USA.
Copyright 2003 ACM 1-58113-753-2/03/0009 ...
$5.00.
1. Introduction
Much of the recent work in ad hoc routing protocols for wireless
networks [25, 15, 26] has focused on coping with mobile nodes,
rapidly changing topologies, and scalability. Less attention has
been paid to finding high-quality paths in the face of lossy wireless
links. This paper presents measurements of link loss characteris-
tics on a 29-node 802.11b test-bed, and uses these measurements
to motivate the design of a new metric which accounts for lossy
links: expected transmission count (ETX).
The metric most commonly used by existing ad hoc routing pro-
tocols is minimum hop-count. These protocols typically use only
links that deliver routing probe packets (query packets, as in DSR
or AODV, or routing updates, as in DSDV). This approach implic-
itly assumes that links either work well or don’t work at all. While
often true in wired networks, this is not a reasonable approximation
in the wireless case: many wireless links have intermediate loss ra-
tios. A link that delivers only 50% of packets may not be useful for
data, but might deliver enough routing update or query packets that
the routing protocol uses it anyway.
Minimizing the hop-count maximizes the distance traveled by
each hop, which is likely to minimize signal strength and maximize
the loss ratio. Even if the best route is a minimum hop-count route,
in a dense network there may be many routes of the same min-
imum length, with widely varying qualities; the arbitrary choice
made by most minimum hop-count metrics is not likely to select
the best. One contribution of this paper is to quantify these effects
(Section 2).
One approach to fixing this problem is to mask transmission er-
rors. For example, the 802.11b ACK mechanism resends lost pack-
ets, making all but the worst 802.11b links appear loss-free. How-
ever, retransmission does not make lossy links desirable for use
in paths: the retransmissions reduce path throughput and interfere
with other traffic. Another approach might be to augment minimum
hop-count routing with a threshold that ignores lossy links, but a
lossy link may be the only way to reach a certain node, and there
might be significant loss ratio differences even among the above-
threshold links.
The solution proposed and evaluated in this paper is the ETX
metric. ETX finds paths with the fewest expected number of trans-
missions (including retransmissions) required to deliver a packet
all the way to its destination. The metric predicts the number of re-
transmissions required using per-link measurements of packet loss
ratios in both directions of each wireless link. The primary goal
of the ETX design is to find paths with high throughput, despite
losses.
In order to demonstrate that ETX is effective, this paper presents
measurements taken from the test-bed network. These measure-
134
5
19
18
5
5
10
6
37
5
23
22
6
5
15
6
25
13
5
5
17
6
14
6
6
20
24
5
21
6
26
1
5
Approx. 79 m
Approx. 22 m
11
5
3
2
2
8
4
9
3
3
12
5
16
4
5
27
6
5
6
28
36
6
7
4
5
5
Figure 1: A map of the test-bed. Each circle is a node; the large number is the node ID, and the superscript indicates which floor of
the building the node is on.
ments show that ETX improves the throughput of multi-hop routes
by up to a factor of two over a minimum hop-count metric. ETX
provides the most improvement for paths with two or more hops,
suggesting that transmission count offers increased benefit as net-
works grow larger and paths become longer.
This paper makes the following main contributions. First, it ex-
plores the details of the performance of minimum hop-count rout-
ing on a wireless test-bed, and explains why minimum hop-count
often finds routes with significantly less throughput than the best
available. Second, it presents the design, implementation, and eval-
uation of theETX metric. Third, itdescribes a set of detailed design
changes to the DSDV [25] and DSR [15] protocols (to which ETX
is an extension), that enable them to more accurately choose routes
with the best metric.
This work is part of an effort to deploy a production-quality
multi-hop rooftop 802.11b network. The initial version of that net-
work was almost unusable due to the effects detailed in Section 2.
The larger goal of this work is to help make such networks a prac-
tical reality.
The paper proceeds in Section 2 with an analysis of the problems
with minimum hop-count routing. Section 3 describes the design
of the new ETX metric, and Section 4 describes how ETX is imple-
mented, including changes to DSDV and DSR. Section 5 evaluates
ETX using experiments on the test-bed. Section 6 describes related
work, and Section 7 concludes the paper.
2. Performance of Minimum-Hop-Count
Routing
This section shows that minimum hop-count routing typically
finds routes with significantly lower throughput than the best avail-
able. The evidence comes from measurements of DSDV on a test-
bed network. We explain why minimum hop-count does poorly by
looking at the distribution of route throughputs and link loss ratios.
2.1 Experimental Test-Bed
All the data in this paper are the result of measurements taken
on a 29-node wireless test-bed. Each node consists of a stationary
Linux PC with a Cisco/Aironet 340 PCI 802.11b card and an omni-
directional 2.2 dBi dipole antenna (a “rubber duck”). The nodes are
placed in offices on five consecutive floors of an office building.
Their positions are shown in Figure 1.
The 802.11b cards are configured to send at one megabit per sec-
ond (Mbps) with one milliwatt (mW) of transmit power. RTS/CTS
is turned off, and the cards are set to “ad hoc” (IBSS, DCF) mode.
Each data packet in the following measurements consists of 24
bytes of 802.11b preamble, 31 bytes of 802.11b and Ethernet en-
capsulation header, 134 bytes of data payload, and 4 bytes of frame
check sequence: 193 bytes in total. An 802.11b ACK packet takes
304 microseconds to transmit, the inter-frame gap is 60 microsec-
onds, and the minimum expected mandatory back-off time is 310
microseconds, resulting in a total time of 2,218 microseconds per
data packet. This gives a maximum throughput of 451 unicast pack-
ets per second over a loss-free link.
While the test-bed itself carried only the data and control traffic
involved in each experiment, interference of various kinds was in-
evitably present. In particular, each floor of the building has four
802.11b access points, on various different channels.
The DSDV implementation used in this paper is new, with mod-
ifications described in Section 4.
2.2 Path Throughputs
Figure 2 compares the throughput of routes found with a min-
imum hop-count metric to the throughput of the best routes that
could be found. Each curve shows the throughput CDF (in pack-
ets per second) for 100 node pairs; the pairs are randomly selected
from the 29×28 = 812 total ordered pairs in the test-bed. A point’s
x value indicates throughput, in packets per second; the y value in-
dicates what fraction of pairs had less throughput. The left curve
is the throughput CDF achieved by routing data using DSDV with
the minimum hop-count metric. The right curve is the throughput
CDF for the best known route between each pair of nodes. Packets
were only sent between one pair at a time. For each pair, the DSDV
and best-path tests were run immediately after one another, to limit
variation in link conditions over time.
The “best” path between each pair of nodes was found by send-
ing data along ten potential best paths, one at a time, and select-
ing the path with the highest throughput. Potential best paths were
identified by running an off-line routing algorithm, using as input
measurements of per-link loss ratios, similar to those in Section 2.4,
135
0
0.2
0.4
0.6
0.8
1
0 50 100 150 200 250 300 350 400 450
Cumulative fraction of node pairs
Packets per second delivered
Run R1: 1 mW, 134-byte packets
Max 4-hop
throughput
2-hop3-hop
Best static route
DSDV hopcount
Figure 2: When using the minimum hop-count metric, DSDV
chooses paths with far less throughput than the best available
routes. Each line is a throughput CDF for the same 100 ran-
domly selected node pairs. The left curve is the throughput
CDF of DSDV with minimum hop-count. The right curve is
the CDF of the best throughput between each pair, found by
trying a number of promising paths. The dotted vertical lines
mark the theoretical maximum throughput of routes of each
hop-count.
and with a penalty to reflect the reduction in throughput caused by
interference between successive hops of multi-hop paths. New link
measurements were collected roughly every hour during the exper-
iment; the best paths for each pair were generated using the most
recently available loss data.
The values in Figure 2 are split into two main ranges, above and
below 225 packets per second. The values above 225 correspond to
pairs that communicated along single-hop paths; those at or below
225 correspond to multi-hop paths. A single-hop direct route can
deliver up to about 450 packets per second, but the fastest two-hop
route has only half that capacity. The halving is due to transmis-
sions on the successive hops interfering with each other: the middle
node cannot receive a packet from the first node at the same time
it is sending a packet to the final node. Similar effects cause the
fastest three-hop route to have a capacity of about 450/3 = 150
packets per second.
Minimum hop-count performs well whenever the shortest route
is also the fastest route, especially when there is a one-hop link with
a low loss ratio. A one-hop link with a loss ratio of less than 50%
will outperform any other route. This is the case for all the points
in the right half of Figure 2. Note that the overhead of DSDV route
advertisements reduces the maximum link capacity by about 15 to
25 packets per second, which is clearly visible in this part of the
graph.
The left half of the graph shows what happens when minimum
hop-count has a choice among a number of multi-hop routes. In
these cases, the hop-count metric usually picks a route significantly
slower than the best known. The most extreme cases are the points
at the far left, in which minimum hop-count is getting a through-
put close to zero, and the best known route has a throughput of
0
50
100
150
200
23-19-24-36
23-37-24-36
23-37-19-36
23-12-19-36
23-19-11-36
23-19-36
23-19-20-36
23-19-7-36
Packets per second delivered
Run R1: 1 mW, 134-byte packets
Max 3-hop throughput
Max 4-hop
Figure 3: Throughput available between one pair of nodes, 23
and 36, along the best eight routes tested. The shortest of the
routes does not perform the best, and there are a number of
routes with the same number of hops that provide very differ-
ent throughput.
about 100 packets per second. The minimum hop-count routes are
slow because they include links with high loss ratios, which cause
bandwidth to be consumed by retransmissions.
2.3 Distribution of Path Throughputs
Figure 3 illustrates a typical case in which minimum hop-count
routing would not favor the highest-throughput route. The through-
put of eight routes from node 23 to node 36 is shown. The routes
are the eight best which were tested in the experiments described
above.
The graph shows that the shortest path, a two-hop route through
node 19, does not yield the highest throughput. The best route
is three hops long, but there are a number of available three-hop
routes which provide widely varying performance.
A routing protocol that selects randomly from the shortest hop-
count routes is unlikely to make the best choice, particularly as the
network grows and the number of possible paths between a given
pair increases.
2.4 Distribution of Link Loss Ratios
Figure 4 helps explain why high-throughput paths are difficult to
find. Each vertical bar corresponds to the direct radio link between
a pair of nodes; the two ends of the bar mark the broadcast packet
delivery ratio in the two directions between the nodes. To measure
delivery ratios, each node took a turn sending a series of broadcast
packets for five seconds, and counted the number of packets that
the 802.11b hardware reported as transmitted. Packets contained
134 bytes of 802.11b data payload. Every other node recorded the
number of packets received. The delivery ratio from node X to each
node Y is calculated by dividing the number of packets received by
Y by the number sent by X. The loss ratio of a link is one minus
its delivery ratio. We use the term “ratio” instead of “rate” to avoid
confusion with throughput delivery rates, which are expressed in
packets per second.
Note that 802.11b broadcasts don’t involve acknowledgements
or retransmissions. Because 802.11b retransmits lost unicast pack-
ets, the unicast packet loss ratio as seen by higher layers is far lower
than the underlying loss ratio (depending on the maximum number
of retransmissions allowed).
Three features of Figure 4 are important. First, a large fraction
of the links have an intermediate delivery ratio in at least one di-
rection. That is, they are likely to deliver some routing protocol
136
0
0.25
0.5
0.75
1
0 100 200 300 400
Delivery Ratio
Link number
(b) Pairwise delivery ratios at 30 mW
0
0.25
0.5
0.75
1
0 100 200 300 400
Delivery Ratio
Link number
(a) Pairwise delivery ratios at 1 mW
Figure 4: One-hop packet delivery ratios between each pair of hosts at 1 mW (above) and 30 mW (below). The top and bottom ends
of each vertical line indicate the delivery ratios in the two directions; the bars in each graph are sorted by the minimum of the two
directions, so the link numbers do not necessarily correspond. The packet size is 134 bytes of 802.11b data payload. Data for all 406
pairs of hosts are shown. Many links are asymmetric, and there is a wide range of loss ratios.
packets, but would lose many packets if used for data. Second,
there is a full spectrum of link delivery ratios, so some advantage
can be expected from making fine-grained choices between links
when choosing paths. Third, many links have asymmetric delivery
ratios.
Of the 406 node pairs in Figure 4a (1 mW), there are 124 with
links which delivered packets in at least one direction. Of those
links, 28 are asymmetric, with forward and reverse delivery ratios
that differ by at least 25%. The 28 asymmetric links involve 22
different nodes. Because 802.11b uses link-level ACKs to confirm
delivery, both directions of a link must work well in order to avoid
retransmissions. Since most nodes in the network are involved in at
least one asymmetric link, routing protocols must cope with asym-
metry to be effective.
3. ETX Metric Design
This section describes the design of the ETX metric. The met-
ric’s overall goal is to choose routes with high end-to-end through-
put. Section 2 suggests that the metric must account for the follow-
ing issues:
• The wide range of link loss ratios.
• The existence of links with asymmetric loss ratios.
• The interference between successive hops of multi-hop paths.
A number of superficially attractive metrics are not suitable. Us-
ing hop-count as the metric while ignoring links with loss ratios
above a certain threshold may cause some destinations to be un-
reachable. Using the product of the per-link delivery ratios as the
path metric, in an attempt to maximize the end-to-end delivery
probability, fails to account for inter-hop interference; this metric
would view a perfect two-hop route as better than a one-hop route
with a 10% loss ratio, when in fact the latter would have almost
twice the throughput. The same objection applies to using the use-
ful throughput of a path’s bottleneck (highest-loss-ratio) link as the
path’s metric. ETX, however, addresses each of these concerns.
End-to-end delay is another potential metric, but changes with
network load as interface queue lengths vary; this can cause routes
to oscillate away from a good path once the path isused. Our goal is
to design a metric that is independent of network load; load balanc-
ing can be performed with separate algorithms that use the infor-
mation provided by ETX. We have implemented ETX as a metric
for the DSDV and DSR routing protocols.
3.1 The Metric
The ETX of a link is the predicted number of data transmissions
required to send a packet over that link, including retransmissions.
The ETX of a route is the sum of the ETX for each link in the
route. For example, the ETX of a three-hop route with perfect links
is three; the ETX of a one-hop route with a 50% delivery ratio is
two.
The ETX of a link is calculated using the forward and reverse
delivery ratios of the link. The forward delivery ratio, d
f
, is the
measured probability that a data packet successfully arrives at the
recipient; the reverse delivery ratio, d
r
, is the probability that the
ACK packet is successfully received. These delivery ratios can
be measured as described below. The expected probability that a
transmission is successfully received and acknowledged is d
f
× d
r
.
A sender will retransmit a packet that is not successfully acknowl-
edged. Because each attempt to transmit a packet can be considered
a Bernoulli trial, the expected number of transmissions is:
ETX =
1
d
f
× d
r
(1)
137
ETX has several important characteristics:
• ETX is based on delivery ratios, which directly affect through-
put.
• ETX detects and appropriately handles asymmetry by incor-
porating loss ratios in each direction.
• ETX can use precise link loss ratio measurements to make
fine-grained decisions between routes.
• ETX penalizes routes with more hops, which have lower
throughput due to interference between different hops of the
same path [20].
• ETX tends to minimize spectrum use, which should maxi-
mize overall system capacity.
In addition, ETX may decrease the energy consumed per packet,
as each transmission or retransmission may increase a node’s en-
ergy consumption.
The delivery ratios d
f
and d
r
are measured using dedicated link
probe packets. Each node broadcasts link probes of a fixed size, at
an average period τ (one second in the implementation). To avoid
accidental synchronization, τ is jittered by up to ±0.1τ per probe.
Because the probes are broadcast, 802.11b does not acknowledge
or retransmit them. Every node remembers the probes it receives
during the last w seconds (ten seconds in our implementation), al-
lowing it to calculate the delivery ratio from the sender at any time
t as:
r( t)=
count(t − w, t)
w/τ
Count(t − w, t) is the number of probes received during the win-
dow w, and w/τ is the number of probes that should have been
received. In the case of the link X → Y , this technique allows X
to measure d
r
, and Y to measure d
f
. Because Y knows it should
receive a probe from X every τ seconds, Y can correctly calculate
the current loss ratio even if no probes arrive from X.
Calculation of a link’s ETX requires both d
f
and d
r
. Each probe
sent by a node X contains the number of probe packets received by
X from each of its neighbors during the last w seconds. This allows
each neighbor to calculate the d
f
to X whenever it receives a probe
from X.
The ETX of a route is the sum of the link metrics. DSR and
DSDV accumulate the route metric as they forward updates and
queries, respectively.
3.2 Discussion
ETX makes at least two assumptions about the link layer. First,
ETX only makes sense for networks with link-layer retransmis-
sion, such as 802.11b. Second, ETX assumes that radios have a
fixed transmit power level. With variable power radios, it might be
preferable to maximize hop-count, thereby decreasing interference
and minimizing the energy used by each packet [29, 12, 18].
ETX does not attempt to route around congested links, and thus
should not suffer from the oscillations that sometimes plague load-
adaptive routing metrics such as end-to-end delay. To a first ap-
proximation, the loss measurements used by ETX do not reflect
how busy a link is; a busy link may cause a probe broadcast to be
deferred, but won’t ordinarily cause it to be lost. This won’t always
be true, however, since 802.11b broadcasts are vulnerable to colli-
sions from hidden terminals, and the 802.11b MAC can be unfair
under high load [4]. As a result, a node might never be able to send
its probes, causing its neighbors to believe that the reverse delivery
ratio had become zero.
If the highest-throughput path has three or fewer hops, ETX is
likely to choose it: the throughput of such paths is determined by
the total number of transmissions, since all of the hops interfere
with each other [20]. If the best path has four or more hops, ETX
may choose a slower path that has fewer hops, since the increased
number of transmissions required by extra hops does not slowdown
throughput beyond three hops.
ETX does not specifically account for mobility. ETX may choose
good paths despite mobility if the underlying routing protocol can
propagate route metrics quickly enough, and if accurate link mea-
surements are available. There isa tradeoff between the accuracy of
link measurements and the protocol’s responsiveness to mobility.
4. Implementation
The routing system in which ETX is implemented has four main
parts: the Click toolkit [19], and Click-based implementations of
DSDV, DSR, and the ETX link measurement algorithms. The im-
plementations can run in user-space, but running in the kernel al-
lows use of the priority queuing described below, as well as easy
access to transmission failure notification from the 802.11 MAC
layer.
The DSDV protocol is implemented following the description
by Perkins and Bhagwat [25], with ambiguities resolved by con-
sulting Broch et al. [5] and the Rice/CMU implementation in the
ns simulator [1, 27]. The DSR implementation follows the IETF
Internet-Draft, version 9 [16].
4.1 Operation of DSDV
DSDV is a distance-vector protocol, which uses sequence num-
bers to ensure freshness, and a settling time mechanism to avoid
unnecessarily propagating routes with inferior metrics. We made
four changes to the original DSDV design in order to ensure that it
uses the path with the best known metric. Before describing those
changes, we present an overview of how the published version of
the protocol selects routes.
Every node has a routing table entry for each destination it knows
about. This entry contains four fields: the destination’s identifier
(IP address), the next hop on the route to that destination, the latest
sequence number heard for that destination, and the metric. A node
forwards packets to the next hop specified by the current contents
of its routing table.
Every node periodically broadcasts a route advertisement packet
containing its complete routing table. This advertisement is known
as a full dump, and occurs at the full dump period.
Each node maintains asequence number for itself, which it incre-
ments and includes in its own entry in every full dump it originates.
The node copies the sequence numbers for the other entries in the
full dump from its routing table. The effect is that the sequence
number field in a routing table entry or advertisement entry reflects
the age of that entry’s routing information.
When a node receives another node’s route advertisement broad-
cast, it updates its own route entries as follows. Suppose node X
receives an advertisement from Y for destination D with metric m
and sequence number n.Ifn is newer than the sequence number
in X’s current entry for D, X replaces its current entry with the
new route through Y . X also accepts the new route if the sequence
number is the same, but m is better than the metric of the current
route. If X has no route to D, it accepts the new route. Otherwise X
ignores the advertised route.
Each route entry has an associated weighted settling time (WST).
The settling time of a route entry with a given sequence number is
138