an enhanced available bandwidth estimation technique for an end to end network path

1932-4537 (c) 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. Seehttp://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSM.2016.2572212, IEEETransactions on Network and Service Management

IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. , NO. , 2015 1

An Enhanced Available Bandwidth EstimationTechnique for an End-to-End Network Path

Anup Kumar Paul, Member, IEEE, Atsuo Tachibana, and Teruyuki Hasegawa

Abstract—This paper presents a unique probing scheme, a rateadjustment algorithm, and a modified excursion detection algo-rithm (EDA) for estimating the available bandwidth (ABW) ofan end-to-end network path more accurately and less intrusively.The proposed algorithm is based on the well known concept ofself-induced congestion and it features a unique probing trainstructure in which there is a region where packets are sampledmore frequently than in other regions. This high-density regionenables our algorithm to find the turning point more accurately.When the dynamic ABW is outside of this region, we readjust thelower rate and upper rate of the packet stream to fit the dynamicABW into that region. We appropriately adjust the range betweenthe lower rate and the upper rate using spread factors, whichenables us to keep the number of packets low and we are thus ableto measure the ABW less intrusively. Finally, to detect the ABWfrom the one-way queuing delay, we present a modified EDA fromPathChirps’ original EDA to better deal with sudden increaseand decrease in queuing delays due to cross traffic burstiness.For the experiments, an Android OS-based device was used tomeasure the ABW over a commercial 4G/LTE mobile network ofa Japanese mobile operator, as well as real testbed measurementswere conducted over fixed and WLAN network. Simulations andexperimental results show that our algorithm can achieve ABWestimations in real time and outperforms other stat-of-the-artmeasurement algorithms in terms of accuracy, intrusiveness, andconvergence time.

Index Terms—Available Bandwidth, Probe Rate Model, Queu-ing Delay, Rate Adjustment, Modified Excursion Detection Algo-rithm, 4G/LTE Network.

I. INTRODUCTION

Available bandwidth (ABW) estimation is crucial for traffic

engineering, quality-of-service (QoS) management, multime-

dia streaming, server selection in application services, con-

gestion management, and network capacity provisioning in

wireless mobile networks. ABW measurement can be con-

sidered essential to ensure that wireless mobile operators can

achieve the QoS standard guaranteed by them while providing

desired data rates to users. This can also be considered

when comparing the performance index of various Telecom

operators in a specific region.Let us first define the terms ABW, bottleneck link (BL)

or narrow link,1 and tight link precisely. Consider an end-to-

end path that includes n links L1, L2, · · · , Ln. Their capac-

ities are B1, B2, · · · , Bn and the traffic loads on these links

are C1, C2, · · · , Cn respectively. The BL can be defined as

Lb(1 ≤ b ≤ n), where

Bb = min(B1, B2, · · · , Bn).

Anup Kumar Paul, Atsuo Tachibana and Teruyuki Hasegawa are with theKDDI R&D Laboratories Inc., Fujimino, Saitama, 356-8502 JAPAN; e-mail:([email protected],[email protected],[email protected]).

1We will use both the terms interchangeably throughout this paper.

The tight link can be defined as Lt(1 ≤ t ≤ n), where

Bt − Ct = min(B1 − C1, B2 − C2, · · · , Bn − Cn).

The unused bandwidth on the tight link, Bt−Ct is called the

ABW of the path. There could be different possible definitions

of ABW, depending on whether we use an approach based

on unused capacity [1] or an approach based on achievable

rate [2]. In wired networks, these two approaches are equiva-

lent, leading to the widely accepted definition of ABW as the

unused capacity of the tight link. But in wireless networks,

interference makes the two concepts quite different. Due to

radio interference, the unused capacity may not be completely

available. On the other hand, when a new flow is established

in the given path to occupy some of that unused capacity, the

interfering cross traffic can re-accommodate itself in response

to the new flow, changing the perception of the new regarding

its ABW [3]. Since, in wireless settings, the unused capacity

approach does not take into account this possible adaptation to

network conditions, in this paper, we use the achievable rate

approach.

Ideally, a probing scheme should provide an accurate mea-

surement of ABW while requiring less time and imposing as

light a load as possible [1]. ABW estimation tools add traffic

to the network path under measurement. This may adversely

affect application traffic and measurement accuracy [4]. The

amount of probe traffic is proportional to the rate of sampling

and the number of concurrent measurement sessions. As a

result, the effect of probe packets on cross traffic exacerbates

with traffic increase. Thus less intrusive approach is desirable.

Therefore researchers have developed several end-to-end ABW

estimation techniques that infer the network characteristics

by transmitting a few packets and observing the effects of

intermediate routers or links on these probing packets. Nev-

ertheless, there are a variety of challenges that we need to

take into account: the estimation technique should be accurate,

non-intrusive, and robust at the same time. Moreover, the

estimation technique should be adaptively applied in different

types of networks and cross traffic (CT) and must be able to

produce accurate periodic estimations in a reasonable amount

of time in order to track bandwidth fluctuations. Therefore, as

mentioned in [5][6], the current ABW estimation techniques

are far from being ready to be applied to many applications

and scenarios.

Measuring end-to-end ABW in a LTE network can be a

challenging task due to varying wireless channel conditions,

scheduling and modulation techniques, and pre-configured

QoS parameters as well as the requirement for dedicated

hardware and underlying operating system (OS) at both the

www.redpel.com +917620593389





Fig. 1. One-way queuing delay signature.

end measurement points. The limitations to accurate network

performance measurements may be the air interface as well

as the transport network. In real networks, the CT patterns

become highly bursty, and this results in a significant deviation

from the preferred fluid model of CT. Our proposed method

determines the ABW by analyzing the one-way queuing delay

curve (Fig. 1), also called the rate-response curve in [7].

In [7], it is shown that fluid-like analysis can give erroneous

bandwidth estimates when taking packet-level interactions in

the router queues into account, especially if the CT is bursty

(e.g., Pareto distributed) or if there are several secondary

bottlenecks [8]. By using longer probe-packet trains instead of

probe-packet pairs, the obtained rate-response curve asymptot-

ically moves towards the fluid curve [7]. In our experiments,

we used long probe-packet trains. Furthermore, we studied

the effect of one wireless bottleneck link on the rate-response

curve. Thus, for the objectives in this paper, we believe that

the original fluid model is sufficient.

By taking into account the various challenges mentioned

above, our goal is to estimate the ABW with good accuracy

and less intrusively, i.e., without interrupting other network

traffic to the extent possible. The main contributions of this

paper are as follows:

• To detect the ABW from the one-way queuing delay

curve, we have proposed a modified excursion detection

algorithm. Due to the bursty arrival of CT, a sudden

increase in queuing delays of packets (called excursion)

occur for a short period of time in the router even though

the packet’s rate is much below the ABW of the tight link.

So, to detect these sudden increases in queuing delays and

to filter out them is the main purpose of the excursion

detection algorithm.

• We have proposed the packet loss recovery algorithm to

make our algorithm more robust.

• We have measured, analyzed, and described how the

characteristics of our proposed method act in wireless

scenarios.

• We have conducted a real test over a commercial 4G/LTE

network of a Japanese mobile operator as well as in a

real fixed and Wireless Local Area Network (WLAN) to

evaluate the effectiveness of our proposed method.

The extended idea from our previous work NEXT [9] enables

to find the turning point2 more accurately.

The rest of the paper is organized as follows. Related work

is discussed in Sect. II. Our proposed algorithm is presented

in Sect. III. Simulation and real testbed results are shown and

ABW measurement performances are discussed in Sect. IV.

Finally, our conclusions presented in Sect. V.

II. RELATED WORK

Available bandwidth measurement techniques can be classi-

fied into two broad categories [10][11], one being passive esti-

mation and the other being active probing. Passive estimation

is done on the basis of congestion situation, packet loss, and

delay performance to estimate the ABW. Active probing on

the other hand sends probe-packets over a network to estimate

the ABW. Due to the efficiency and reliability of estimations,

active probing is usually considered. Active probing further

consists of the Probe Gap Model (PGM) [12] and Probe Rate

Model (PRM) [13].

The PGM generates an estimate of the ABW by estimation

of the CT rate in the link. Tools developed following this

technique require previous knowledge of the capacity of the

network path to be measured. The working behind this probing

technique is where the sender sends a pair of packets to

the receiver. The pair packets are transmitted close enough

together in time for packets to queue together at the BL.

Measuring the change in packet spacing, the receiver can make

an estimate of the amount of CT during the measurement

time in the bottleneck link [12] and then compute the ABW

as the difference between the BL capacity and the CT rate:

ABW = Bb − Ct. Examples of algorithms in this category

are Spruce [14] and IGI [15].

On the other hand, the PRM techniques operate on the

basis of self-induced congestion where this mechanism sends

a stream of packets where the input rate of each stream is

varied either iteratively or exponentially. Cross traffic follows

a fluid model and average rates of CT change slowly. If a

source sends probes to a destination at rate R less than the

ABW, probes will experience similar delays. On the other

hand, if R is greater than the ABW, probes will queue in the

network and experience increasing delays. This technique is

based on the observation that the delays of successive probing

packets increase when the probing rate exceeds the ABW in

the path. It consists in probing the network at different rates

and detecting (at the destination) the point at which delays

start to increase. At this point, the probing rates are equal to

the ABW. The PRM model has proved to be accurate and

it is used in many estimation tools, such as PathChirp [1],

TOPP [12], Pathload [13], etc.

It is evident from the research that the structure of the

probing packet sequence is a major player in estimating the

ABW. This is why several probing packet structures have

been proposed in various approaches. For example, the single

packet concept is used by simple protocols such as ping

and traceroute. Packet pairs are used by Spruce [14]. The

2A turning point is the point from where all other successive packets queu-ing delay shows an increasing trend until the last packet. The correspondingpacket’s rate is treated as the ABW.






packet train structure is used by TOPP [12] and Pathload [13].

TOPP and Pathload use a constant bit rate stream, sending

pairs of trains of packets at a given rate and changing this

rate every round. TOPP linearly increases the sending rate

in successive streams, trying to find the exact turning point.

Pathload on the other hand varies the probing rate using

a binary search scheme and the final output, the result of

multiple measurements, is a variation range rather than a

single estimate. Since multiple trains are required to produce

a single estimation, the intrusiveness of these techniques is

quite high and the measurement process is time-consuming.

Packet chirp, another type of packet train structure, is used by

PathChirp [1]. The difference among these algorithms is their

probing structure methods, which are based on the number of

packets and the spacing between the packets. Packet spacing

values can be either fixed in size [13] or follow the exponential

distribution [1].

PathChirp [1] sends a variable bit-rate stream called chirp,

which consists of exponentially spaced packets with rates gen-

erated by the equation r = L×γi, ∀r ≤ H , where L and H are

the lower rate and the upper rate respectively, γ is the spread

factor, and i = {1, 2, . . .}. Whenever the estimated ABW

feedback by the receiver is close to L or H , PathChirp adjusts

the new L and H heuristically. This heuristic rate adjustment is

inefficient and results in inaccurate ABW estimation in many

cases. A detailed description of PathChirp’s rate adjustment

algorithm can be found in their ns-2 [16] simulation code or

real network implementation code [1]. Although PathChirp has

several problems, after extensive empirical evaluation by [17],

the author has come to the conclusion that among various

ABW estimation tools, PathChirp has the lowest overhead and

comparably good accuracy in multihop network paths with

different CT and multiple tight links. Even when PathChirp

runs continuously on a given path, it has virtually no impact

on the response times of TCP connections sharing the same

path.

PathChirp can estimate the ABW by sending multiple

chirps, and for each chirp, the receiver sends back the esti-

mated ABW. Apart from several advantages, PathChirp has

two major drawbacks. First, the packets are all exponentially

spaced and second, the rate adjustment algorithm is not

optimum. The first drawback lies in the probe-packet train

structure. Starting from the L, the packets are closely spaced

but as we go towards the H , the packet density significantly

decreases. In PRM techniques, the mechanism for estimating

the ABW is self-congestion and the ABW is determined by

detecting the turning point at which the queuing delay starts

increasing for all successive packets, i.e., the rate at which the

packets start facing queuing delays, and the previous packet

rate is the expected ABW. The problem arises when the

sampling rate of the previous packet is sparsely spaced from

the packet rate at which the queuing delay starts increasing.

In other words, PathChirp samples the lower rates more

frequently than the higher rates. Therefore, the tool is less

accurate if the actual ABW is not located near L. As an

example, for a spread factor of 1.2, if a chirp uses L as 1Mbps and the H as 100 Mbps, then the total number of packets

generated by PathChirp is 26. The first 13 packet rates are all

0 5 10 15 20 25 300

10

20

30

40

50

60

70

80

90

100

Number of packets

Pack

et R

ate

in M

bps

PathChirp

Fig. 2. PathChirp’s rate generation.

less than 10 Mbps, while the rest of the 13 packet rates lie in

the range from 10 Mbps to 100 Mbps (Fig. 2).

An alternative methodology to the active end-to-end

approach would be to query every network element

(switch/router) along a network path. A network administrator

could collect the statistical counters from all related ports, via

protocols such as sFlow [18]; or infer from Openflow control

messages such as PacketIn and FlowRemoved messages [19].

However, obtaining an accurate, consistent, and timely reading

from multiple switches in an end-to end manner can be very

difficult [20]. Further, collecting counters from switches is

done at per-second granularity basis and requires network ad-

ministrative privileges, which makes the approach less timely

and useful for building distributed systems, improving network

protocols, or improving application performance. Another art

of technique, such as MinProbe [21] used application traffic

implicitly as available bandwidth probes, and are able to

remove all the traditional costs and overheads. A similar idea

was proposed in MGRP [4], which has the same goal of

lower overhead. However, in both cases, one problem may

occur when a TCP cwnd (congestion window) is small, e.g.,

when cwnd = 1. In this case, the application does not create

enough traffic and dummy probe needs to be created. As a

result, accurate and less intrusive active ABW measurement

technique is important [22][23].

Recent researches include ABW measurement specific to

wireless networks [2][24][25][26][27][28]. Exact [24] and

IdleGap [25] assume that the RTS/CTS is always enabled

and present only the simulation results. CapProbe [26] tries to

avoid the influence of cross traffic by only estimating capacity.

ProbeGap [27] measures the ABW in WLANs indirectly from

the idle time fraction using one-way delay samples over

the wireless link, but requires third-party capacity estimation

tools. DietTOPP [28] uses a reduced TOPP algorithm with a

modified search algorithm to determine the ABW in wireless

networks. WBest [2] measures the ABW in two stages. In the

first stage, the packet pair technique estimates the effective

capacity over a flow path where the last hop is a wireless

LAN. In the second stage, a packet train technique estimates

achievable throughput to infer the ABW.

Several comparisons of ABW measurement tools have

been made using simulation and measurement studies [14],






[17], [29]. However, none of these active measurement tools

considers measurement over a 4G/LTE mobile network. A

recent work [30] developed a passive measurement tool for

ABW estimation over a 4G/LTE network. Our paper reports

the first study where the performance of an active ABW

measurement tool is investigated over a 4G/LTE commercial

mobile network.

III. THE PROPOSED ALGORITHM

Our proposed idea is divided into four sub-ideas:

1. Packet Generation Algorithm

2. Rate Adjustment Algorithm

3. Modified Excursion Detection Algorithm

4. Packet Loss Recovery Algorithm

A preliminary version of our proposed idea, New Enhanced

Available Bandwidth Measurement Technique (NEXT), has

appeared in [9]. In NEXT, we proposed the first two sub-

ideas (1 and 2). Our extended version of NEXT includes the

modified excursion detection algorithm and the packet loss

recovery algorithm, and we refer to it as NEXT-V2 throughout

this paper.

A. Packet Generation Algorithm

Our proposed idea, NEXT, estimates the ABW along a net-

work path by launching a number of packet chirps (numbered

c = 1, 2, 3 · · · ) from sender to receiver. Each chirp consists

of a certain number of packets that depend on the lower rate,

the upper rate, and the spread factor. Assume that chirp cconsists of N packets. The ratio of successive packet inter-

spacing times within a chirp is defined as the spread factor.

In NEXT, we have two spread factors: γ1 and γ2. First, we

divided the total range of rates into three regions in terms of

the number of packets. NEXT used N number of packets to

test the values between the lower rate L and the upper rate H .

We have divided the entire range of rates from L to H into

three portions, where P1 and P2 are two region intersecting

points.

We set the initial rate as R0 = L. Successive rates will be

generated according to whether the current rate falls within

the P1 to P2 region or not. If the current rate Ri falls within

P1 to P2, then the next rate is calculated as Ri = Ri−1 × γ2.

Otherwise, the rate will be Ri = Ri−1 × γ1. We used spread

factor γ2 = 1.1 from the P1 to P2 region of the total range

of rates. In other areas, we used the spread factor γ1 = 1.2(adapted from PathChirp’s experimental evaluation). NEXT is

based on the observation that the smaller the sample intervals

around the turning point, the more accurately we can find the

ABW at that turning point. NEXT is more accurate if the

ABW falls in the region from one-third to two-thirds of the

stream.3 The larger the number of packets in a small range of

rates i.e., the higher the sampling density around the turning

point, the more quickly the queue will build up and so the

accuracy will be precisely detected. Compared to PathChirp,

3When the ABW is not inside that region, we readjust the low rate andhigh rate in such a way that the ABW fits into that region in the followinground of measurement.

our proposed scheme NEXT is different—both algorithms use

a sequence of packets of increasing delays, but the shape of

the traffic and the spacing between packets, especially from

the P1 to P2 portion of the total range, are not the same.

Readers interested more about the details of the algorithm’s

pseudo-code can refer to [9].

Here we will additionally describe how quickly our algo-

rithm measures a single estimate of the ABW. In order to do

that, it is necessary to formulate the total chirp duration of a

single chirp. Let Sp be the size of a single packet and tp be

the time duration of a single packet in a packet train (chirp).

We know that

tp =SP

chirpRate

Here chirpRate is the number of bits sent per second. Packet

sizes are considered in computing the chirp rate. Since we have

three different regions in a single packet train, we calculate the

time duration of three separate portions and then add them

together to obtain the total chirp duration.

Thus, the total duration up to P1 is calculated as follows:

tP1p =

Sp

L+

Sp

L× γ1+

Sp

L× γ21

+ · · ·+ Sp

L× γi−11

tP1p =

Sp

L

⎛⎜⎜⎝1−

(1γ1

) logP1−logL

logγ1

1− 1γ1

⎞⎟⎟⎠ (1)

where, i = logP1−logLlogγ1

(from Eq.2 in [9]).

In a similar way, we can obtain the chirp duration from P1

to P2 as

tP2p =

Sp

P1

⎛⎜⎜⎝1−

(1γ2

) logP2−logP1logγ2

1− 1γ2

⎞⎟⎟⎠ (2)

and from P2 to H as

tHp =Sp

P2

⎛⎜⎜⎝1−

(1γ1

) logH−logP2logγ1

1− 1γ1

⎞⎟⎟⎠ (3)

Therefore, total chirp duration T chirpNEXT can be expressed as

the sum of tP1p , tP2

p , and tHp .

T chirpNEXT = tP1

p + tP2p + tHp (4)

B. Rate Adjustment Algorithm

Rate adjustments of L and H are performed if the feedback

value from the NEXT receiver to the NEXT sender is close to

L or H , and whenever the feedback value does not fit into the

P1 to P2 region. Details of the algorithms pseudo-code can be

found in [9]. For the ease of understanding, we will describe

our rate adjustment algorithm pictorially in this paper.

Figure 3 describes our rate adjustment algorithm. The

horizontal line represents the rate of packets where L is the

lower rate and H is the higher rate. The individual red dots

represent the packets with different rates. For comprehensive






Fig. 3. Rate adjustment of NEXT.

understanding, the spacing’s between the red dots are set equal

here but this does not mean that the rate intervals among

these red dots (packets) are equal. The consecutive actions for

any particular conditions are going from top to bottom and

are indicated by line 1, 2, · · · , 6, where line 2 and line 3 are

derived from line 1, and line 5 is derived from line 4, and line

6 is derived from line 5. The rate adjustment algorithm stops

adjusting rates when the condition of P1 < ABWinst < P2 is

satisfied, where ABWinst is the instantaneous ABW feedback

by the receiver to the sender in a single round of measurement.

This can be seen from the last horizontal line (line 6) of the

figure.

Rate adjustments of L and H of each chirp are conducted

if the feedback value (ABWinst) from the receiver is close to

L or H and even when ABWinst does not fit into the P1 to

P2 region. As described in Fig. 3, line 2, if ABWinst exists

near H (ABWinst > (P2+H)/2), then the new low rate LN

and the new high rate HN are calculated as follows:4

LN =H

γn1

(5)

HN = H × γn1 × γn

2 (6)

If ABWinst exists near L (ABWinst < (L + P1)/2), see

line 3, and then the new low rate LN and the new high rate

HN is calculated as follows:

LN =L

γn1 γ

n2

(7)

HN = L× γn1 (8)

Now, if either of the above two conditions is not satisfied,

then the ABWinst feedback by the receiver falls somewhere

else. In this case, we need to ensure that the ABWinst falls

within the P1 to P2 region. Let us assume that the ABWinst

falls in the position indicated in Fig. 3, line 4. In this case, the

rate adjustment of LN and HN (see line 5) can be calculated

as follows:

LN =ABWinst

γn1 γ

n/22

(9)

HN = ABWinst × γn1 γ

n/22 (10)

4Here, n is the number of packets in each intersecting region. As anexample, for a total of 10 packets, the value of n will be 3 and is calculatedby Eq.7 in [9].

The physical meaning of the above two equations is that if

we divide the ABWinst by γn1 ×γ

n/22 , we are actually shifting

L to the rate that is n+n/2 = 3n/2 packets5 before the current

rate. And if we multiply ABWinst by γn1 × γ

n/22 , we shift H

to the rate that is 3n/2 packets after the current rate. Now

it is obvious that, with this new rate (LN and HN ), if we

recalculate P1 and P2, as can be seen from the last horizontal

line (line 6) of Fig. 3, then ABWinst feedback by the receiver

falls between P1 and P2 and thus the algorithm converges.

C. Modified Excursion Detection Algorithm

From PathChirp’s original Excursion Detection Algorithm

(EDA), we have developed a modified EDA. PathChirp [1]

estimates the ABW by launching a series of particular packet

trains, called chirps, each of which consists of k packets

sent with an inter-packet gap that is exponentially reduced.

The chirps (numbered m = 1, 2, 3 · · · ) are sent from sender

to receiver and then statistical analysis is conducted at the

receiver by taking into account the one-way-delays (OWD),

q(m)k , faced by each packet k on the intermediate router. A

typical OWD signature is shown in Fig. 1. PathChirp uses

the shape of the signature to make an estimate E(m)k of the

per-packet available bandwidth B[t(m)k , t

(m)k+1], where t

(m)k is the

sender transmission time of packet k. It then takes a weighted

average of the E(m)k corresponding to each chirp m to obtain

estimates D(m) of the per-chirp ABW:

D(m) =

∑N−1k=1 E

(m)k Δk∑N−1

k=1 Δk

where Δk is the inter-spacing time between packet k and k+1at the receiver and N is the total number of packets in a chirp.

Finally, it makes estimates ρ[t−τ, t] of the ABW by averaging

the estimates D(m) obtained in the time interval[t−τ, t], where

τ is the total time interval of the measurement.

In order to accurately compute E(m)k , PathChirp segments

each signature into regions belonging to excursions and re-

gions not belonging to excursions. Typically, a queuing delay

signature consists of excursions from the zero axis (q(m)k > 0

for several consecutive packets) caused by a burst of cross

traffic. The ultimate goal of PathChirp’s EDA is to identify

potential starting and ending packet numbers j and i respec-

tively for an excursion. Every packet j where q(m)j < q

(m)j+1

is the potential starting point of an excursion. The end of

excursion i is defined as the first packet, where

[q(i)− q(j)] <maxj≤k≤i[q (k)− q (j)]

F(11)

Here, F is a parameter called the decrease factor. The above

equation means that, at i, the queuing delay relative to q(j) has

decreased by a factor of F from the maximum queuing delay

increase after j and upto i. If i−j > Z, that is, if the signature

is long enough (Z is the busy period threshold, with a default

value of 5 [1]), then all packets between j and i form an

excursion. The last excursion of a signature does not usually

5Assuming n = 3 will give 4.5 packets as indicated in Fig. 3. The valueof n depends on the setting of the value of L and H






terminate: that is, there is some packet l with q(m)l < q

(m)l+1

such that there is no i > l for which Eq. 11 holds (replacing

j with l in Eq. 11). This point of interest is treated as the

turning point of the queuing delay and the corresponding rate

of packet l is the desired ABW. Readers interested in greater

detail can refer to the original paper [1].

Our modified EDA is also based on the principle of self-

induced congestion like PathChirp and we also assume that

increasing queuing delay implies less ABW than the instanta-

neous packets rate and that decreasing queuing delay signifies

the opposite. Mathematically,

E(m)k ≥ Rk, ifq

(m)k ≥ q

(m)k+1 (12)

E(m)k ≤ Rk, otherwise (13)

where Rk is the corresponding rate of the packet k.

To justify the above equations, our basic idea is that, if a

certain packet’s queuing delay is less than the average queuing

delay faced by all other packets before it, and not including it,

then this packet’s rate is not considered as one of the packets

inside the excursion region or the turning point of the queuing

delay signature. Each chirp packet k that falls into one of the

following three conditions according to our algorithm decides

the value of E(m)k .

Case1: If a certain packet k belongs to an excursion that

terminates and q(m)k < q

(m)k+1, provided that q

(m)k >

avg(q(m)1 : q

(m)k−1), then we set E

(m)k = Rk.

Case2: If k belongs to an excursion that does not terminate,

provided that q(m)k > avg(q

(m)1 : q

(m)k−1), then set

E(m)k = Rl, ∀k > l, where l is the start of the

excursion.

Case3:For all those k not belonging to any excursions and

the successive k’s’ queuing delay shows a decreasing

trend until the last packet of the chirp, then we set

E(m)k = RN−1, where N is the last packet number.

This case usually happens when we send a chirp

whose maximum rate is less than the ABW.

In Case 1 and Case 2, if q(m)k < avg(q

(m)1 : q

(m)k−1), then we

simply move on to the next k and start over again. For the

pseudo-code of the algorithm see Algorithm 1, where line 4,5and line 31,32 describe Case 3. Lines 9 to 16 calculate the

average queuing delay of all other packets before the current

packet. Line 17 imposes the condition (q(m)k > avg(q

(m)1 :

q(m)k−1)) for Case 1 and Case 2. Lines 18 to 30 deal with

detecting the excursion due to cross traffic burstiness.

To better understand the modified EDA and how it differs

from PathChirp’s EDA, we present one simple example.

Figure. 4 represents one typical queuing delay of a single

chirp obtained from the simulation. We have simulated a single

bottleneck scenario where we set the actual ABW as 50 Mbps.

From the simulation outcome, we have picked a single chirp’s

queuing delay to better understand how PathChirp detects

ABW from this typical queuing delay and how our modified

approach differs from PathChirp.

In Fig. 4, we have marked the packet’s number as 1, 2, · · · 21for ease of explanation. The horizontal axis represents the

packet rates and the vertical axis represents the queuing delays

Algorithm 1 Modified Excursion Detection Algorithm

Require: qd: Queuing Delay; F : Decrease Factor;

1: Set i = 0 (current location in chirp);

2: Set j = 0 (current location where queuing delay in-

creases);

3: Set N=Total number of packets in a chirp.

Ensure: TP : Turning point of the queuing delay signature.

4: while qd[j] ≥ qd[j + 1] and j < N do5: increment j by 1;

6: end while7: Set i = j + 18: while i ≤ N do9: Set qdsum = 0 and count = 1;

10: for k = 0 to j do11: qdsum = qdsum+ qd[k];12: increment count by 1;

13: end for14: if count > 1 then15: avgQdelay = qdsum/(count− 1);16: end if17: if qd[j] > avgQdelay then18: maxQdelay = max(maxQdelay, qd[i]− qd[j])19: if (qd[i]− qd[j]) < (maxQdelay/F ) then20: j = i;21: while qd[j] ≥ qd[j + 1] and j < N do22: increment j by 1;

23: end while24: Set i = j;

25: end if26: else27: increment j by 1;

28: end if29: increment i by 1;

30: end while31: if j = N then32: decrement j by 133: end if34: Set TP = j;

of packets. We define a particular packet’s queuing delay as

qd[i], the queuing delay difference between two points as

Diff(i, j) and the maximum queuing delay between two

points as max, where i, j = 1, 2, · · · 21 and i �= j. Let’s

apply PathChirp’s EDA on these points from the beginning.

Since qd[1] ≥ qd[2], move to point 2. Now qd[2] < qd[3],so set max = Diff(2, 3), move to point 4, and calcu-

late Diff(2, 4). Since the default value of the busy period

threshold is 5, these three points (2, 3, 4) will not make any

excursion. So continue with calculating Diff(2, 5) and since

Diff(2, 5) > max, now set max = Diff(2, 5) and move to

point 6. If we continue, we will then find that Diff(2, 10) <max/F (F=decrease factor, the default value is 1.5) and the

algorithm will move its pointer from point 2 to point 10. If we

run the algorithm for the next excursion consisting of point 11to point 16, we can find that PathChirp will detect the ABW

as packet number 14’s rate (approximately 40 Mbps) instead






Fig. 4. A typical queuing delay signature obtained from the simulation.

of 50 Mbps, because in this case, max = Diff(14, 15) and

Diff(14, 16) > max/F . All other successive points (after

point 16) show an increasing trend in queuing delay.

Now, let’s apply our modified EDA to these points of

Fig. 4. Here, we additionally define avg(i) as the average

queuing delay faced by all other packets before packet i and

not including i, where i = 2, 3, · · · 21. Starting from the

beginning, since qd(1) ≥ qd(2), move to point 2 [line 4,5in Algorithm. 1]. Now qd(2) < avg(2), so ignore this point

and move on to the next point [line 17 and 27]. In a similar

way, we can find that our algorithm moves the pointer to point

6. Since qd(6) > avg(6), we calculate Diff(6, 7) and set

max = Diff(6, 7). Then, calculate Diff(6, 8) and since

Diff(6, 8) > max, we set max = Diff(6, 8). Continuing,

Diff(6, 9) < max/F , so the pointer moves to point 9 [lines

18 − 25]. Here, at this point, qd(9) ≥ qd(10) ≥ qd(11), and

the pointer moves to point 11. Now, qd(11) < avg(11), so we

move to point 12 without considering any other condition. If

we continue, we can then find that our pointer will move to

point 14 and then to points 15 and 16 because of the same

condition that we imposed. So, finally, our pointer is fixed at

point 16, because if we proceed forward after point 16, we

can see that the maximum value of the queuing delay will

be max = Diff(16, 18) and since Diff(16, 19) > max/Fand all other points after 19 show an increasing trend until

the last packet, our algorithm will detect the ABW as packet

number 16’s rate (approximately 48.24 Mbps, close enough to

the actual value of 50 Mbps).

This is a simple example from a single chirp’s queuing delay

to better understand how our modified algorithm achieves bet-

ter accuracy than PathChirp. However, we do understand that,

during our whole simulation time, we have sent many chirps

and every chirp’s queuing delay signature is not the same. So,

we are not justifying our idea based on this single queuing

delay signature. We have conducted large-scale simulation

and we have measured the ABW by averaging out all the

detected turning point values from all the chirp’s queuing delay

signatures and based on the large-scale simulation results, we

can claim that our idea works better than PathChirp and other

related state-of-the-art ABW measurement tools.

Fig. 5. Packet loss recovery.

D. Packet Loss Recovery (PLR) Algorithm

Packet loss in a wireless network is an inevitable issue

that impacts the accuracy of ABW estimation. Some tools,

e.g., PathChirp and Pathload, discard estimates when packet

loss occurs to avoid errors in ABW estimation computation.

However, this results in longer and more variable measurement

times. So, instead of discarding estimates when packet loss

occurs, we reconstruct the one-way queuing delay curve

(Fig. 1) by considering whether a single packet loss occurs or

multiple packet losses occur. We recover the possible queuing

delay information of the lost packet (Ld) based on the previous

packet’s queuing delay (Pd) and the next packet’s queuing

delay (Nd) information in case of single packet loss, and

in case of multiple packet losses, we used packet loss rate

information. We consider three cases:

Case1: In case of a single packet loss and if Pd ≥ Nd,

then there will be three possible delay values for

the lost packet. It may be equal to Pd or Nd or

greater than either or less than either. If we set

it to be equal to both or greater than either, then

there is no logical meaning, because in both cases,

Algorithm 1 will move the pointer to Nd (see lines

4,19 and 27 of Algorithm 1). So we set Ld to

be less than Nd as Ld = Nd/2. We set it to be

less because if this lost packet plays some role in

bandwidth estimation, then we will underestimate

the ABW rather than overestimate the rate of the

next packet. This underestimation will help other

applications to prevent further packet loss by sending

packets with an underestimated rate. On the other

hand, if Pd < Nd, then we set the lost packet’s delay

as the average of Pd and Nd to construct the trend

of queuing delays of successive packets. Figure 5

describes the idea.

Case2: If a single packet loss occurs in separate positions,

i.e., multiple packet losses occur independently and

not successively in a train of packets, then we apply

the recovery idea considering case 1 to separate

positions to reconstruct the one-way queuing delay.

Case3: If multiple packet losses occur successively in a

packet train, we calculate the packet loss rate rlas the number of lost packets divided by the total

number of packets and adjust the available bandwidth

as ABWm = ABWm−1 × (1− rl). Where m is the






current chirp number. If chirp m−1 also faced packet

losses then we simply discard chirp m− 1 from the

calculation of ABW, otherwise the accuracy will be

affected severely. The final ABW is the average over

all chirps measurement result.

IV. SIMULATION AND TESTBED RESULTS

In order to study the performance of NEXT-V2, simulations

and practical measurements were conducted over wired and

commercial 4G/LTE networks respectively.

A. Simulation Test and Performance Evaluation

In this section, we used a simulation environment to imple-

ment and evaluate the performance of NEXT-V2 and other

prominent ABW estimation techniques. By doing so, we

ensured that the only variables that impact the performance of

an ABW estimation technique are the algorithmic design of its

probe-stream and inference logic. Specifically, issues related

to time-stamping accuracy, timer granularity, CPU load, and

interrupt processing are taken out of the equation—a simulator

allows for perfect time-stamping and spacing of probe-packets.

We selected several prominent ABW estimation techniques—

namely, PathChirp [1], Pathload [13], and Spruce [14] that

represent existing diversity in their algorithmic design used

for inferring end-to-end ABW. We implemented each of these

algorithms in the ns-2 [16] network simulation environment.

We relied on published literature as well as publicly available

implementations [31] to extract the details of each algorithm.

Performance Metrics: We characterized the performance

of the NEXT-V2 algorithm using two types of metrics:

• Intrusiveness

• Accuracy

Intrusiveness is defined as the average bit rate of a tool.

The intrusiveness of PathChirp and NEXT as well as NEXT-

V2 can be easily compared.6 We compared the total packet

size of a single packet train between both methods. From [1],

we know that a chirp that has lower rate L, upper rate H , and

spread factor γ, consisting of N packets can be calculated as

NPathChirp =

⌊1 +

1

log (γ)log

(H

L

)⌋

On average, PathChirp sends 22 packets. With a packet size

of 1200 bytes, the total packet size of a single packet train is

22×1200 = 26.4KB. On the other hand, the average number

of packets sent by the NEXT algorithm is 10. So NEXT’s total

packet size of a single packet train is 10 × 1200 = 12KB.

Thus, the intrusiveness of NEXT is 26.4/12.0 = 2.2 times

lower as that of PathChirp. Because of our rate adjustment

Algorithm, NEXT sends a lower number of packets compared

to PathChirp.

Each run of an ABW estimation algorithm should yield a

good estimate of the end-to-end ABW. In order to quantify

the accuracy of an ABW estimation algorithm, we projected

the actual ABW and the estimated ABW in the simulation

6Since the packet structure of NEXT and NEXT-V2 is the same, theintrusiveness is also the same. So hereafter, mentioning intrusiveness to referto one implies another.

Fig. 6. Network topology with a single bottleneck link.

0

10

20

30

40

50

60

70

80

90

100

10 20 30 40 50 60 70 80 90

Estim

ated

Ava

ilabl

e B

andw

idth

Cross Traffic Rate

PathChirpPathLoad

SpruceNEXT

NEXT-V2Actual AB

Fig. 7. ABW accuracy comparison in a single bottleneck link.

results. We conducted several types of experiments to evaluate

the accuracy—we describe these next.

Single Bottleneck—One Tight Link: The accuracy of most

ABW estimation algorithms is established by their proponents

by running them on links shared by CT with a constant bit

rate (CBR). We validated our ns-2 implementation by using

the network topology depicted in Fig. 6. We ran an ABW

estimation algorithm between node Snd. and Recv. and CT

went from CBRsender to CBRreceiver. We varied the CT

load from 10 Mbps to 90 Mbps and for each load, we recorded

the estimated ABW averaged over the total simulation run.

Figure 7 plots the average of the estimated ABW against the

actual ABW. From the figure, we can see that Spruce is quite

accurate in estimating the ABW because it assumes knowledge

of the BL capacity and it is the same as the tight-link capacity

in this scenario (which is quite an impractical assumption in

many Internet paths; we will explain this in the next subsec-

tion). Our extended idea NEXT-V2 outperforms PathChirp in

most cases and is an improved version of our previously pro-

posed idea NEXT. NEXT-V2 achieves almost the same level

of accuracy as Pathload; however, the intrusiveness of NEXT

as well as our extended version NEXT-V2 is significantly less

than PathChirp as well as Pathload. Due to the fairness of

comparison, we have set L = 1 Mbps and H = 4 Mbps

for NEXT, NEXT-V2, and PathChirp. However, due to the

heuristic rate adjustment, PathChirp performs poorly. Similar

conclusions can be drawn from Fig. 9.

Single Bottleneck—Two Potential Tight Links: The in-

ference logic of the PGM techniques (e.g., Spruce) is based

on the assumption that, on the path for which the ABW is

to be estimated, the tight as well as the narrow link are the

same. In practice, this may not be the case with many Internet






Fig. 8. Multihop network topology with multiple tight link.

0

10

20

30

40

50

60

70

80

90

100

0 20 40 60 80 100 120 140 160

Ava

ilabl

e B

andw

idth

Simulation Time

PathChirpPathload

SpruceNEXT

NEXT-V2Actual ABW

Fig. 9. ABW accuracy comparison in multiple tight links.

paths—indeed, an ISP access link that is shared among a

large user population may have a lower ABW than the last-

mile narrow link for many users. In order to evaluate the

performance on such paths, we simulated the topology as

shown in Fig. 8. We ran NEXT, NEXT-V2, and other ABW

estimation algorithms for a total of 160 simulation seconds.

We started the CBR traffic (between CBRsender − 1 and

CBRreceiver− 1) of 60 Mbps at 20 simulation seconds and

stopped at 100 simulation seconds. We also started another

CBR traffic (between CBRsender−2 and CBRreceiver−2)

of 30 Mbps at 70 simulation seconds and stopped at 130simulation seconds. Note that, in this topology, the BL capacity

is 80 Mbps, whereas the tight link capacity is 100 Mbps (since

it carries most traffic). During other times, the tight link and

the BL are the same and hence the actual ABW is 80 Mbps.

The results are plotted in Fig. 9. From the figure, we see

that, under these dynamic network conditions, NEXT, NEXT-

V2, and PathChirp perform well, whereas Spruce perform

badly due to their impractical assumption (that the tight link

and narrow link are the same7). NEXT outperforms PathChirp

in most cases and tracks changes of the ABW quickly and

NEXT-V2 is improved in comparison with NEXT in terms of

accuracy. This is because our chirp structure has fine granular-

ity from the one-third to two-thirds portions and we adjusted

the L and H appropriately to fit the possible ABW into that

region. NEXT-V2 outperforms because of the modified EDA.

Cross Traffic Burstiness: The ABW is determined by the

7The tight link of a path is the one with the least amount of ABW, whilethe narrow link is the one with the least transmission capacity [13]. The tightlink of a path may not be the same as the narrow link if it carries a significanttraffic load.

0

20

40

60

80

100

120

0 50 100 150 200 250 300 350

Ava

ilabl

e B

andw

idth

Simulation Time

PathChirpNEXT

NEXT-V2

Fig. 10. ABW accuracy comparison with exponential distribution of ON-OFFcross traffic. Mean ON-OFF period 10 sec.

CT arrival process. The CT influences the probe-packet train

via dynamics in the shared packet queue at the tight link. In

particular, the queue-size grows when the collective bit rate of

the probe-packet trains and the arriving CT exceeds the link

capacity. Bursty CT creates transient queue dynamics. Since

NEXT-V2 is capable of estimating and adapting to the ABW

at fine time scales, it also reacts to the transient queue build-up

caused by the CT bursts. In order to study the impact of such

bursts on the NEXT-V2s probe stream and the steady-state

throughput it achieves, we consider Pareto and Exponential CT

models. These models generate ON/OFF traffic. During ON

periods, packets are generated at a constant bit rate. During

OFF periods, no traffic is generated. Burst times and idle

times are taken from the Pareto distribution and Exponential

distribution for Pareto and Exponential CT respectively.

The different traffic models each have their own pros and

cons. The type of network under study and the traffic char-

acteristics strictly influence the choice of traffic model used

for analysis. Traffic models that cannot capture or describe the

statistical characteristics of the actual traffic on the network are

to be avoided, since the choice of such models will result in

under-estimation or over-estimation of network performance.

There is no single model that can be used effectively for

modeling traffic in all kinds of networks. In case of high-

speed networks with unexpected demand on packet transfers,

Pareto and Exponential-based traffic models are excellent

candidates since these models take into consideration the long-

term correlation in packet arrival times [32]. Similarly, with

Marcov models, though they are mathematically tractable, they

fail to fit the actual traffic of high-speed networks [33].

To carry out this experiment, we generated a simple dumbell

topology as shown in Fig. 6. The CT and the probe packets

of PathChirp, NEXT and NEXT-V2 share the 100 Mbps tight

link (Bt). All other links have a transmission capacity of 1Gbps. We used Pareto and Exponential CT to evaluate the

performance. We used 1000 bytes packet, 1 sec and 10 sec

duration for the ON/OFF period for Pareto and Exponential

CT respectively, and a 60 Mbps CT rate.

We ran this experiment over 350 simulation seconds. We

started the CT from the beginning to the end of the simulation.

So, during this interval, the actual ABW is 40 Mbps when






TABLE IREAL TESTBED RESULTS OVER FIXED NETWORK AND WLAN

Fixed Network Testbed WLAN Testbed

Cross TrafficRate (Mbps)

Actual ABW(Mbps)

Estimated ABW(Mbps) Cross Traffic

Rate (Mbps)Actual ABW

(Mbps)

Estimated ABW(Mbps)

Pathchirp NEXT NEXT-V2 Pathchirp NEXT NEXT-V210 90 102.83 98.23 93.48 2 18 20.47 20.25 18.2620 80 92.22 76.06 79.08 4 16 18.81 18.90 17.7530 70 87.99 74.74 68.84 6 14 17.32 16.90 16.5740 60 78.31 64.19 60.88 8 12 8.07 9.23 12.2750 50 58.69 56.96 51.37 10 10 13.60 11.98 8.4760 40 38.85 36.40 37.96 12 8 10.81 10.20 7.3070 30 31.87 27.25 28.62 14 6 8.74 8.70 6.0480 20 23.75 16.36 17.73 16 4 6.78 5.90 4.0790 10 7.50 8.16 9.04 18 2 4.10 3.20 2.87

0

20

40

60

80

100

120

0 50 100 150 200 250 300 350

Ava

ilabl

e B

andw

idth

Simulation Time

PathChirpNEXT

NEXT-V2

Fig. 11. ABW accuracy comparison with pareto distribution of ON-OFFcross traffic. Mean ON-OFF period 1 sec and shape 1.5.

there is a CT burst, otherwise, the actual ABW is 100 Mbps8.

We presented the ABW fluctuation in Fig. 10 and Fig. 11

for Exponential and Pareto CT respectively. From the figure,

we can see that when the CT arrives at the tight link, it

interacts with the NEXT-V2 probe-packets and NEXT-V2

sender learns that the ABW has decreased to Bt − Ct after

a delay of 1 RTT. The NEXT-V2 sender immediately adjusts

its sending rate and fits the ABW into the high-density region

of the sending probe pattern to achieve higher accuracy. The

important point to notice here is the quick rate adaptation of

our algorithm (with two spread factors) to the sudden arrival

of CT to detect the ABW in a short time interval as well as

the responsiveness to the bursty CT nature. As compared to

PathChirp, we can see that NEXT and NEXT-V2 performed

better in tracking the ABW with the ON and OFF period of CT

since NEXT and NEXT-V2 both have only 10 probe-packets in

a single chirp. We also noticed that, while bursty CT resulted

in noisier measurement data than CBR CT, we were able to

compensate for the noise by perfoming additional processing

in ABW estimation as described in Alg. 1. Specifically, we

found that the modified excursion detection algorithm that

smoothed the measurement data by moving average resulted

in higher accuracy than PathChirp as well as our previous

approach NEXT that also used PathChirp’s EDA. We noticed a

8We could not provide the actual ABW as a ground truth in the simulationresult in Fig. 10 and Fig. 11, because the ON/OFF period is dependent onPareto and Exponential distribution and is not predetermined.

slight delay in ABW tracking in the case of PathChirp because

it generates on average 22 packets in a single chirp and the

rate adjustment of PathChirp is heuristic. So we realized the

fact that, to cope with the burstiness of CT in a real network

situation, an ABW estimation algorithm should have a lower

number of packets with optimal rate adjustment algorithm

while achieving comparably good accuracy to properly track

the changes in the ABW during the ON and OFF period of

CT. As a result, we observed that CT burstiness had limited

impact on NEXT-V2.

B. Real Testbed Results and Performance Evaluation

In order to evaluate our estimation method, the performance

of NEXT and NEXT-V2 have been studied in a controlled

testbed environment and compared with PathChirp over fixed

network topology (Fig. 6) and WLAN topology (Fig. 12). We

used the default configurations for all the probing tools. In

addition, results in [1] show that pathChirp generally performs

better with larger packets; therefore we set the packets size

of all the tools to 1000 byte. In a fixed network topology,

two CISCO Catalyst 3750 series switches using CISCO IOS

are connected together through a CAT-5e cross-cable and by

creating VLAN they served as routers; two other machines of

the testbed served as a source of controlled traffic flows using

the IXIA tool [34]. Finally, the sender and the receiver for

each measurement tool used additional PCs running Ubuntu

GNU/Linux. The bottleneck link between two switches are set

as 100 Mbps. In a WLAN topology, we deployed two wireless

nodes, one base station as an access point, one router, and two

server PCs. IXIA is used for generating CT. We measured

the wireless link capacity between the wireless node and the

access point when two wireless nodes are active and found that

the maximum throughput that each node can get is roughly

20 Mbps. So we used this rough value as the true ABW for

accuracy comparison with the ABW estimation tools.

We tested PathChirp, NEXT and NEXT-V2 in the presence

of real CT generated by IXIA. Table. I shows a measurement

performed in our testbed while the network path is loaded

with real CT varying from 10 Mbps to 90 Mbps in fixed

network and from 2 Mbps to 18 Mbps in WLAN testbed.

Each measurement result is the average over 10 repeating

measurement process for each tool. Our experiments show that

PathChirp constantly overestimates ABW and measurements






TABLE IIPACKET LOSS RECOVERY ALGORITHM EVALUATION WITH 5% PACKET

LOSS RATE.

Cross TrafficRate (Mbps)

Actual ABW(Mbps)

Estimated ABW(Mbps)

NEXT-V2(Without PLR) Error (%) NEXT-V2

(With PLR) Error (%)

10 90 49.59 44.9 59.71 33.6520 80 42.86 46.42 53.67 32.9130 70 40.96 41.48 50.35 28.0740 60 36.7 38.83 44.1 26.550 50 30.98 38.04 35.88 28.2460 40 21.19 47.02 28.97 27.5770 30 17.73 40.9 20.11 32.9680 20 11.86 40.7 13.04 34.890 10 4.57 54.3 5.69 43.1

Fig. 12. WLAN network topology.

are quite unstable. This is a well-know problem of PathChirp:

similar results have been obtained in [35][36]. On the other

hand, the accuracy and stability of NEXT-V2 is notable: we

found that 80% of estimations exhibit a relative error lower

than 5% in fixed network testbed and 70% of estimations

exhibit a relative error lower than 15% in WLAN testbed.

We evaluated the packet loss recovery (PLR) algorithm with

the topology as shown in Fig. 6. We used Dummynet [37] in

the network to control the packet loss rate. We set 5% packet

loss rate, sent CT with different rates and estimated the ABW.

First we evaluated NEXT-V2 without PLR algorithm9 and then

with PLR algorithm. The result is shown in Table II. In the

experiment, we saw that with 5% packet loss rate, multiple

packet loss occurred in a single chirp and only in few cases

single packet loss occurred. So the result shown in Table II

mostly reflects the adjustment of Case3 in PLR algorithm.

Furthermore, Dummynet internally round times to multiples

of the quantum of the system timer, which runs HZ times per

second (in our case HZ = 1000). This introduces a timing

error of 1/HZ = 1 ms that is randomly added to some packets

queuing delay and cause further misdetections of turning point

(Fig. 1). Although multiple and successive packet loss is a rare

phenomenon in today’s real network, we firmly believe that,

our algorithm can detect the ABW with greater accuracy for

Case1 and Case2 as described in PLR algorithm.

We have also evaluated our algorithm using LTE connection

of a commercial Japanese mobile operator. NEXT-V2 is imple-

mented in an Android OS-based mobile terminal and Linux

OS-based server PC and evaluated over a 4G/LTE network

9Without PLR algorithm, the NEXT-V2 receiver simply returns zero ABWwhen a packet loss occurs. Thus the probe packet sender do not update thelow rate and high rate, which further wrongly estimates the ABW in the nextround. As a result, the final estimated ABW which is an average over allestimates is affected by the packet loss rate.

Fig. 13. 4G/LTE network topology.

in real cross traffic scenarios. We created an Android OS-

based ABW measurement tool using Android SDK tools that

initiate a NEXT-V2 session by generating UDP probe traffic

towards the Linux-based server PC located in KDDI R&D

Labs during up-link measurement and vice versa during down-

link measurement. The measurement tool takes the address

to the server and an associated port number as an input to

exchange packets. The configurable parameters of NEXT-V2

such as the low rate and high rate of the packet trains, spread

factors, etc., can be specified. The results are displayed on

a graph as well as stored in a log file. The graph displays

the total probe traffic send for a single estimate in Kilobytes

(KB). Further, the log file generated shows the details of

the measurement. The measurement tool utilizes Androids

telephony API to display the network type and connection

state. It displays values such as RSSI, MCC, MNC, and

LAC of the network to which we are currently connected. A

basic essential requirement for creating network maps is usage

of geolocation services. Android location API is utilized to

associate the ABW estimation with the measurement location.

The latitude and longitude values are displayed on screen as

well as added to the log file with associated ABW estimation.

Android application installation and test performance was

conducted on an HTC smartphone consisting of a Quadcore

processor. It consists of 2 GB RAM with support for LTE,

HSDPA, HSUPA, and HSPA+.

The 4G/LTE network topology for experimentation is shown

in Fig 13. All wired links have a capacity of at least 100Mbps. According to the mobile operator, 4G/LTE networks can

achieve more than 100 Mbps in the physical layer theoretically.

But carrier aggregation-supported mobile terminals (MTs) are

very rare now. The MTs used in the experiments support a

data rate up to 75 Mbps in a 4G/LTE network. To validate

the measurement results, we compared the true ABW and the

estimated ABW produced by NEXT-V2. In a fixed-wireline

testbed, the true ABW can be measured using tools such as

tcpdump. Then, the ABW can be computed as the difference

between the fixed-line capacity and the cross traffic rate.

However, in a 4G/LTE network, the capacity of the link

at the IP layer is very difficult to determine, because it

varies with the radio quality. Furthermore, even if the radio

quality is known, there is no simple formula to calculate the

cross traffic impact on the capacity since it depends on the

packet size, varying wireless channel conditions, scheduling

and modulation techniques, pre-configured QoS parameters,

etc. Instead, in this paper, we used the maximum achievable

FTP throughput as a ground truth of the true ABW. To inves-

tigate the accuracy of our results obtained, we also compared






TABLE IIIRELATIONSHIP BETWEEN ABW AND FTP THROUGHPUT IN LTE

NETWORK

ResourceBlock

TheoreticalCapacity (Mbps)

ActualCapacity(Mbps)

CBR CrossTraffic

FTPThroughput (Mbps)

6 6.04 4.53 2 2.4115 15.12 11.24 5 5.3525 25.2 18.9 10 7.6750 50.4 37.8 15 20.3675 75.6 56.7 25 29.19100 100.8 75.6 40 33.29

them with the FTP throughput measurement results on both

links using the XCAL Speedtest available for the Android

smartphone in Google Play Store [38]. The XCAL Speedtest,

developed by Accuver Communications is an Android OS-

based tool that provides a solution for wireless network testing.

It supports major wireless technologies displaying current

wireless network characteristics and investigates cell coverage

and capacity.

Furthermore, to understand the relationship between FTP

throughput and the ABW, we conducted simulation in ns-

3 [39] with the topology shown in Fig. 13, where there is

an FTP session running between a TCP cubic server and

a mobile client and the mobile client is downloading files

using the TCP connection. We used TCP Cubic as it is the

default TCP congestion control algorithm in Linux OS in real

networks. We also used the SISO transmission mode for the

MTs and eNB. We would like to see how closely the FTP

throughput resembles the ABW with the presence of cross

traffic. The result is shown in Table III. For different resource

blocks (RBs), we ran different amounts of CBR cross traffic

from the cross traffic source to the cross traffic destination

and calculated the FTP throughput. From Table III, the first

column represents different RBs in eNB. The second and

the third column represents the theoretical capacity and the

actual capacity of the corresponding RBs respectively. Due

to the overhead used for controlling and signaling, which is

approximately 25% introduced by PDCCH, the down-link RS

signal, and other control signals [40], the actual capacity is

75% of the theoretical capacity. For different resource blocks

(RBs), we ran different amounts of CBR cross traffic from

the cross traffic source to the cross traffic destination and

calculated the FTP throughput. We can see that FTP through-

put closely resembles the ABW.10 Table IV summarizes the

results obtained by NEXT-V2 over a 4G/LTE network. We

conducted measurement on different dates and times over

different places to ensure the fairness of the measurement.

We conducted measurement while walking through the streets

and sometimes by car at an average speed of 30 to 40km/h. During the measurement, the mobile terminals received

signal strength indicator (RSSI) was between −67 dbm to

−92 dbm in different locations. We used 1.1 and 1.05 as

the two spread factors for ABW measurement. For each run,

we measured the ABW in the up-link (UL) and down-link

(DL) directions. We compared the estimated ABW with the

FTP throughput and XCAL Speedtest. For this purpose, we

uploaded and downloaded a 10 MB file to and from one of

the servers in KDDI Labs and measured the FTP throughput

10The remaining actual capacity after the amount of CBR cross traffic.

Fig. 14. Estimated error while varying packet size.

(UL) and FTP throughput (DL) respectively. From the table,

we can see that NEXT-V2 achieves very good accuracy as

the estimates closely follow the FTP throughput and XCAL

Speedtest results.

The results achieved using our measurement tool does not

provide information regarding existing cross traffic in a com-

mercial LTE network. While the experiments were run during

particular times of the day, an assumption is made of the

existence of constant cross traffic in the network during a short

time measurement period. We compared our measurement tool

results with FTP throughput values where the FTP results

provides achievable throughput of the network being less than

the theoretical capacity of the LTE network. It is possible

to run experiments to estimate available bandwidth in LTE

networks with existing measurement tools on a computer

tethered via a LTE-enabled phone or Dongle. However, our

objective is to consider the development and deployment of a

measurement tool for Android OS-based devices, and existing

tools for computers are not an option for measuring available

bandwidth on LTE networks. The existence of FTP sessions

on the Android device allows us to select it as a benchmarking

option against our measurement tool.

Figure 14 shows the estimated error11 of NEXT-V2 while

varying the probe-packet size. For each packet size, we re-

peated the measurement 20 times and then averaged the 20measurement results for fairness. For each run, we measured

the ABW in the down-link, the total time required for NEXT-

V2 to produce a single estimate, and the total bytes sent

during the measurement. We can see from the figure that the

estimated error varies with the probe-packet size. The reason

for the varying measurement estimates of the ABW can be

derived from the link-level acknowledgments. If the probe-

packet size is small, then the extra overhead introduced by

the link-level acknowledgments is relatively large compared

to a larger probe-packet size. This will affect probe-packet

separation and hence the rate-response curve, which is the

basis of accurately determining the ABW by detecting the

turning point. Thus, the ABW produced by NEXT-V2 varies

with varying probe-packet size. We can see that the estimation

11The difference between the estimated value of NEXT-V2 and FTPthroughput.






TABLE IVREAL TEST-BED RESULTS OVER 4G/LTE NETWORK

Location Date/TimeNEXT-V2 FTP Throughput XCAL Speedtest

UL(Mbps) DL(Mbps) UL(Mbps) DL(Mbps) UL(Mbps) DL(Mbps)

KDDI Lab(Fujimino City)

Aug 5/15:59 5.25 7.24 6.54 7.16 7.17 9.33Aug 5/16:36 6.32 10.42 7.98 10.52 8.00 12.34Aug 6/9:36 7.47 7.52 9.15 7.21 8.5 9.2Aug 6/13:26 8.43 9.28 7.21 8.77 7.5 9.7Aug 7/ 9:21 5.36 7.18 4.59 6.53 6.54 8.31Aug 7/17:06 3.92 15.82 3.96 15.86 4.2 16.2Aug 14/11:27 8.94 12.60 7.82 12.38 10.12 13.3Aug 14/11:42 7.86 6.94 8.08 8.00 6.89 7.23

Kawagoe CityAug 10/17:53 7.43 17.54 6.00 18.80 9.35 19.24Aug 10/18:38 6.71 8.38 5.73 10.37 7.5 10.2Aug 10/19:42 10.88 16.08 7.39 16.54 9.8 18.90

Tokyo City

Aug 19/13:45 3.89 6.46 4.77 6.72 5.5 7.8Aug 19/14:44 8.74 15.55 8.62 13.89 10.75 16.65Aug 19/15:31 9.8 16.47 7 15.44 9.23 17.3Aug 19/16:57 8.19 9.38 6.78 7.43 8.45 11.23

TABLE VTOTAL TRAFFIC SENT AND TOTAL ELAPSED TIME FOR DIFFERENT PACKET

SIZE

Packet Size (Byte) Total Traffic Sent (KB) Total Elapsed Time (sec)600 189 2.32700 227 2.22800 225 3.45900 277 2.541000 221 2.891100 389 2.231200 356 3.111300 281 2.291400 447 4.891500 399 3.99

error produced by NEXT-V2 with a probe-packet size from

1000 bytes and above is around 10% and less than 10%

respectively. We also measured the total traffic sent and the

total elapsed time for a single ABW estimate as can be seen

from Table V. The result indicates that NEXT-V2 estimates

the ABW with less intrusiveness and within 2 to 4 seconds

depending on the probe-packet size. For this experiment, we

downloaded a 10 MB size file and it took about 10 to 12seconds.

V. CONCLUSION

In this work, we presented the details of NEXT-V2, an

extended version of NEXT, an active probing algorithm that

features an efficient measurement scheme for end-to-end ABW

estimation in a fixed, WLAN and 4G/LTE network. We have

proposed a unique packet train structure, an optimal rate

adjustment algorithm, and a modified excursion detection

algorithm to identify the ABW with higher accuracy, less

convergence time, and less overhead. NEXT-V2 is compared

with other existing ABW estimation tools in a simulation and

real testbed to prove its algorithmic strength. From the real

testbed results, we can see that 90% of the cases NEXT-V2

reports a less than 10% error. On the other hand, NEXT and

PathChirp reports a less than 10% error for 70% and 20% of

the cases respectively. Intrusiveness of NEXT-V2 is reduced

by 50% as compared to that of PathChirp.

From experiments on a 4G/LTE network, a few conclusions

can be drawn. First, current bandwidth estimation tools are

significantly impacted by wireless network conditions, such as

contention from other traffic and rate adaptation. This yields

inaccurate estimates, high and varying convergence times, and

intrusiveness. Thus, current tools are generally impractical

for applications such as streaming multimedia that require

fast, accurate, and non-intrusive bandwidth estimations even

when the last hop is over a WLAN. Second, the experiments

conducted and results achieved on a commercial 4G/LTE

network show the ability of NEXT-V2 to make quick and less

intrusive ABW estimations, with higher accuracy. A less than

15% error is reported in case of down-link ABW estimation

for 80% of the cases and a less than 20% error is reported in

case of up-link ABW measurement for 70% of the cases.

Comparison of ABW estimations with FTP throughput

measurements on Android OS is carried out using various

real-time cross traffic. This measurement tool uses minimal

power and network resources making it possible to conduct

multiple test sessions. It provides rich data sets with ABW

estimations with associated geo-location values. The ideal tool

would be one that provides accurate estimations, less overhead,

quick response time, and 100% reliability, whereas there is

no mandatory requirement for the ideal tool in all scenarios.

Tool selection is based on the application and the network

environment.

REFERENCES

[1] V. J. Ribeiro, R. H. Riedi, R. G. Baraniuk, J. Navratil, and L. Cottrell,“Pathchirp: Efficient available bandwidth estimation for network paths,”in Passive and Active Measurement Workshop, 2003.

[2] M. Li, C. Mark, and K. Robert, “Wbest: A bandwidth estimation toolfor IEEE 802.11 wireless networks,” in 33rd IEEE Conference on LocalComputer Networks, 2008. LCN 2008., Oct 2008, pp. 374–381.

[3] M.A. Alzate, J.C. Pagan, N.M. Pena, and M.A. Labrador, “End-to-end bandwidth and available bandwidth estimation in multi-hop IEEE802.11b ad hoc networks,” in 42nd Annual Conference on InformationSciences and Systems, 2008. CISS 2008., March 2008, pp. 659–664.

[4] P. Papageorge, J. McCann, and M. Hicks, “Passive aggressive measure-ment with mgrp,” SIGCOMM Comput. Commun. Rev., vol. 39, no. 4,pp. 279–290, Aug. 2009.

[5] M. Jain and C. Dovrolis, “Ten fallacies and pitfalls on end-to-endavailable bandwidth estimation,” in Proceedings of the 4th ACMSIGCOMM Conference on Internet Measurement, New York, NY, USA,2004, IMC ’04, pp. 272–277.

[6] C. D. Guerrero and M. A. Labrador, “On the applicability of availablebandwidth estimation techniques and tools.,” Computer Communica-tions, vol. 33, no. 1, pp. 11–22, 2010.






[7] X. Liu, K. Ravindran, and D. Loguinov, “Multi-hop probing asymptoticsin available bandwidth estimation: Stochastic analysis,” in Proceedingsof the 5th ACM SIGCOMM Conference on Internet Measurement, 2005,IMC ’05.

[8] B. Mel, M. Bjrkman, and P. Gunningberg, “Regression-based availablebandwidth measurements,” in International Symposium on PerformanceEvaluation of Computer and Telecommunications Systems, 2002.

[9] A.K. Paul, A. Tachibana, and T. Hasegawa, “NEXT: New enhancedavailable bandwidth measurement technique, algorithm and evaluation,”in IEEE PIMRC, Washington DC, USA, 2014, pp. 443–447.

[10] D. Gupta, D. Wu, P. Mohapatra, and C. N. Chuah, “Experimentalcomparison of bandwidth estimation tools for wireless mesh networks,”in INFOCOM 2009, IEEE, April 2009, pp. 2891–2895.

[11] P. Zhao, X. Yang, C. Dong, S. Yang, S. Bhattarai, and W. Yu, “On an ef-ficient estimation of available bandwidth for ieee 802.11-based wirelessnetworks,” in Global Telecommunications Conference (GLOBECOM2011), 2011 IEEE, Dec 2011, pp. 1–5.

[12] B. Melander, M. Bjorkman, and P. Gunningberg, “A new end-to-endprobing and analysis method for estimating bandwidth bottlenecks.,” inIEEE GLOBECOM, 2000, pp. 415–420.

[13] M. Jain and C. Dovrolis, “Pathload: A measurement tool for end-to-end available bandwidth,” in Proceedings of Passive and ActiveMeasurements (PAM) Workshop, 2002, pp. 14–25.

[14] J. Strauss, D. Katabi, and F. Kaashoek, “A measurement study ofavailable bandwidth estimation tools,” in Proceedings of the 3rd ACMSIGCOMM Conference on Internet Measurement, New York, NY, USA,2003, IMC ’03, pp. 39–44.

[15] N. Hu and P. Steenkiste, “Evaluation and characterization of availablebandwidth probing techniques.,” IEEE Journal on Selected Areas inCommunications, vol. 21, no. 6, pp. 879–894, 2003.

[16] “The Network Simulator NS-2,” http://www.isi.edu/nsnam/ns/.[17] A. Shriram and J. Kaur, “Empirical evaluation of techniques for measur-

ing available bandwidth,” in INFOCOM 2007. 26th IEEE InternationalConference on Computer Communications. IEEE, 2007, pp. 2162–2170.

[18] “sflow, version 5.,” http://www.sflow.org/sflow version 5.txt.[19] C. Yu, C. Lumezanu, Y. Zhang, V. Singh, G. Jiang, and H.V. Mad-

hyastha, “Flowsense: Monitoring network utilization with zero mea-surement cost,” in Proceedings of the 14th International Conference onPassive and Active Measurement, Berlin, Heidelberg, 2013, PAM’13,pp. 31–41, Springer-Verlag.

[20] “High frequency sflow v5 counter sampling,”ftp://ftp.netperf.org/papers/high freq sflow/hf sflow counters.pdf.

[21] H. Wang, K.S. Lee, E. Li, C.L. Lim, A. Tang, and H. Weather-spoon, “Timing is everything: Accurate, minimum overhead, availablebandwidth estimation in high-speed wired networks,” in Proceedings ofthe 2014 Conference on Internet Measurement Conference, New York,NY, USA, 2014, IMC ’14, pp. 407–420, ACM.

[22] S. Joel and B. Paul, “An active measurement system for sharedenvironments,” in Proceedings of the 7th ACM SIGCOMM Conferenceon Internet Measurement, New York, NY, USA, 2007, IMC ’07, pp.303–314, ACM.

[23] S. Joel, B. Paul, and C. Mark, “Router primitives for programmableactive measurement,” in Proceedings of the 2Nd ACM SIGCOMM Work-shop on Programmable Routers for Extensible Services of Tomorrow,New York, NY, USA, 2009, PRESTO ’09, pp. 13–18, ACM.

[24] S. Shah, K. Chen, and K. Nahrstedt, “Available bandwidth estimation inIEEE 802.11-based wireless networks,” in First ISMA/CAIDA Workshopon Bandwidth Estimation (BEst), San Diego, CA, USA, December 2003.

[25] H.K. Lee, V. Hall, K.H. Yum, K. Kim, and E.J. Kim, “Bandwidthestimation in wireless LANS for multimedia streaming services,” inIEEE International Conference on Multimedia and Expo, 2006, July2006, pp. 1181–1184.

[26] R. Kapoor, L. Chen, L. Lao, M. Gerla, and M. Y. Sanadidi, “Capprobe: Asimple and accurate capacity estimation technique,” SIGCOMM Comput.Commun. Rev., vol. 34, no. 4, pp. 67–78, Aug. 2004.

[27] K. Lakshminarayanan, V.N. Padmanabhan, and J. Padhye, “Bandwidthestimation in broadband access networks,” in Proceedings of the 4thACM SIGCOMM Conference on Internet Measurement, New York, NY,USA, 2004, IMC ’04, pp. 314–321, ACM.

[28] A. Johnsson, B. Melander, and M. Bjorkman, “Bandwidth measurementin wireless networks,” in Mediterranean Ad Hoc Networking Workshop,Porquerolles, France, June 2005.

[29] A. Shriram et al., “Comparison of public end-to-end bandwidthestimation tools on high-speed links,” in Passive and Active NetworkMeasurement Workshop (PAM), Boston, MA, Mar 2005, vol. 3431, pp.306–320, PAM 2005.

[30] J. Huang, F. Qian, Y. Guo, Y. Zhou, Q. Xu, Z. M. Mao, S. Sen, andO. Spatscheck, “An in-depth study of LTE: effect of network protocoland application behavior on performance,” in Proceedings of the ACMSIGCOMM, New York, USA, 2013, pp. 363–374.

[31] “Tools for bandwidth estimation,” http://www.icir.org/models/tools.html.[32] A. Adas, “Traffic models in broadband networks,” Comm. Mag., vol.

35, no. 7, pp. 82–89, July 1997.[33] A.M. Mohammed and A.F. Agamy, “A survey on the common network

traffic sources models,” IJCN, vol. 3, no. 2, pp. 103–115, 2011.[34] “The IXIA Traffic Generator,” http://www.ixiacom.com/.[35] A. Shriram, M. Murray, Y. Hyun, N. Brownlee, A. Broido,

M. Fomenkov, and k. claffy, “Comparison of public end-to-endbandwidth estimation tools on high-speed links,” in Proceedings of the6th International Conference on Passive and Active Network Measure-ment, Berlin, Heidelberg, 2005, PAM’05, pp. 306–320, Springer-Verlag.

[36] S. Ekelin, M. Nilsson, E. Hartikainen, A. Johnsson, J.-E. Mangs,B. Melander, and M. Bjorkman, “Real-time measurement of end-to-end available bandwidth using kalman filtering,” in Network Operationsand Management Symposium, 2006. NOMS 2006. 10th IEEE/IFIP, April2006, pp. 73–84.

[37] M. Carbone and L. Rizzo, “Dummynet revisited,” SIGCOMM Comput.Commun. Rev., vol. 40, no. 2, pp. 12–20, Apr. 2010.

[38] “Xcal Speedtest,” https://play.google.com/store/apps.[39] “The network simulator ns-3,” https://www.nsnam.org.[40] NGMN Alliance, “Guidelines for lte backhaul traffic estimation[white

paper],” 2011.

Anup Kumar Paul received a B.Sc.(hons), Mastersdegree in information and communication engineer-ing from the University of Rajshahi, Bangladesh, anda Ph.D. degree in global information and telecom-munication studies from Waseda University, Tokyo,Japan, in 2004, 2006, and 2013 respectively. Heis a research engineer at KDDI R&D LaboratoriesInc., Japan. He joined the KDDI R&D Lab, Japan,in 2013. Since then, he has been actively involvedin research and development activities in the fieldof high-performance transport protocols, network

measurement and traffic management in 4G/LTE mobile networks. He is amember of IEEE.

Atsuo Tachibana received B.E. and M.E. degreesfrom Osaka University, Japan in 2000 and 2002,respectively, and received his PhD in InformationEngineering from the Graduate School of ComputerScience and Systems Engineering, Kyushu Instituteof Technology, Japan in 2012. His research interestsinclude issues related to network measurement andtraffic management in wired and wireless networks.Currently, he is a research manager at KDDI R&DLaboratories Inc.

Teruyuki Hasegawa received B.E. and M.E. de-grees in electrical engineering from Kyoto Univer-sity in 1991 and 1993 respectively, and a Ph.Ddegree in information science and technology fromthe University of Tokyo in 2008. Since joining KDD(now KDDI) in 1993, he has been working in thefield of high-speed communication protocols, multi-cast systems, and future Internet. He is currently thesenior manager of the IP Communication QualityLab. at KDDI R&D Laboratories, Inc. He receivedthe Meritorious Award on Radio of ARIB in 2003.

He is a member of IEICE and IPSJ.



an enhanced available bandwidth estimation technique for an end to end network path

Design