Content uploaded by Emanuele Goldoni
Author content
All content in this area was uploaded by Emanuele Goldoni on Sep 26, 2018
Content may be subject to copyright.
Assolo, a New Method for Available Bandwidth Estimation
Emanuele Goldoni
University of Pavia
Department of Electronics
27100 - Pavia, Italy
emanuele.goldoni@unipv.it
Giuseppe Rossi, Alberto Torelli
University of Pavia
Department of Computer Eng. and System Science
27100 - Pavia, Italy
giuseppe.rossi@unipv.it, alberto.torelli01@ateneopv.it
Abstract
End-to-end available bandwidth estimation is very im-
portant for bandwidth dependent applications, quality of
service verification and traffic engineering. Although sev-
eral techniques and tools have been developed in the past,
producing reliable estimations in real-time still remains
challenging – it is necessary to ensure that the measure-
ment process is accurate, non-intrusive and robust to non-
deterministic delays or traffic bursts. This paper presents
ASSOLO, a new active probing tool for estimating avail-
able bandwidth based on the concept of “self-induced con-
gestion”. ASSOLO features a new probing traffic profile
called RE AC H (Reflected ExponentiAl Chirp), which tests a
wide range of rates being more accurate in the center of
the probing interval. Moreover, the tool runs inside a real-
time operating system and uses some de-noising techniques
to improve the measurement process. Experimental results
show that ASSOLO outperforms pathChirp, a state-of-the-
art measurement tool, estimating available bandwidth with
greater accuracy and stability.
1. Introduction
The available bandwidth of a network path is a crucial
metric in quality-of-service management, traffic engineer-
ing or congestion control. Moreover, VOIP, peer-to-peer
and video-streaming are examples of widely-used applica-
tions that could greatly benefit from the knowledge of the
available bandwidth along an Internet path.
Researchers have proposed several end-to-end measure-
ment algorithms during the last years. These techniques
can be divided in two categories: active probing and pas-
sive measurements. The latter infer the required infor-
mation from existing data transmissions; for this reason,
these methods can be applied only to network paths that
recently carried traffic. Active probing techniques, on the
other hand, inject end-to-end test traffic into the network
and estimate available bandwidth observing the effect of
cross-traffic on probe packets. Several active probing tools
have emerged in recent years, such as IGI [10], Delphi [20],
Spruce [24], Pathload [11], TOPP [14, 12], pathChirp [21],
FEAT [27] and BART [4]. They differ mainly in the struc-
ture of probe streams and in the algorithms used to estimate
available bandwidth.
This paper presents ASSOLO (Available-bandwidth
Smart Sampling On-Line Tool), a new self-induced conges-
tion probing tool for estimating available bandwidth. AS-
SOLO features a new probing traffic profile called REA CH
(Reflected ExponentiAl Chirp). A R EACH tests a wide
range of rates and is more accurate in the center of the
probed interval. Moreover, ASSOLO uses a combination
of new and existing techniques to improve the accuracy and
stability of results.
The rest of the paper is organized as follows. First, Sec-
tion 2 introduces the related work on available bandwidth
estimation and focuses on the Probe Gap Model, the general
measurement scheme adopted by our tool. Next, Section 3
illustrates the main algorithm and the new features intro-
duced in ASSOLO. Evaluation of ASSOLO is presented in
Section 4, including results obtained comparing our solu-
tion to the state-of-the-art tool pathChirp. Finally, in Sec-
tion 5 we conclude and we outline future works.
2. Related Work
The majority of existing tools belong either the Probe
Gap Model (PGM) or the Probe Rate Model (PRM). In the
former model, a tool sends probe packet pairs or trains; then
it exploits the dispersion of packets on the receiver side to
calculate available bandwidth. The main assumption is that
the link with the minimum available bandwidth is also the
link having the minimum capacity. However, this is proba-
bly also the biggest limit of this approach: the hypothesis is
not valid for many Internet paths. Most notable tools based
on the Probe Gap Model are Spruce [24], IGI [10] and Del-
phi [20].
The Probe Rate Model, instead, is based on the concept
of self-induced congestion. The underlying idea is quite
simple: if a sequence of packets is sent at a rate lower than
the available bandwidth along the network path, then the ar-
rival rate of packets at the receiver will not exhibit any no-
table variation and it will match with the sender’s rate. On
the other hand, if the sending rate exceeds available band-
width, one or more intermediate queues will fill up and the
probe traffic will experience delays. Thus, the measurement
is performed through the research of the turning point at
which the probe stream starts seeing an increasing trend.
The PRM model has proved to be accurate and it is used in
many estimation tools, such as TOPP, Pathload, pathChirp,
FEAT and BART.
TOPP [14, 12] and Pathload [11] use a constant bit-rate
stream, sending pairs or trains of packets at a given rate
and changing this rate every round. TOPP increases linearly
the sending rate in successive streams, trying to find out
the exact turning point. On the other hand, Pathload varies
the probing rate using a binary search scheme and the final
output, obtained after 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.
pathChirp [21] sends a variable bit-rate stream called
chirp, which consists of exponentially spaced packets. A
chirp allows to probe the network path over a wide range
of rates injecting only one stream – if the delays show an
increasing trend starting from a particular packet, the asso-
ciated rate is used to infer the unused capacity. pathChirp
can estimate available bandwidth sending only one chirp:
this feature makes the measurement process quite fast and
lightweight. However, pathChirp samples the lower rates
more frequently than the higher rates. Therefore the tool
is less accurate if actual available bandwidth is not located
nearby the beginning of the probing range.
BART [4] relies on sequences of packet pairs sent at ran-
domized rates. This tool uses also a Kalman filter to track
the evolution of available bandwidth in real-time and to fil-
ter out noisy observations. BART is lightweight, efficient
and non-intrusive; however, the tool is still in development
and it is not freely available.
FEAT [27] is a recent tool which features a probe pattern
called fisheye stream. A fisheye stream consists of pack-
ets of equal size which are sent at a changing rate, from a
lower bound to a maximum probing rate. The tool identifies
also an interval, called “focus region”, where the available-
bandwidth is most likely to be. Inside this region the sam-
pling frequency is higher and the number of packets sent
for each sampling rate is larger. This approach creates a
more identifiable turning point but it also makes FEAT’s
measurement process intrusive. While preliminary results
look quite promising, to our knowledge the developed soft-
ware has never been released.
3. ASSOLO
ASSOLO is based on the concept of “self-induced con-
gestion”. Unique to this tool is a new probe traffic profile
called RE AC H (Reflected ExponentiAl Chirp), which tests
a wide range of rates being more accurate in the center of
the probing interval. A R E ACH combines both the advan-
tages of a pathChirp’s chirp and a FEAT’s fisheye stream: it
probes multiple rates with a single stream, it is more accu-
rate at the center of the stream, where the actual available
bandwidth is likely to be, and it injects a negligible amount
of traffic into the network . ASSOLO introduces also some
techniques to minimize the impact of different sources of
errors on the estimation process.
3.1. Probing stream
ARE AC H tests different rates increasing the instanta-
neous packet rates from a lower bound Lto a maximum
rate U. The first kpackets of the stream probe values lower
than the center H=U+L
2; additional kpackets test values
between Hand the maximum probing rate U. However,
the probing rates do not increase linearly in a RE AC H. In-
stead, the density of the stream increases as well as values
approach the center of the interval [L, U]. Then, once the
rate Hhas been tested, probing density start decreasing.
The same can be said for the accuracy of the estimation,
since it is proportional to the density of the probing stream.
A similar idea is proposed in [27], but our method intro-
duces a different spacing algorithm.
The parameter σdefines the maximum relative accuracy
of ASSOLO’s estimations. Given the probing range, the
absolute error Saround the center of the probing interval is
defined as:
S=σU−L
2.(1)
Moreover, the algorithm uses a coefficient γto control how
fast the density of streams changes. This parameter is simi-
lar to the spread factor used by pathChirp, although the two
resulting trains are quite different. ASSOLO uses by default
σ= 5% and it sets γto 1.2. However, it is important to
note that the choice of these parameters is arbitrary – values
would typically be set based on the specific requirements of
the target application.
To better describe the REACH stream generated by AS-
SOLO we introduce an additional parameter ∆xdefined as
follows:
∆x=S·γ|x−1|=σU−L
2γ|x−1|(2)
Starting from the center Hof the probing interval towards
the upper bound U, instantaneous packet rates in a RE AC H
are H, H + ∆1, H + ∆1+ ∆2, H + ∆1+ ∆2+ ∆3, .... On
the other hand, from the center towards the lower bound,
probing rates are H, H −∆1, H −∆1−∆2, H −∆1−
∆2−∆3, ... A more formal description of instantaneous
probing rates Rjtested by a RE AC H is:
(Rj=H−Pk
j=1 ∆j,if j= 1
Rj=Rj−1+ ∆j,∀j > 1(3)
The resulting stream is shown in Figure 1. As also the name
RE AC H (Reflected ExponentiAl CHirp) suggests, the profile
is symmetric: the right and the left part look like two mir-
rored exponential functions.
ASSOLO uses kpackets to test values between Hand
the upper bound U. Thus, the instantaneous rate Rkassoci-
ated with the kth packet is the maximum probing rate. We
can write this condition as:
U=Rk=H+
k
X
i=1
∆i→U−H=
k
X
i=1
∆i(4)
If we substitute the values of ∆iand H, we get:
U−H=σU−L
2
k
X
i=1
γ|i−1|(5)
U−U+L
2=σU−L
2
k
X
i=1
γ|i−1|(6)
U−U+L
2
σU−L
2
=
k
X
i=1
γ|i−1|(7)
U−L
2
σU−L
2
=
k
X
i=1
γ|i−1|(8)
1
σ=
k
X
i=1
γ|i−1|(9)
In addition, the value of the truncated sum is:
k
X
i=1
γ|i−1|=γk+1 −1
γ−1(10)
Combining Equations 9 and 10, we get:
γk+1 =γ−1
σ+ 1 (11)
and so
k=logγγ−1
σ+ 1−1(12)
To be more precise, Rkis the maximum sending rate not
exceeding the upper bound. Equation 4 should then take the
form:
U≥H+
k
X
i=1
∆i(13)
Hence the correct value of kis:
k=logγγ−1
σ+ 1−1(14)
ARE AC H uses the first kpackets of the stream to probe
values lower than the center H=U+L
2. Then, the stream
probes the rate H; finally, other kpackets tests values be-
tween Hand the maximum probing rate U. As a result, a
RE AC H probes 2k+ 1 rates exploiting relative delays be-
tween probe packets.
The use of information only on the relative delays is
adopted also by other notable tools, such as pathChirp and
BART. This approach has a twofold advantage: it allows
to require neither clock synchronization nor clock-offset
knowledge between the two end hosts probing the network.
However, it is important to consider that the first packet of
the train itself does not have any associated rate. Instead, it
is used as a reference value to calculate all successive rela-
tive queuing delays within a stream.
Therefore, out tool needs to send an additional packet at
the beginning of the RE AC H. The total number Nof packets
used by ASSOLO to probe 2k+ 1 rates is:
N= 1 + (2k+ 1) = 2k+ 2 (15)
Since we know the size of a RE AC H, we can now rewrite
Equation 3 as:
Rj=H+sign j−N
2·S·γ|j−N
2|−1
γ−1(16)
3.2. End-hosts predictability
A fundamental difficulty with existing measurement
tools rise from a number of issues on both end-hosts
and network paths [30]: system timing, hardware errors
and end-to-end pathologies could produce a considerable
amount of noise in the individual network observations. In
[17] the impact of context switching on the measurement
process is analyzed in depth. Some tests conducted in our
lab confirmed that a significant amount of noisy observa-
tions is due to the non-deterministic behavior of the operat-
ing systems hosting the sender and the receiver.
Network measurement tools have strict operational dead-
lines between the arrive of a packet and the application’s re-
sponse to that event – the same can be said for the sender
side, where the packet sent by the application should ideally
Figure 1. Distribution of packets in a REACH
stream.
start with no delay. To meet these deadlines we decided to
use a real-time operating systems (RTOS), which can guar-
antee predictability and accurate system timings for appli-
cations. ASSOLO run inside a GNU/Linux kernel with RT-
Preempt [15] patch enabled, thus using a fully preemptible
kernel with high-resolution kernel timers. In order to min-
imize the impact of context switches on the bandwidth es-
timation process, the tool gets the highest priority on both
the end-host systems while probing the network.
Compared to other Linux real-time approaches, such as
RTAI [5] and Xenomai [6], RT-Preempt is not a hard real-
time approach in strict sense: processes can incur a latency
that is not deterministic and no guarantees are usually pro-
vided on the feasibility of a given task set. Although this
real-time extension to the Linux kernel suffers from the
above-mentioned limitations, it greatly improves the perfor-
mances of many applications and the responsiveness of the
whole system. Moreover, no special programming libraries
are required – the applications compiled for RT-Preempt
Linux can be used also on a standard, non real-time Linux
system.
ASSOLO could be easily ported to other real-time op-
erating systems, since it is written in C language and uses
standard system calls. However we decided to use the RT-
preempt approach, which makes the software much more
portable and easier to deploy and maintain over a large net-
work infrastructure.
3.3. Observations filtering
Intermediate routers can be heavily affected by similar
problems: interrupt coalescence, timer and clock resolution
and context-switching delays are all factors that can poten-
tially modify timings of the probe traffic, therefore intro-
ducing errors. Moreover, quite all existing tools assume the
hypothesis of fluid cross-traffic [18], ignoring the discrete
nature of packets. However this non-deterministic behav-
ior of intermediate nodes depends on the specific network
path and it can not be easily controlled or even described.
As a result, most of existing available bandwidth estimation
techniques produce noisy observations [9, 7].
Filtering techniques could be introduced in order to con-
vert noisy values into a reliable estimate. The idea of using
such a solution in this context is based on the predictabil-
ity and long-term stability of the Internet. Typically, the
available bandwidth of an Internet path shows strong cor-
relation and a certain degree of stability over intervals that
span from several minutes to a few hours [29, 27]. Given
a new observation, an effective filtering technique can pro-
duce a new estimate of the available bandwidth combining
both the most recent observation and the old values.
A vast majority of available bandwidth estimation tools
introduce filtering techniques: Moving Average [21], Ex-
ponential Weighted Moving Average (EWMA) [16, 23],
Wavelets [13] or Kalman filters [4, 3, 26] have been suc-
cessfully adopted to attenuate noise and local random fluc-
tuations.
Ideally, sharp and non-persistent changes can at first be
treated as noise and they should be ignored. However, if the
change persists, the available bandwidth is likely to have
changed and the filter should quickly converge to the new
measured value. In [8] we proposed Vertical Horizontal Fil-
ter (VHF), a new de-noising technique designed around this
principle and borrowed from the financial world [28]. VHF
is a modified EWMA filter which can dynamically modify
its behavior according to trends identified in the temporal
evolution of available bandwidth.
The simulations we performed revealed that linear filter-
ing techniques can effectively reduce the impact of noise
on the estimation of the available bandwidth. Compared to
filtering methods mentioned above, the VHF filter leads to
better results in many cases and shows greater stability. Our
experiments indicate also that there is no need to fine tune
the VHF filter every time some network conditions change.
A detailed description of VHF and a comparison between
different linear filtering techniques can be found in [8].
Results convinced us to adopt the Vertical Horizontal Fil-
ter inside our tool. ASSOLO does not output raw measures;
instead, it uses the VHF filter to cope with noisy observa-
tions and to estimate the actual available bandwidth.
3.4. Excursions segmentation
According to the basic principle of self-induced conges-
tion, increasing queuing delays indicates an instantaneous
rate higher than available bandwidth; otherwise, packets
sent by ASSOLO will experience no delays. This model
is valid also for tools which probe multiple rates with a sin-
gle train – the last instantaneous probing rate which does
not result in an increasing queuing delay is considered a
simple estimate of available bandwidth. However, this ap-
proach oversimplifies reality, lacking, for example, to con-
sider cross-traffic bursty behavior.
The typical profile of queuing delays in a train is of-
ten non-monotonic. For example, Fig. 2 shows the typical
queuing delays of a chirp sent by pathChirp: one or more
excursions produced by bursts return to zero, while a final
excursion ends with increasing queuing delays.
The authors of pathChirp introduced a smart segmen-
tation algorithm to cope with this kind of burstiness ef-
fects, detecting increasing delays belonging to a cross-
traffic bursty transient. A detailed description of the algo-
rithm can be found in [21]. ASSOLO adopts exactly the
same technique to analyze the queuing delays of each sin-
gle RE AC H and to identify the correct turning point.
Figure 2. Typical queuing delays in a chirp.
4. Results
In order to evaluate our estimation method, the perfor-
mance of ASSOLO has been studied in a controlled testbed
environment. In addition, we compared the intrusiveness
and the accuracy of our solution with pathChirp, a similar
state-of-the-art measurement tool.
The testbed configuration is shown in Fig. 3. Two
computers using Ubuntu GNU/Linux are connected to-
gether through a Fast Ethernet cross-cable and they serve as
routers; two other machines of the testbed simulate a source
of controlled traffic flows using the D-ITG tool [2]. Finally,
the sender and the receiver for each measurement tool use
additional PCs running Ubuntu GNU/Linux. Moreover, we
provisioned the network two Fast Ethernet switches to in-
troduce a source of errors. Prasad et al. in [19] showed that
each store-and-forward device introduces an additional se-
rialization latency in a packet’s delay, which results in con-
sistent underestimation of that hop’s capacity. We used the
default configurations for both probing tools. In addition,
results in [21] show that pathChirp generally performs bet-
ter with larger packets; therefore we set the packets size of
both tools to 1000 byte.
More details of the configuration of testbed’s devices and
tools are provided in [25].
Switch Switch
Router Router
Probe traffic
Cross traffic
ITG Sender ITG Receiver
Sender Receiver
Figure 3. Testbed configuration.
4.1. Intrusiveness
The intrusiveness of pathChirp and ASSOLO can be eas-
ily compared. From [7] we know that a chirp is composed
of Npackets, where Ncan be calculated as follows:
Nchirp =2 + 1
logγ log U
L.
The size of the stream sent by pathChirp depends on the
upper (U) and lower (L) rate bounds. However, pathChirp
automatically reduce or increase the probing range if it is
too wide or too narrow: as a result, the tool sends on average
15-20 packets.
On the other hand, the length of a REACH only depends
on the two parameter σand γ. In Section 3.1 we calculated
the size of a REAC H probe as:
Nreach = 2 ·logγγ−1
σ+ 1−1+ 2.
Hence, ASSOLO send always 18 packets using default val-
ues of σand γ.
Our experiments show the amount of traffic injected by
ASSOLO and pathChirp is comparable and extremely lim-
ited. Using default parameters on a 100 Mbps link, both
tools takes less than one second to produce an estimation
and the average probe traffic intensity is about 300 Kbps.
4.2. Accuracy
We tested both pathChirp and ASSOLO in the pres-
ence of different sources of Constant Bit Rate (CBR) cross-
traffic. Fig. 4 shows a measurement performed in our
testbed while the network path is loaded with a CBR cross-
traffic of 32 Mbps.
Figure 4. An example of estimation using
pathChirp and ASSOLO.
We generated CBR cross-traffic of 32, 16 and 8 Mbps
and, finally, we turned off the traffic source. We tested both
pathChirp and ASSOLO in each cross-traffic scenario, re-
peating the measurement process 10 times for each tool:
averaged results are shown in Fig. 5.
Our experiments show that pathChirp constantly over-
estimates available bandwidth and measurements are quite
unstable. This is well-know problem of pathChirp: similar
results have been obtained in [22, 4, 1]. On the other hand,
the accuracy and stability of ASSOLO is notable: we found
that 80% of estimations exhibit a relative error lower than
15%.
5. Conclusion
We presented ASSOLO, an active probing tool which
features an efficient measurement scheme for end-to-end
available bandwidth estimation. Moreover, the tool runs in-
side a real-time operating system and uses some de-noising
techniques to improve the estimation process. Testbed ex-
periments reveal that ASSOLO is non-intrusive and accu-
rate. In addition, ASSOLO outperforms pathChirp, a state-
of-the-art measurement tool, estimating available band-
width with greater accuracy and stability.
The testbed we used is quite simple and the synthetic
cross-traffic does not fully catch the complexity of actual
communication flows. We will test intensively the perfor-
mance of ASSOLO over Internet paths. We will also study
the impact on the measurement process of different cross-
traffic sources and real-world devices. An extensive com-
parison of ASSOLO with other existing tools is also needed.
Above all, BART and FEAT are recent tools which perform
better than the original pathChirp: a comparative study will
be conducted as soon as the code of these software will be
freely available.
Figure 5. Available bandwidth test results us-
ing pathChirp and ASSOLO
Since the bounds of ASSOLO’s probing interval have
to be set manually at start up, a coarse estimation of the
current available bandwidth is required prior using our tool.
We plan to introduce a self-configuring feature, avoiding the
needs for any prior knowledge of the network path. More-
over, the tool could automatically repeat the same tuning
process when the measured available bandwidth changes
significantly.
Additional implementation details and the source
code of ASSOLO are all available at http://netlab-
mn.unipv.it/assolo or through the authors. Future develop-
ments and data reports will be published at the same loca-
tion.
References
[1] A. A. Ali, F. Michaut, and F. Lepage. End-to-end available
bandwidth measurement tools : A comparative evaluation of
performances. ArXiv e-prints, June 2007.
[2] S. Avallone, S. Guadagno, D. Emma, A. Pescap`
e, and
G. Ventre. D-itg distributed internet traffic generator. In
QEST, pages 316–317. IEEE Computer Society, 2004.
[3] A. Cabellos-Aparicio, F. J. Garcia, and J. Domingo-Pascual.
A novel available bandwidth estimation and tracking algo-
rithm. In Network Operations and Management Symposium
Workshops, 2008. NOMS Workshops 2008. IEEE, pages 87–
94, Apr. 2008.
[4] S. Ekelin, M. Nilsson, E. Hartikainen, A. Johnsson, J.-E.
Mangs, B. Melander, and M. Bjorkman. Real-time mea-
surement of end-to-end available bandwidth using kalman
filtering. In NOMS 2006, 10th IEEE/IFIP, 2006.
[5] M. et al. RTAI: Realtime application interface for linux.
https://www.rtai.org/.
[6] P. G. et al. Xenomai: Real-time framework for linux.
http://www.xenomai.org.
[7] E. Goldoni. Nuovi approcci nella stima della banda disponi-
bile in una rete a pacchetto. Master’s thesis, Universit`
a degli
Studi di Pavia, 2007.
[8] E. Goldoni and G. F. Rossi. Improving available bandwidth
estimation using averaging filtering techniques. Technical
report, University of Pavia, 2008.
[9] C. D. Guerrero and M. A. Labrador. Experimental and an-
alytical evaluation of available bandwidth estimation tools.
In Local Computer Networks, Proceedings 2006 31st IEEE
Conference on, pages 710–717, Nov. 2006.
[10] N. Hu and P. Steenkiste. Evaluation and characterization
of available bandwidth probing techniques. IEEE Journal
on Selected Areas in Communications, 21(6):879–894, Aug.
2003.
[11] M. Jain and C. Dovrolis. Pathload: A measurement tool for
end-to-end available bandwidth. In Proceedings of 3rd PAM
Workshop, Fort Collins, CO, USA, March 2002.
[12] A. Johnsson, B. Melander, and M. Bj ¨
orkman. Diettopp:
A first implementation and evaluation of a simplified band-
width measurement method. In 2nd Swedish National Com-
puter Networking Workshop, Karlstad, November 2004.
[13] S.-R. Kang and D. Loguinov. IMR-Pathload: Robust
available bandwidth estimation under end-host interrupt de-
lay. In Passive and Active Network Measurement, volume
4979/2008 of Lecture Notes in Computer Science, pages
172–181. Springer Berlin / Heidelberg, 2008.
[14] B. Melander, M. Bjorkman, and P. Gunningberg. A new
end-to-end probing and analysis method for estimatingband-
width bottlenecks. In GLOBECOM ’00, San Francisco, CA,
USA, 2000.
[15] I. Molnar and T. Gleixner. Gnu/linux real-time wiki.
http://rt.wiki.kernel.org/.
[16] J. Navratil and R. L. Cottrell. Abwe: A practical approach to
available bandwidth. In Proceedings of 4th PAM Workshop,
San Diego, CA, USA, April 2003.
[17] Y. Ozturk and M. Kulkarni. Dichirp: direct injection band-
width estimation. Int. J. Netw. Manag., 18(5):377–394,
2008.
[18] R. Prasad, M. Murray, C. Dovrolis, and K. Claffy. Band-
width estimation: Metrics, measurement techniques, and
tools, 2003.
[19] R. S. Prasad, C. Dovrolis, and B. A. Mah. The effect of
layer-2 store-and-forward devices on per-hop capacity esti-
mation. In INFOCOM 2003. Twenty-Second Annual Joint
Conference of the IEEE Computer and Communications So-
cieties. IEEE, volume 3, pages 2090–2100, Mar./Apr. 2003.
[20] V. Ribeiro, M. Coates, R. Riedi, S. Sarvotham, B. Hen-
dricks, and R. Baraniuk. Multifractal cross-traffic estima-
tion. In Proc. of ITC Specialist Seminar on IP Traffic Mea-
surement, Sept. 2000.
[21] V. Ribeiro, R. Riedi, R. Baraniuk, J. Navratil, and L. Cot-
trell. pathchirp: Efficient available bandwidth estimation for
network paths. In Proceedings of 4th PAM Workshop, San
Diego, CA, USA, April 2003.
[22] A. Shriram, M. Murray, Y. Hyun, N. Brownlee, A. Broido,
and K. C. M. Fomenkov. Comparison of public end-to-end
bandwidth estimation tools on high-speed links. In Proceed-
ings of 6th PAM Workshop, Boston, MA, USA, March 2005.
[23] J. Sommers, P. Barford, and W. Willinger. A proposed
framework for calibration of available bandwidth estimation
tools. iscc, 0:709–718, 2006.
[24] J. Strauss, D. Katabi, and F. Kaashoek. A measurement
study of available bandwidth estimation tools. In IMC’03:
Proceedings of the 3rd ACM-SIGCOMM Conference on In-
ternet Measurement, New York, NY, USA, 2003.
[25] A. Torelli. Sviluppo di una tecnica innovativa per la stima
della banda disponibile. Master’s thesis, Universit`
a degli
Studi di Pavia, 2008.
[26] G. Urvoy-Keller, T. En-Najjary, and A. Sorniotti. Opera-
tional comparison of available bandwidth estimation tools.
SIGCOMM Comput. Commun. Rev., 38(1):39–42, 2008.
[27] Q. Wang and L. Cheng. FEAT: Improving accuracy in
end-to-end available bandwidth measurement. In Global
Telecommunications Conference GLOBECOM ’06, pages
1–4, Nov. 2006.
[28] A. White. The vertical horizontal filter. Futures Magazine,
Volume XX, No. 10:1–10, 1991.
[29] Y. Zhang and N. Duffield. On the constancy of internet path
properties. In IMW ’01: Proceedings of the 1st ACM SIG-
COMM Workshop on Internet Measurement, New York, NY,
USA, 2001.
[30] H. Zhou, Y. Wang, X. Wang, and X. Huai. Difficulties in
estimating available bandwidth. In Communications, 2006.
ICC ’06. IEEE International Conference on, June 2006.