Content uploaded by Emanuele Goldoni
Author content
All content in this area was uploaded by Emanuele Goldoni on Sep 25, 2018
Content may be subject to copyright.
ASSOLO: an Efficient Tool for Active End-to-end
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 a cru-
cial metric for bandwidth-dependent services such as multimedia
streaming, peer-to-peer and gaming applications; it is also useful
for quality of service verification and traffic engineering. This
paper presents the details of ASSOLO, an efficient active probing
tool for estimating the available bandwidth of a network path.
The tool is based on the well-known concept of “self-induced
congestion”, and it features a new probing traffic profile called
REACH (Reflected ExponentiAl Chirp) to test a wide range of
possible rates with a single stream of packets. In addition, the
program runs inside a real-time operating system and uses some
de-noising techniques to improve the measurement process. Ex-
perimental results show that ASSOLO outperforms pathChirp,
a state-of-the-art measurement tool, estimating the available
bandwidth with greater accuracy and stability in presence of
different cross-traffic sources. Moreover, we demonstrate that the
use of a real-time operating system can increase the stability of
the estimations lowering the impact of software context switches.
Keywords-Available bandwidth, active network measurement,
performance evaluation, real-time.
I. INT ROD UC TI ON
ASSOLO is a novel tool for available bandwidth estimation
in packet-switched networks which has been originally intro-
duced in [1]. This work extends some of the results presented
in the original paper by investigating the performance of our
tool in presence of poissonian cross-traffics. We also study
the actual impact of a real-time operating system on the
measurement process, and we provide more details on the
filtering technique implemented into the program.
The available bandwidth of a network path is a crucial
metric in quality-of-service management, traffic engineering
or congestion control. Voice over IP (VoIP), peer-to-peer and
video-streaming are examples of widely-used applications that
could greatly benefit from the knowledge of the available
bandwidth along an Internet path. For example, in [2] and
[3] the importance of the available bandwidth is investigated
respectively for peer-to-peer (P2P) networks and gaming-on-
demand services. In [4] the authors focus instead on improving
the perceived quality of video streaming through a dynamic
path selection based on the measurement of network-layer
metrics. Similarly, in [5] the authors propose a live broadcast
platform where the video source is distributed to a number
of clients organized in a peer-to-peer tree-structured overlay
network. In this network the root node is also responsible
for organizing and maintaining the position of each peer
within the tree according to the available bandwidth and the
latency between peers. The knowledge of the actual available
bandwidth is also exploited in [6] to improve video streaming
rate- and quality-adaptation decisions; results obtained through
simulations show that an estimation algorithm can substan-
tially increase streaming performance.
The same approach is also adopted in existing commercial
products: Microsoft Windows Media Server includes a tech-
nology called Intelligent Streaming for on-demand and live
media streaming over IP. This solution identifies the actual
maximum throughput allowed by the network path using a
end-to-end client/server system. This value is used to choose
the best encoding rate which maximizes the quality of received
media without overloading the network [7].
In principle, it would be possible to obtain estimates of the
available bandwidth directly from intermediate routers along
the network path; however, this is not feasible in practice due
to technical and security reasons. Therefore, researchers have
proposed several end-to-end measurement algorithms which
infer the network characteristic transmitting a few packets
and observing the effects of intermediate routers or links on
these probe frames. Examples of probing tools which have
emerged in recent years are IGI [8], Spruce [9], Pathload [10],
TOPP [11], [12], pathChirp [13], FEAT [14] and BART [15].
They differ mainly in the structure of probe streams and in
the algorithms used to estimate available bandwidth from the
received packets. Nevertheless, producing reliable estimations
in real-time still remains challenging: the measurement process
should be efficient, accurate, non-intrusive and robust at the
same time. Moreover, the algorithm should adaptively apply
to different types of networks and cross-traffics, and must be
able to produce fast periodic estimations in order to track
bandwidth fluctuations. As a result, as noted in [16]–[18],
current available bandwidth estimation techniques and tools
are far from being ready to be applied in many applications
and scenarios.
Compared to the tools mention above, our novel tool AS-
SOLO (Available-bandwidth Smart Sampling On-Line Tool)
features a new probing traffic profile called REACH (Reflected
ExponentiAl Chirp). A RE ACH tests a wide range of rates
283
International Journal
o
n Advances in Systems and Measurements
, vol 2
no
4
, year 200
9
,
http://www.iariajournals.org/systems_and_measurements/
and is more accurate in the center of the probed interval.
Moreover, ASSOLO uses a combination of new and existing
filtering techniques to improve the accuracy and stability of
results. Finally, our tool runs inside a real-time operating
system in order to minimize the impact of context switches
on the measurement process.
The rest of the paper is organized as follows. In Section
II we introduce the related work on available bandwidth
estimation and we focus on the Probe Gap Model, the general
measurement scheme adopted by our tool. Next, Section III
illustrates the algorithm used to generate the REAC H probing
stream and the additional features introduced in our tool. An
evaluation of ASSOLO is presented in Section IV, including
results obtained comparing our solution to the state-of-the-
art tool pathChirp both in terms of intrusiveness and accuracy.
Finally, in Section V we conclude and we outline future works.
II. RELATED WORK
Techniques for end-to-end available bandwidth estimation
can be divided into two categories: active probing and passive
measurements. The latter infer the required information from
existing data transmissions while active probing techniques
produce an estimation injecting dedicated probe traffic into
the network.
Passive measurements do not require dedicated packets to
perform the estimation: useful information is obtained from
traffic originated by active connections providing a particular
service. In this context, the idea of using TCP for network
measurements has attracted a lot of studies: RTT values [19]
or ACK arrival times [20], [21] have been used on the
sender’s side to infer the available bandwidth from existing
transmissions. These methods are lightweight and fast but they
can be applied only to network paths that have recently carried
traffic. Moreover, congestion control algorithms, buffers and
competing connections may influence the achievable through-
put of a single TCP connection, thus altering the accuracy of
estimations [22].
Active measurement techniques use probe packets to mea-
sure the end-to-end delays introduced by existing cross-traffic
(Figure 1). These methods require instrumentation at both
ends of the path; moreover, the probe traffic injected into the
network may affect the performance of other applications and
actually alter the available bandwidth. In addition, some tools
require a long measurement time and use hundred of packets
before producing an estimation. The majority of existing tools
belong to the Probe Gap Model (PGM) or the Probe Rate
Model (PRM).
In the Probe Gap Model, a tool sends a single probing
pair or train; it exploits then the dispersion of packets on
the receiver side to calculate the available bandwidth. The
main assumption of this model is that the link with the
minimum available bandwidth is also the link having the
minimum capacity. This is probably the biggest limit of this
approach: the hypothesis is not valid for many Internet paths
and can results in significant underestimations of the available
Fig. 1. The spacing effect on multiple traffics over a congested network
path.
bandwidth over multi-hop links [23]. Notable tools based on
the Probe Gap Model are Spruce [9], IGI [8] and Delphi [24].
Delphi [24] assumes a multi-fractal model for the cross-
traffic. The main idea in this tool is that the spacing of two
probing packets at the receiver can provide an estimate of the
amount of traffic at a link. Spruce [9] is based too on direct
probing and it uses tens of packet pairs to collect available
bandwidth estimations. The input rate of pairs is chosen to be
roughly around to the capacity of the path, which is assumed
to be known. Moreover, packets are spaced with exponential
intervals to emulate a poissonian sampling process. IGI [8]
uses a sequence of about 60 unevenly space packets to probe
the network and the gap between two consecutive packets is
increased until the average output and initial gaps match.
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 arrival
rate of packets at the receiver will not exhibit any notable
variation and it will match with the sender’s rate. On the other
hand, if the sending rate exceeds available bandwidth, 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 [11], [12] and Pathload [10] 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. Pathload on the other hand varies the probing
rate using a binary search scheme and the final output, 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.
PTR [8] is an active probing algorithm which sends several
probing packets to detect background traffic. The method com-
284
International Journal
o
n Advances in Systems and Measurements
, vol 2
no
4
, year 200
9
,
http://www.iariajournals.org/systems_and_measurements/
pares the time interval at the source with that of destination
and then uses the timings to estimate the value of available
bandwidth.
pathChirp [13] 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 associated rate is used to
infer the unused capacity. pathChirp can estimate available
bandwidth sending only one chirp: this feature makes the
measurement process fast and lightweight. However, path-
Chirp 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. Smoothed-chirp (S-chirp) is a similar approach based
on iterative probing and originally proposed in [25].
BART [15] relies on sequences of packet pairs sent at
randomized rates. This tool uses also a Kalman filter to track
the evolution of available bandwidth in real-time and to filter
out noisy observations. BART is lightweight, efficient and
non-intrusive; however, the tool is still in development and
it is not freely available. MR-BART is a extension of the
original BART method which employs multi-rate probe packet
sequences to achieve faster convergence and more accurate
estimations.
FEAT [14] is a recent tool which features a probe pattern
called fisheye stream. A fisheye stream consists of packets 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 sampling 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 the measurement process intrusive.
While BART and FEAT look quite promising, it is difficult
to compare them to other state-of-the-art tools: the results
presented by the respective authors have been obtained only
through simulations or using specific Internet paths, and to our
best knowledge the two programs have never been released
publicly.
III. ASSOLO
ASSOLO is an available bandwidth estimation tool which
has been originally presented in [1]. Unique to this tool is a
new probe traffic profile called REA CH (Reflected ExponentiAl
Chirp), which tests a wide range of rates using a single stream
of packets and injecting a negligible amount of traffic into the
network. The tool introduces also some techniques to minimize
the impact of different sources of errors on the estimation
process.
A. Probing stream
ASSOLO is based on the concept of “self-induced con-
gestion” – it tests different rates using a single stream of
packets, and then infers the available bandwidth harnessing
the information about the relative delays. This approach has
a twofold advantage: it requires 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 relative queuing delays within a stream.
The novel RE ACH probing traffic profile tests multiple rates
with a single stream, and it is more accurate at the center
of the stream, where the actual available bandwidth is likely
to be. A similar idea was originally proposed in [14], but our
method introduces a different spacing algorithm and sends less
packets. Compared to pathChirp, the stream used by ASSOLO
is different too – both tools use a sequence of packets with
increasing delays, but the shape of the traffic and the delays
within a stream are not the same.
The RE ACH stream used by our tool tests different rates
increasing the instantaneous 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 k
packets test values between Hand the maximum probing rate
U. However, the probing rates do not increase linearly in a
RE ACH. Instead, 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, the 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.
The maximum relative accuracy of ASSOLO’s estimations
is defined by the parameter σ. Given the probing range, the
absolute error Saround the center of the probing interval is
calculated as:
S=σU−L
2.(1)
Moreover, the algorithm uses a coefficient γto control how
fast the density of streams changes. This parameter reminds
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 should be
assigned according to the specific requirements of the target
application. Decreasing γand σ, the tool would send more
packets but it should result in a more accurate estimation; sim-
ilarly, increasing these value should reduce both intrusiveness
and accuracy.
An additional parameter ∆xis also needed to better describe
the RE ACH stream generated by ASSOLO. The function of
this auxiliary coefficient is to describe the gap between two
consecutive packets of the stream, and it is defined as follows:
∆x=S·γ|x−1|(2)
and combining Equations 1 and 2 we get:
∆x=σU−L
2γ|x−1|(3)
Starting from the center Hof the probing interval towards
the upper bound U, instantaneous packet rates in a RE ACH are
285
International Journal
o
n Advances in Systems and Measurements
, vol 2
no
4
, year 200
9
,
http://www.iariajournals.org/systems_and_measurements/
0 5 10 15
0
20
40
60
80
100
120
140
160
180
200
Probed rate
Rates [Mbps]
Reach Profile
Fig. 2. Distribution of packets in a REACH stream.
H, H + ∆1, H + ∆1+ ∆2, H + ∆1+ ∆2+ ∆3, .... A more
formal description of instantaneous probing rates Rxtested by
this stream is:
(Rx=H, if x= 1
Rx=Rx−1+ ∆x,∀x > 1, Rx< U (4)
On the other hand, probing rates from the center towards
the lower bound are H, H −∆1, H −∆1−∆2, H −∆1−
∆2−∆3, .... The instantaneous rates tested by a REAC H can
then be described as:
(Ry=H, if y= 1
Ry=Ry−1−∆y,∀y > 1, Ry> L (5)
The resulting stream is shown in Figure 2. As also the name
RE ACH (Reflected ExponentiAl CHirp) suggests, the profile is
symmetric: the right and the left part look like two mirrored
exponential functions.
Since the function is symmetric, we can analyze only the
right part of the stream – the same considerations would also
apply to the left one. ASSOLO uses kpackets to test values
between Hand the upper bound U. Thus, the instantaneous
rate Rkassociated 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(6)
If we substitute the values of ∆iand H, we get:
U−H=σU−L
2
k
X
i=1
γ|i−1|(7)
U−U+L
2=σU−L
2
k
X
i=1
γ|i−1|(8)
U−U+L
2
σU−L
2
=
k
X
i=1
γ|i−1|(9)
U−L
2
σU−L
2
=
k
X
i=1
γ|i−1|(10)
1
σ=
k
X
i=1
γ|i−1|(11)
In addition, the value of the truncated sum is:
k
X
i=1
γ|i−1|=γk+1 −1
γ−1(12)
Combining Equations 11 and 12, we get:
γk+1 =γ−1
σ+ 1 (13)
which leads to
k=logγγ−1
σ+ 1−1(14)
Actually we should define Rkas the maximum sending rate
not exceeding the upper bound U. Equation 6 should then take
the form:
U≥H+
k
X
i=1
∆i(15)
Hence the correct value of kis:
k=logγγ−1
σ+ 1−1(16)
As we mentioned before, a REAC 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 between Hand the maximum probing rate U. As
a result, a REAC H probes 2k+1 rates exploiting relative delays
between probe packets. Therefore, our tool needs to send an
additional packet at the beginning of the RE ACH. The total
number Nof packets used by ASSOLO to probe 2k+ 1 rates
is:
N= 1 + (2k+ 1) = 2k+ 2 (17)
Since we know the size of a RE ACH , we can also combine
Equations 4 and 5 and describe the rates probed by a RE ACH
profile as:
Rj=H+sign j−N
2·S·γ|j−N
2|−1
γ−1(18)
286
International Journal
o
n Advances in Systems and Measurements
, vol 2
no
4
, year 200
9
,
http://www.iariajournals.org/systems_and_measurements/
B. End-hosts predictability
A fundamental difficulty with the existing measurement
tools stems from a number of issues on both end-hosts and
network paths [26]: system timing, hardware errors and end-
to-end pathologies could produce a considerable amount of
noise in the individual network observations. For example, the
Linux kernel is a time sharing operating system designed to
give a fair share of the CPU in a multi-user environment [27] –
even some kernel services like memory allocation and system
calls exhibit some non-deterministic timing behavior.
Network measurement tools have strict operational dead-
lines between the arrive of a packet and the application’s
response to that event – the same can be said for the sender
side, where the packet sent by the application should ideally
start with no delay. In [28] 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 observations is due to the non-deterministic behavior of
the operating systems hosting the sender and the receiver. To
accommodate deadlines on both the end-hosts we decided to
use a real-time operating systems (RTOS), which can guaran-
tee predictability and accurate system timings for applications.
ASSOLO runs inside a GNU/Linux system with RT-Preempt
[29], [30] patch enabled, thus using a fully preemptible ker-
nel with high-resolution kernel timers. In order to minimize
the impact of context switches on the bandwidth estimation
process, the tool gets the highest priority on both the end-host
systems while probing the network.
Our program could be easily ported to other real-time
operating 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
network infrastructure.
Compared to other Linux real-time approaches, such as
RTAI [31] and Xenomai [32], 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 provided
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 performances
of many applications and the responsiveness of the whole
system, thus providing adequate service for most applications
that need real-time determinism [33]. Moreover, no special
programming libraries are required: the applications compiled
for RT-Preempt Linux can be also used on a standard, non
real-time Linux system with negligible adaptations.
C. Observations filtering
Like the end-hosts, also intermediate routers can be heavily
affected by predictability issues: interrupt coalescence, clock
resolution and context-switching delays are all factors that
can potentially modify timings of the probe traffic, therefore
introducing errors. Moreover, almost all existing tools assume
the hypothesis of fluid cross-traffic [34], ignoring the discrete
nature of packets. However this non-deterministic behavior 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 [35], [36].
A vast majority of available bandwidth estimation tools
introduce filtering techniques: for example, Moving Average,
Exponential Weighted Moving Average (EWMA), Wavelets or
Kalman filters have been successfully adopted in [13], [15],
[37]–[40] to attenuate noise and local random fluctuations,
converting noisy values into a reliable estimate.
The idea of using such a solution in this context is based
on the predictability and long-term stability of the Internet.
Typically, the available bandwidth of an Internet path shows
strong correlation and a certain degree of stability over inter-
vals that span from several minutes to a few hours [14], [41].
Given a new observation, an effective filtering technique can
produce a new estimate of the available bandwidth combining
both the most recent observation and the old values.
For example, the Exponentially Weighted Moving Average
(EWMA) filter uses one or more observed values Okand
outputs a new estimation Eicalculated as follows:
Ei=αEi−1+ (1 −α)Oi. (19)
This filter is used by some estimation tool like Abing [37] and
Yaz [38]. However, the difficulty with the EWMA technique
lies in the choice of the exponential weight α. With large
values of α, the old estimates are given more importance and
the filter is slow but stable; agility is instead achieved by
keeping αsmall. Ideally, the filter should be adaptive, setting
the value of αaccording to the current circumstances: sharp
and non-persistent changes can at first be treated as noise
using lower weights αi. However, if the change persists, the
filter should quickly converge to the new value. Equation (19)
should then take the form:
Ei=αiEi−1+ (1 −αi)Oi. (20)
Lowpass EMA [42], Stability [43] and Error Based Filters [43]
are three existing techniques designed around this philosophy.
Although they have been proposed a couple of years ago, to
our best knowledge none of them has actively been employed
in an available bandwidth estimation tool.
In [44] we originally proposed the use of Vertical Horizontal
Filter (VHF) in such a context. The VHF filter is a modified
EWMA technique borrowed from the financial world [45]
which can dynamically modify its behavior according to trends
identified in the temporal evolution of available bandwidth
according to the same principles of the three above mentioned
filters. The dynamically exponential weight αiin (20) is
computed as:
αi=β∆max
Pi
t=i−M|Ot−Ot−1|(21)
where ∆max is the gap between the maximum and the
minimum values in the Mmost recent observations. We set β
as 1
3and the window size M= 10, although these parameters
287
International Journal
o
n Advances in Systems and Measurements
, vol 2
no
4
, year 200
9
,
http://www.iariajournals.org/systems_and_measurements/
were obtained empirically and a careful choice could bring
further improvements.
We performed a series of simulations to investigate the
effectiveness of different filtering techniques on the avail-
able bandwidth estimation process. Compared to the methods
mentioned above, we found that the VHF filter leads to
better results in many cases and shows greater stability. Our
experiments also indicated 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 [44].
Results persuaded us to employ the Vertical Horizontal
Filter, which is used inside our tool to cope with noisy
observations and to estimate the actual available bandwidth
from raw measurements.
D. Excursions segmentation
According to the basic principle of PRM’s self-induced
congestion, an instantaneous sending rate higher than the
actual available bandwidth results in increasing queuing delays
at receiver; otherwise, packets sent by a tool will experience
no delays. This model is valid also for tools which probe
multiple rates with a single train, like ASSOLO does – the
last instantaneous probing rate which does not result in an
increasing queuing delay is considered a simple estimate of
available bandwidth. However, this approach oversimplifies
reality, lacking, for example, to consider cross-traffic bursty
behavior and end-host interrupt coalescence effects.
Traditional network adapters generate an interrupt for each
received frame, thus generating up to thousands of internal
signals per second in high-speed networks. These interrupts
consume a lot of system’s resources and introduce a significant
amount of context switches, resulting in a CPU overhead. [46]
To mitigate the effects of this issue, some network adapters
recently introduced the support for Interrupt Coalescence (IC)
[47]. This solution decreases the processing overhead buffering
multiple packets before generating a single interrupt for the
burst of frames. A similar approach has been introduced
in NAPI [48], a modification to the device driver packet
processing framework of Linux kernel. NAPI mixes interrupts
with a polling approach to implement an adaptive interrupt
coalescing which modifies its behavior according to the actual
network load. This solution usually results in improved per-
formances for high-speed networking. Although IC decreases
the per-packet processing overhead, it introduces also non-
deterministic queueing delays, thus altering the time spacing
of packets in a probing train. As noted in [49], IC can be
detrimental to TCP self-clocking making the traffic more
bursty, and it has a negative effect on the accuracy of active
and passive bandwidth measurements.
The typical profile of queuing delays in a train is often non-
monotonic. For example, Figure 3 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.
Fig. 3. Typical queuing delays in a chirp.
function EXC UR SI ON(q, i, F, L)
j←i+ 1
qmax ←0
while (j≤N)AND (qj−qi> qmax/F )do
j++ ⊲Count excursion’s packets
end while
if j≥Nthen
return j ⊲ Non-ending Excursion
end if
if j−1≥Lthen
return j ⊲ Excursion
else
return i ⊲ Not an excursion
end if
end function
Fig. 4. Pseudo-code for the pathChirp’s excursion segmentation algorithm
[13].
The authors of pathChirp introduced a smart segmentation
algorithm to cope with this kind of burstiness effects, detecting
increasing delays belonging to a cross-traffic bursty transient.
The main goal of pathChirp’s excursion segmentation algo-
rithm is to identify potential starting and ending packet iand j
respectively for an excursion. Potentially, every packet iwhere
queueing delay qistarts increasing could be a starting point
of an excursion. We define the end of the excursion as the
point where the queuing delay returns to zero or where it has
decreased by a factor Ffrom the maximum queueing delay
experienced during this interval. Moreover, if the distance
between these two packet is long enough, for example longer
than a threshold L, then all packets between iand jform
an excursion. On the other hand, the last excursion identifies
the congested region and it does not terminate. The pseudo-
code of the procedure is presented in Figure 4 while a detailed
description of the whole algorithm can be found in the original
paper of pathChirp [13]. Since this solution proved to be quite
effective to cope with burstiness, ASSOLO adopts exactly the
same technique to analyze the queuing delays of each single
RE ACH and to identify the correct turning point.
288
International Journal
o
n Advances in Systems and Measurements
, vol 2
no
4
, year 200
9
,
http://www.iariajournals.org/systems_and_measurements/
E. Availability
Additional implementation details and a copy of the source
code of ASSOLO are all freely available at http://netlab-
mn.unipv.it/assolo/ or through the authors. Future develop-
ments and data reports will be published at the same location.
IV. RES ULT S
In order to evaluate our estimation method, the performance
of ASSOLO has been studied in a controlled testbed envi-
ronment. In addition, we compared the intrusiveness and the
accuracy of our solution with pathChirp, a similar state-of-the-
art measurement tool, in presence of poissonian or constant bit
rate (CBR) cross-traffics.
The testbed configuration is shown in Figure 5. Two
computers using Ubuntu GNU/Linux are connected together
through a Fast Ethernet cross-cable and serve as routers. Two
other machines of the testbed simulate a source of controlled
traffic flows using the D-ITG tool [50], which loads the
network generating synthetic flows of known properties and
statistical distributions. Finally, the sender and the receiver for
each measurement tool use additional PCs running Ubuntu
GNU/Linux with a standard or real-time kernel. Prasad et al.
in [51] showed that each store-and-forward device introduces
an additional serialization latency in a packet’s delay. This can
result in a consistent underestimation of the hop’s capacity.
Therefore, we provisioned the network with two Fast Ethernet
switches in order to introduce an additional potential source
of errors during tests.
The topology of the testbed is quite simple but sufficient to
evaluate the performance of a measurement tool: for example,
the same configuration has been used in [52] and [17] to
perform an experimental comparison of different available
bandwidth estimation tools.
We adopted the default configurations for both probing
tools: ASSOLO uses σ= 5% and γto 1.2while the γof
pathChirp has been initially set to 1.2. Since results obtained
in [13] showed that pathChirp generally performs better with
larger packets, the packet size for both tools was 1000 byte.
Finally, the upper and the lower bandwidth bounds Uand L
were respectively equal to 200 and 10 Mbps; however, both
tools automatically adjust the values if the range is too narrow.
A complete list of all the configuration parameters of testbed’s
devices and tools is provided in [53].
A. Intrusiveness
The intrusiveness of pathChirp and ASSOLO can be easily
compared. From [36] we know that a chirp is composed of N
packets, 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 reduces or increases the probing range if it is
too wide or too narrow: as a result, the tool sends on average
15-20 packets.
Fig. 5. Testbed configuration.
On the other hand, the length of a RE ACH only depends on
the two parameter σand γ. In section III-A we calculated the
size of a REAC H probe as:
Nreach = 2 ·logγγ−1
σ+ 1−1+ 2.
Hence, our algorithm send always 18 packets using default
values of σand γ.
Our experiments show that the amount of traffic injected by
ASSOLO and pathChirp is comparable and extremely limited.
Using the default parameters, the measurement process of both
tools takes less than one second to produce an estimation over
links with a capacity higher than 1 Mbps. However, the two
methods are based on the concept of self-induced congestion,
i.e., the estimation is performed by injecting probe traffic at a
rate higher than the available bandwidth of the network path.
The drawback of this approach is that the bottleneck node is
congested by the probe traffic – the existing cross-traffic is
delayed, and its packets’ timings can be significantly affected
by the measurement process.
B. Accuracy
We tested both pathChirp and ASSOLO in the presence
of different sources of cross-traffic with varying intensity. We
generated CBR cross-traffic of 64, 32 and 16 Mbps and, finally,
we turned off the traffic source. We evaluated both tools in
each cross-traffic scenario, repeating the measurement process
10 times for each algorithm: averaged results are shown in
Figure 6. Then, we repeated the same tests simulating different
sources of poissonian cross-traffic with increasing average
traffic load. The results obtained after 10 runs are shown in
Figure 7.
Our experiments show that pathChirp constantly overesti-
mates available bandwidth and measurements are quite un-
stable. This is a well-know problem of pathChirp: similar
results have been obtained in [15], [17], [54]. On the other
hand, we found that 80% of ASSOLO’s estimations exhibit a
relative error lower than 15%. Figure 8 shows an example of
289
International Journal
o
n Advances in Systems and Measurements
, vol 2
no
4
, year 200
9
,
http://www.iariajournals.org/systems_and_measurements/
0 10 20 30 40 50 60 70
0
20
40
60
80
100
120
CBR Cross−Traffic [Mbps]
Available Bandwith [Mbps]
PathChirp Measurements
0 10 20 30 40 50 60 70
0
20
40
60
80
100
120
CBR Cross−Traffic [Mbps]
Available Bandwith [Mbps]
ASSOLO Measurements
Fig. 6. Measurements obtained in presence of different CBR cross-traffics
measurement performed in our testbed while the network path
is loaded with a Constant Bit Rate cross-traffic of 32 Mbps:
the difference between the two tools is notable both in terms
of accuracy and stability.
It is worthy of remark that the accuracy of the two tools
does not seem to depend on the nature of the cross-traffic
– the performances are almost identical using either a CBR
source or poissonian distributed packets.
C. Stability
We have analyzed the impact of a real-time operating system
on the ASSOLO’s measurement process. We performed a
few tests with the real-time feature enabled and then we
disabled it before repeating the estimation procedure with our
tool. A sample comparison of the measurements obtained in
the two cases is shown in Figure 9: the average value is
correct in both configurations but the real-time feature provides
much more stability. Although more investigations would be
required, preliminary results confirm that the use of a real-time
0 10 20 30 40 50 60 70
0
20
40
60
80
100
120
Poisson Cross−Traffic [Mbps]
Available Bandwith [Mbps]
PathChirp Measurements
0 10 20 30 40 50 60 70
0
20
40
60
80
100
120
Poisson Cross−Traffic [Mbps]
Available Bandwith [Mbps]
ASSOLO Measurements
Fig. 7. Measurements obtained in presence of different poissonian cross-
traffics
environment can effectively reduce the impact of different non-
deterministic sources of error.
The same experiments could also be repeated for a longer
observation interval, in order to catch possible long-term
oscillations or biases in the estimations obtained with a non
real-time system.
V. CO NC LU SI ON A ND FU TU RE WO RK
In this work we presented the details of ASSOLO, an
active probing tool which features an efficient measurement
scheme for end-to-end available bandwidth estimation in
packet-switched networks. Moreover, we described some de-
noising techniques and detailed the real-time operating system
used by our tool to improve the estimation process.
Preliminary experiments revealed that our algorithm is non-
intrusive and accurate, estimating the available bandwidth with
greater accuracy and stability with respect to the pathChirp
measurement tool developed by the Rice university.
290
International Journal
o
n Advances in Systems and Measurements
, vol 2
no
4
, year 200
9
,
http://www.iariajournals.org/systems_and_measurements/
0 10 20 30 40 50
0
20
40
60
80
100
Time [s]
Estimated Available Bandwidth [Mbps]
Real
ASSOLO
PathChirp
Fig. 8. An example of estimation using pathChirp and ASSOLO.
10 15 20 25 30 35 40
0
20
40
60
80
100
Time [s]
Estimated Available Bandwidth [Mbps]
Avail−BW
Real−Time
Non Real−Time
Fig. 9. Measurements using either a real-time operating system or not.
The testbed we used is quite simple and the synthetic cross-
traffic does not fully catch the complexity of actual commu-
nication flows. We plan to test intensively the performance
of our tool over actual Internet paths and in presence of
realistic cross-traffic traces. We will also include a study of the
actual accuracy, intrusiveness and robustness when dynamic
traffic patterns are presents. An extensive comparison of our
approach with other state-of-the-art tools is needed too. Above
all, BART and FEAT are recent tools which seem to perform
better than the original pathChirp: a comparative study will be
conducted as soon as the code of these software will be freely
available.
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 an initial self-configuring feature as proposed in
[55], thus avoiding the need for any prior knowledge of the
network path.
ACK NOW LE DG ME NT
We would like to thank Dr. Davide Cavalca for giving the
paper a critical reading and for providing us several helpful
comments. We acknowledge also Dr. Alberto Savioli and
Marco Schivi for their help during the setup of the laboratory
testbed and the analysis of experimental results.
REF ER EN CE S
[1] E. Goldoni, G. Rossi, and A. Torelli, “Assolo, a new method for available
bandwidth estimation,” in Proc. IARIA International Conference on
Internet Monitoring and Protection (ICIMP 2009), May 2009, pp. 130–
136.
[2] C. Wu, B. Li, and S. Zhao, “Characterizing peer-to-peer streaming
flows,” IEEE Journal on Selected Areas in Communications, vol. 25,
no. 9, pp. 1612–1626, December 2007.
[3] J. P. Laulajainen, T. Sutinen, and S. Jarvinen, “Experiments with QOS-
aware gaming-on-demand service,” in IEEE International Conference
on Advanced Information Networking and Applications (AINA 2006),
vol. 1, Apr. 2006, pp. 805–810.
[4] M. Jain and C. Dovrolis, “Path selection using available bandwidth
estimation in overlay-based video streaming,” Computer Networks,
vol. 52, no. 12, pp. 2411–2418, August 2008.
[5] M. Favalli, L.and Folli, A. Lombardo, D. Reforgiato, and G. Schembra,
“A bandwidth-aware p2p platform for the transmission of multipoint
multiple description video streams,” in Proc. Italian Networking Work-
shop Reti.it 2009, Jan. 2009.
[6] T. Tunali and K. Anar, “Adaptive available bandwidth estimation for
internet video streaming,” Signal Processing: Image Communication,
vol. 21, no. 3, pp. 217–234, March 2006.
[7] M. Topic, Streaming Media Demystified. New York, NY: McGraw-Hill
Professional, 2002.
[8] N. Hu and P. Steenkiste, “Evaluation and characterization of available
bandwidth probing techniques,” IEEE Journal on Selected Areas in
Communications, vol. 21, no. 6, pp. 879–894, August 2003.
[9] J. Strauss, D. Katabi, and F. Kaashoek, “A measurement study of
available bandwidth estimation tools,” in Proc. ACM SIGCOMM
Conference on Internet Measurement (IMC’03), pp. 39–44, October
2003.
[10] M. Jain and C. Dovrolis, “Pathload: A measurement tool for end-to-
end available bandwidth,” in Proc. Passive and Active Measurement
Conference (PAM 2002), Mar. 2002, pp. 14–25.
[11] B. Melander, M. Bjorkman, and P. Gunningberg, “A new end-to-end
probing and analysis method for estimating bandwidth bottlenecks,” in
Proc. IEEE Global Communications Conference (GLOBECOM 2000),
Nov. 2000, pp. 415–420.
[12] A. Johnsson, B. Melander, and M. Björkman, “Diettopp: A first
implementation and evaluation of a simplified bandwidth measurement
method,” in Proc. Swedish National Computer Networking Workshop
(SNCNW 2004), Nov. 2004.
[13] V. Ribeiro, R. Riedi, R. Baraniuk, J. Navratil, and L. Cottrell, “path-
Chirp: Efficient available bandwidth estimation for network paths,” in
Proc. Passive and Active Measurement Conference (PAM 2003), Apr.
2003.
[14] Q. Wang and L. Cheng, “FEAT: Improving accuracy in end-to-end avail-
able bandwidth measurement,” in Proc. IEEE Global Communications
Conference (GLOBECOM 2006), Nov. 2006, pp. 1–4.
[15] 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 Proc. IEEE/IFIP
Network Operations and Management Symposium (NOMS 2006), Apr.
2006, pp. 73–84.
[16] M. Jain and C. Dovrolis, “Ten fallacies and pitfalls on end-to-end
available bandwidth estimation,” in Proc. ACM SIGCOMM Conference
on Internet Measurement (IMC’04). Oct. 2004, pp. 272–277.
[17] A. A. Ali, F. Michaut, and F. Lepage, “End-to-end available bandwidth
measurement tools : A comparative evaluation of performances,” in
Proc. International Workshop on Internet Performance, Simulation,
Monitoring and Measurement (IPS-MoMe 2006), Feb. 2006.
[18] C. D. Guerrero and M. A. Labrador, “On the applicability of available
bandwidth estimation techniques and tools,” in Computer Communica-
tions, vol. 33, no. 1, pp. 11–22, January 2010.
291
International Journal
o
n Advances in Systems and Measurements
, vol 2
no
4
, year 200
9
,
http://www.iariajournals.org/systems_and_measurements/
[19] L. S. Brakmo, S. W. O’Malley, and L. L. Peterson, “Tcp vegas: new
techniques for congestion detection and avoidance,” in Proc. ACM
SIGCOMM Conference on Communications Architectures, Protocols
and Applications (SIGCOMM’94). Aug. 1994, pp. 24–35.
[20] M. Gerla, M. Y. Sanadidi, R. Wang, A. Zanella, C. Casetti, and
S. Mascolo, “TCP westwood: congestion window control using band-
width estimation,” in Proc. IEEE Global Communications Conference
(GLOBECOM 2001), Nov. 2001, vol. 3, pp. 1698–1702.
[21] C. L. T. Man, G. Hasegawa, and M. Murata, “A new available band-
width measurement technique for service overlay networks,” in Proc.
IEEE/IFIP International Conference on Management of Multimedia
Networks and Services (MMNS 2003), Sep. 2003, pp. 436–448.
[22] M. Jain and C. Dovrolis, “End-to-end available bandwidth: measurement
methodology, dynamics, and relation with tcp throughput,” IEEE/ACM
Transactions on Networking, vol. 11, no. 4, pp. 537–549, August 2003.
[23] L. Lao, C. Dovrolis, and M. Y. Sanadidi, “The probe gap model
can underestimate the available bandwidth of multihop paths,” ACM
SIGCOMM Computer Communication Review, vol. 36, no. 5, pp. 29–
34, October 2006.
[24] V. Ribeiro, M. Coates, R. Riedi, S. Sarvotham, B. Hendricks, and
R. Baraniuk, “Multifractal cross-traffic estimation,” in Proc. ITC
Specialist Seminar on IP Traffic Measurement, Sep. 2000.
[25] A. Pasztor, “Accurate active measurement in the internet and its
applications,” Ph.D. dissertation, Department of Electrical and Electronic
Engineering, The University of Melbourne, 2003.
[26] H. Zhou, Y. Wang, X. Wang, and X. Huai, “Difficulties in estimating
available bandwidth,” in Proc. IEEE International Conference on
Communications (ICC’06), Jun. 2006, pp. 704–709
[27] The Linux Kernel. [Online]. Available: http://www.kernel.org
[28] Y. Ozturk and M. Kulkarni, “Dichirp: direct injection bandwidth
estimation,” International Journal of Network Management, vol. 18,
no. 5, pp. 377–394, September 2008.
[29] Gnu/Linux Real-Time. [Online] Available: http://rt.wiki.kernel.org/
[30] K. Koolwal, “Myths and realities of real-time linux software systems,”
in Proc. Real-Time Linux Workshop (RTLWS 2009), Oct. 2009. [Online].
Available: http://lwn.net/images/conf/rtlws11/papers/proc/p20.pdf
[31] RTAI: Realtime application interface for linux. [Online] Available: http:
//www.rtai.org/
[32] Xenomai: Real-time framework for Linux. [Online] Available: http:
//www.xenomai.org
[33] K. Yaghmour, J. Masters, G. Ben-Yossef, and P. Gerum, Building
Embedded Linux Systems. Sebastopol, CA: O’Reilly & Associates, 2008.
[34] R. Prasad, M. Murray, C. Dovrolis, and K. Claffy, “Bandwidth
estimation: Metrics, measurement techniques, and tools,” IEEE Network,
vol. 17, no. 6, pp. 27–35, November 2003.
[35] C. D. Guerrero and M. A. Labrador, “Experimental and analytical
evaluation of available bandwidth estimation tools,” in Proc. IEEE
Conference on Local Computer Networks (LCN 2006), Nov. 2006, pp.
710–717.
[36] E. Goldoni, “Nuovi approcci nella stima della banda disponibile in una
rete a pacchetto,” Master thesis, University of Pavia, 2007.
[37] J. Navratil and R. L. Cottrell, “Abwe: A practical approach to available
bandwidth,” in Proc. Passive and Active Measurement Conference (PAM
2003), Apr. 2003.
[38] J. Sommers, P. Barford, and W. Willinger, “A proposed framework
for calibration of available bandwidth estimation tools,” in Proc. IEEE
Symposium on Computers and Communications (ISCC’06), Jun. 2006,
pp. 709–718.
[39] S.-R. Kang and D. Loguinov, “IMR-Pathload: Robust available band-
width estimation under end-host interrupt delay,” in Proc. Passive and
Active Measurement Conference (PAM 2008), Apr. 2008, pp. 172–181.
[40] G. Urvoy-Keller, T. En-Najjary, and A. Sorniotti, “Operational com-
parison of available bandwidth estimation tools,” ACM SIGCOMM
Computer Communication Review, vol. 38, no. 1, pp. 39–42, January
2008.
[41] Y. Zhang and N. Duffield, “On the constancy of Internet path prop-
erties,” in Proc. ACM SIGCOMM Workshop on Internet Measurement
(IMW’01), Nov. 2001, pp. 197-211.
[42] L. Burgstahler and M. Neubauer, “New modifications of the exponential
moving average algorithm for bandwidth estimation,” in Proc. ITC Spe-
cialist Seminar on Internet Traffic Engineering and Traffic Management,
July 2002.
[43] M. Kim and B. Noble, “Mobile network estimation,” in Proc. Interna-
tional conference on Mobile Computing and networking (MobiCom’01),
Jul. 2001, pp. 298–309.
[44] E. Goldoni, G. F. Rossi, and P. Gamba, “Improving available bandwidth
estimation using averaging filtering techniques,” University of Pavia,
Tech. Rep., 2008. [Online]. Available: netlab-mn.unipv.it/publications/
tr-netlab2008-01.pdf
[45] A. White, “The vertical horizontal filter,” Futures Magazine, vol. 20,
no. 10, pp. 1–10, 1991.
[46] J. C. Mogul and K. K. Ramakrishnan, “Eliminating receive livelock in
an interrupt-driven kernel,” ACM Transactions on Computer Systems,
vol. 15, no. 3, pp. 217–252, August 1997.
[47] Intel. (2003) Interrupt moderation using Intel gigabit ethernet controllers.
[Online]. Available: http://download.intel.com/design/network/applnots/
ap450.pdf
[48] J. H. Salim, R. Olsson, and A. Kuznetsov, “Beyond softnet,” in Proc.
USENIX Annual Linux Showcase & Conference (ALC’01). Nov. 2001,
pp. 165-172.
[49] R. Prasad, M. Jain, and C. Dovrolis, “Effects of interrupt coalescence
on network measurements,” in Proc. Passive and Active Measurement
Conference (PAM 2004), Apr. 2004, pp. 247–256.
[50] S. Avallone, S. Guadagno, D. Emma, A. Pescapè, and G. Ventre, “D-
ITG distributed internet traffic generator.” in QEST. IEEE Computer
Society, 2004, pp. 316–317.
A. Botta, A. Dainotti, A. Pescapè, “Multi-protocol and multi-platform
traffic generation and measurement,” in IEEE Conference on Computer
Communications (INFOCOM 2007), Demo Session, May 2007.
[51] R. S. Prasad, C. Dovrolis, and B. A. Mah, “The effect of layer-2
store-and-forward devices on per-hop capacity estimation,” in IEEE
Conference on Computer Communications (INFOCOM 2003), Mar.
2003, vol. 3, pp. 2090–2100.
[52] L. Angrisani, S. D’Antonio, M. Esposito, and M. Vadursi, “Techniques
for available bandwidth measurement in ip networks: a performance
comparison,” Computer Networks, vol. 50, no. 3, pp. 332–349, February
2006.
[53] A. Torelli, “Sviluppo di una tecnica innovativa per la stima della banda
disponibile,” Master thesis, University of Pavia, 2008.
[54] 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 Proc. Passive and Active Measurement
Conference (PAM 2005), Mar. 2005, pp. 306–320
[55] W. Tan, M. Zhanikeev, and Y. Tanaka, “Abshoot: A reliable and efficient
scheme for end-to-end available bandwidth measurement,” in Proc.
IEEE Region 10 Conference (TENCON 2006), Nov. 2006.
292
International Journal
o
n Advances in Systems and Measurements
, vol 2
no
4
, year 200
9
,
http://www.iariajournals.org/systems_and_measurements/
