See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3335135

Using partial differential equations to model TCP mice and elephants in large
IP networks

Article in IEEE/ACM Transactions on Networking · January 2006

DOI: 10.1109/TNET.2005.860102 · Source: IEEE Xplore

CITATIONS READS
56 128

6 authors, including:

Michele Garetto Emilio Leonardi

Università degli Studi di Torino Politecnico di Torino
106 PUBLICATIONS 3,324 CITATIONS 279 PUBLICATIONS 5,075 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

NAPA-WINE EU FP7 View project

Protocol-Independent Switch Fabric View project

All content following this page was uploaded by Emilio Leonardi on 02 December 2014.

The user has requested enhancement of the downloaded file.

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 13, NO. 6, DECEMBER 2005 1289

Using Partial Differential Equations to Model TCP

Mice and Elephants in Large IP Networks
Marco Ajmone Marsan, Fellow, IEEE, Michele Garetto, Member, IEEE, Paolo Giaccone, Member, IEEE,
Emilio Leonardi, Member, IEEE, Enrico Schiattarella, Student Member, IEEE, and
Alessandro Tarello, Student Member, IEEE

Abstract—In this paper we propose a new fluid model approach In this paper, we develop a fluid model approach in which
in which a different description of the dynamics of traffic sources the sources dynamics are described by partial differential equa-
is adopted, exploiting partial differential equations. This new de- tions. This new description of the source dynamics allows the
scription of the source dynamics allows the natural representation
of short-lived as well as long-lived TCP connections, with no sac- natural representation of TCP mice as well as elephants, with
rifice in the scalability of the model. In addition, the use of partial no sacrifice in the scalability of the model. In addition, the use
differential equations permits the description of distributions, in- of partial differential equations permits the description of TCP
stead of averages, thus providing better accuracy in the results. window distributions, instead of averages, thus providing better
The comparison between the performance estimates obtained accuracy in the performance predictions.
with fluid models and with ns-2 simulations proves the accuracy
of the proposed modeling approach. The rest of this paper is organized as follows. Section II
overviews the fluid model of IP networks originally proposed
Index Terms—Computer network performance, differential by Misra, Gong, and Towsley [3]–[5], and Section III discusses
equations, modeling, transport protocol, wide-area networks.
other previous works in the same area. Section IV describes
the modeling methodology that we propose in this paper, based
I. INTRODUCTION on partial differential equations, for the case of TCP elephants.
We first introduce the simplest version of our model, and then
A NEW CLASS of semi-analytical models has recently been
introduced in the networking arena, and today appears to
be the most promising approach for scalable and accurate per-
progressively extend it to cope with finite window sizes, fast
recovery, and drop-tail buffers. Results are shown along the
formance analysis of large IP networks. These new models, way, and compared with performance estimates generated by
often called fluid models, adopt a deterministic description of ns-2 simulations, so as to prove the accuracy of the proposed
the average network dynamics through a set of differential equa- fluid model approach. In Section V, we move on to considering
tions [1]–[5], thus neglecting the detailed, packet-by-packet de- the case of TCP mice, generalizing the equations adopted for
scription of the stochastic network dynamics. The resulting set TCP elephants; we point out the fundamental limitation of a
of differential equations is then solved numerically, obtaining deterministic approach to describe the network behavior, and
estimates of the time-dependent network behavior. suggest solutions to overcome this limitation. Numerical and
The most attractive property of fluid models resides in the fact simulation results are then presented and discussed for different
that their complexity (i.e., the number of differential equations network scenarios. Finally, Section VI concludes the paper.
to be solved) is independent of the number of TCP flows and of
II. MGT FLUID MODEL OF IP NETWORKS
link capacities, when considering traffic scenarios comprising
only long-lived TCP flows (commonly called elephants). In ad- In [3]–[5], Misra, Gong, and Towsley presented simple dif-
dition, fluid models have been recently proved to capture the ferential equations to describe the behavior of TCP elephants
limiting behavior of TCP elephants in single bottleneck topolo- over networks of IP routers adopting a RED (Random Early De-
gies when the number of TCP flows, the bottleneck capacity and tection [10]) active queue management (AQM) scheme. Their
the buffer size jointly grow to infinity [1], [6]–[9]. approach (that we call MGT) spurred several research efforts
An important limit of the fluid model approaches presented aiming at the application of various kinds of fluid models to the
so far in the literature is their poor representation of scenarios performance analysis of packet networks. It is important to note
comprising the short-lived TCP flows (commonly called mice), that the equations of the MGT model heavily rely on the assump-
which are the majority of the flows in the Internet. tions mentioned above (all TCP connections are elephants, and
all IP routers adopt RED), and that the extension to mice and
Manuscript received April 23, 2004; revised January 24, 2005; approved by drop-tail routers may be not simple.
IEEE/ACM TRANSACTIONS ON NETWORKING Editor R. Srikant. This work was Consider a network comprising router output interfaces,
supported by the Italian Ministry for Education, University and Research, within equipped with FIFO buffers, and interfacing data channels at
the TANGO FIRB project. A preliminary version of this paper was presented at
the IEEE INFOCOM 2004, Hong Kong. rate (the extension to nonhomogeneous data rates is straight-
The authors are with the Dipartimento di Elettronica, Politecnico di forward). The network is fed by classes of long-lived TCP
Torino, 10129 Torino, Italy (e-mail: [email protected]; garetto@ flows; all the elephants within the same class follow the same
mail.tlc.polito.it; [email protected]; [email protected];
[email protected]; [email protected]). route through the network, thus experiencing the same round-
Digital Object Identifier 10.1109/TNET.2005.860102 trip time (RTT), and the same average loss probability (ALP).
1063-6692/$20.00 © 2005 IEEE
1290 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 13, NO. 6, DECEMBER 2005

At time all buffers are assumed to be empty. Buffers drop perceived by elephants of class satisfies the following
packets according to their average occupancy, as dictated by a expression:
RED AQM scheme.

A. TCP Source Evolution Equations

Consider the th class of elephants; the temporal evolution (4)
of the average window of TCP sources in the class, , is where is the total propagation delay1 experienced by ele-
described by the following differential equation: phants in class , and is the time when the fluid injected at
time by the TCP sources reaches the th buffer along its path
(1) . We have

(5)
where is the average RTT for class , and is the loss
indicator rate experienced by TCP flows of class . The loss indicator rate is instead given by
The differential equation is obtained by considering the fact
that elephants can be assumed to always be in congestion avoid- (6)
ance (CA) mode, so that the window dynamics are close to ad-
ditive increase, multiplicative decrease (AIMD). The window
where is the instantaneous emission rate of TCP
increase rate in CA mode is approximatively linear, and cor-
sources, is a calibration parameter, and is the instanta-
responds to one packet per RTT. The window decrease rate is
neous loss probability experienced by elephants in class :
proportional to the rate with which congestion indications are
received by the source, and each congestion indication implies
a reduction of the window by a factor two. (7)

B. Network Evolution Equations

Finally
denotes the (fluid) level of the queue in the th buffer
at time ; the temporal evolution of the queue level is described (8)
by

(2) where , if buffer is the first buffer traversed by elephants

of class , and 0, otherwise; is derived by the routing matrix,
where represents the fluid arrival rate at the buffer, being if buffer immediately follows buffer along
the departure rate from the buffer (which equals , provided ; is the number of class active flows.
that ) and the function represents the instanta- It can be observed that the MGT fluid model is extremely
neous loss probability at the buffer, which depends on the RED simple, requiring just one equation per class of elephants, thus
parameters. An explicit expression for is given in [3] for being capable of scaling to quite large network models. How-
RED buffers. ever, we must also note that the description of TCP mice with
If denotes the instantaneous delay of buffer at time , the MGT model is not natural, because (obviously) the start time
we can write of each mouse determines its window dynamics over time. This
aspect is not captured by (1), and one equation has to be written
(3) for each mouse, as in [2]. This means that the independence of
the fluid model complexity with respect to the number of flows
If indicates the set of flows traversing buffer , and is lost. Moreover, the MGT model, due to the fact that it only de-
are respectively the arrival and departure rates at buffer scribes the average dynamics, also has problems in coping with
referred to elephants in class , it results: drop-tail buffers. Finally, the calibration parameter in (6), which
is necessary to compensate for the use of the average window
size, instead of the window size distribution, must be set empir-
ically.

III. PREVIOUS WORK ON FLUID MODELS

To the best of our knowledge, fluid models were first proposed
which means that the total amount of fluid arrived up to time
in [3] to study the interaction between TCP elephants and a
at the buffer leaves the buffer by time , since the buffer
RED buffer in a packet network consisting of just one bottleneck
is FIFO.
link. In [5], the authors have recently extended their model to
C. Source–Network Interactions 1Equation (4) comprises the propagation delay g in a single term, as if it
were concentrated only at the last hop. This is just for the sake of easier reading,
Consider elephants in class . Let be the th buffer since the inclusion of the propagation delay of each hop would introduce just a
traversed by them along their path of length . The RTT formal modification in the recursive equation of t .
AJMONE MARSAN et al.: USING PARTIAL DIFFERENTIAL EQUATIONS TO MODEL TCP MICE AND ELEPHANTS IN LARGE IP NETWORKS 1291

consider general multi-bottleneck topologies comprising RED distribution for the TCP flow population. This major improve-
routers. ment in the representation of the TCP sources dynamics gives
The equations reported in Section II briefly summarize the us the advantage of a greater model flexibility, which: 1) allows
fluid model proposed in [5], which constitutes the starting point TCP mice to be described in a way such that the insensitivity of
for our work. This set of ordinary differential equations must be complexity with respect to the number of TCP flows is main-
solved numerically, using standard discretization techniques. tained and 2) permits the modeling of networks in which AQM
In [1] and [2], an alternative fluid model has been proposed routers coexist with drop-tail routers.
to describe the dynamics of the average window for TCP ele- In other words, rather than just describing the average TCP
phants traversing a network of drop-tail routers. The behavior connection behavior, we statistically model the dynamics of the
of such a network is pulsing: congestion epochs in which some entire population of TCP flows sharing the same path. This ap-
buffers are overloaded (and overflow) are interleaved to periods proach leads to systems of partial derivatives differential equa-
of time in which no buffer is overloaded, and no loss is experi- tions, and produces more flexible models, which scale indepen-
enced, due to the fact that previous losses forced TCP sources dently of the number of TCP flows.
to reduce their sending rate. In such a setup, a careful analysis In this section, we first introduce the basic model for the TCP
of the average TCP window dynamics at congestion epochs is flow population. This basic model can be extended by adding
necessary, whereas sources can be simply assumed to increase several features, which permit a progressively more accurate de-
their rate at constant speed between congestion epochs. This be- scription of the behavior of TCP sources. Such extensions are
havior allows the development of fluid equations and an efficient described one by one for the sake of readability, but they can
methodology to solve them. Ingenious queueing theory argu- be combined at will, to obtain models with the desired level of
ments are exploited to evaluate the loss probability during con- accuracy and numerical complexity.
gestion epochs, and to study the synchronization effect among
sources sharing the same bottleneck link. Also in this case, the A. Basic TCP Sources
complexity of the fluid model analysis is independent of link To begin, consider a fixed number of TCP elephants. We use
capacities and of the number of TCP flows. An extension that to indicate the number2 of elephants of class whose
allows considering TCP mice has also been proposed in [1] and window is at time . For the sake of simplicity, we consider
[2]. In this case, since the dynamics of TCP mice with different just one class of flows, and omit the index from all variables.
size and/or different start times are different, each mouse must The source dynamics are approximately described by the fol-
be described with two differential equations; one representing lowing equation, for :
the average window evolution, and one describing the workload
evolution. As a consequence, one of the nicest properties of fluid
models, the insensitivity of the complexity with respect to the (9)
number of TCP flows, is lost.
In [6]–[9], fluid models have been exploited to prove proper- where is the loss indication rate. A formal derivation of
ties related to the asymptotic behavior of a single RED bottle- (9) is given in Appendix A. Note that this equation is equivalent
neck topology fed by long lived TCP connections in the so called to the deterministic transport equation reported in Corollary 1
“many flows regime”, i.e., when the number of TCP flows, the of [9], which was obtained by applying Mean Field Analysis.
bottleneck capacity, and the buffer-size increase jointly to in- The intuitive explanation of the formula is the following. The
finity. In particular, in [9] the windows size dynamic for a pop- time evolution of the population described by is gov-
ulation of long-lived TCP connections is represented by a erned by two terms: 1) the integral accounts for the growth rate
stochastic process whose time samples are the window size dis- of due to sources with window between and that
tributions (the process is said to be measure-valued), and the experience losses, and 2) the second term is the decrease rate
mean field dynamics of the process are described by a deter- of due to sources increasing their window with rate
ministic “transport equation” which can be approximated by a .
partial differential equation, under mild assumptions. The quantity can be computed by recalling (6):
In this paper, differently from [9], we adopt an approach
which allows us to directly obtain an approximate description
(10)
of the TCP source window size distribution dynamics, based on
partial differential equations. We also show that our description
of the source dynamics allows a natural representation of mice in which the current window size of the sources that emitted the
as well as elephants, with no sacrifice in the scalability of the lost fluid approximates the window size value at which those
model. sources emitted this fluid. Intuitively, this loss model distributes
the lost fluid over the entire population, proportionally to the
window size. Note that this loss model does not require any cal-
IV. MODELING TCP ELEPHANTS ibration parameter, contrary to the MGT model; indeed, statis-
The class of fluid models that we propose in this paper differs tics like the variance of the TCP flow windows size impacts on
from the previous proposals (with the exception of [9]) because, the network stationary behavior.
instead of describing just the evolution of the average window 2P (w; t) is assumed to be a continuous function IR ! IR due to the fluid
size of TCP sources, we model the evolution of the window size nature of the model.
1292 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 13, NO. 6, DECEMBER 2005

Fig. 1. Fluid model (left) and ns-2 (right): average window size distribution for 8 TCP elephants traversing a single bottleneck link with RED buffer, varying
their maximum window size; these TCP flows compete with 8 other TCP elephants with maximum window size 64.

B. Accounting for the Maximum Window Size

W
TABLE I
MAXIMUM WINDOW SIZE AND AVERAGE WINDOW SIZE (AWS) (IN
We now extend the basic model of (9) to account for the max- PACKETS) FOR CLASS 2 FLOWS, AVERAGE QUEUE LENGTH (AQL) (IN
imum window size of TCP sources, that we denote by . PACKETS) AND AVERAGE LOSS PROBABILITY (ALP) FOR THE
EXPERIMENTS OF SECTION IV-C
It holds that

(11)

for , where is the unit step function, and Consider the case of a single bottleneck link topology
is the number of TCP flows whose window is exactly in which a gentle version of the RED AQM algorithm
equal to . ( , , , )
For we can write is implemented, with two classes of eight TCP elephants
saturating the link capacity ( Mb/s), assuming a
propagation delay equal to 30 ms. We compare the results of
(12) three different experiments, in which the first elephant class
with the boundary conditions and (class 1) has always maximum window size 64, while the other
. The derivation of (11) class (class 2) has maximum window size 64, 32, and 24. The
is very similar to that of (9). The first term in (11) is the packet size for this and all other experiments in this paper is
contribution of all TCP sources which experience losses at 10 000 bits. In Fig. 1, we show the window size probability
window size between and ( if exceeds it). The density function of elephants in class 2 predicted by our model
second term of (11) is the contribution of all TCP sources at and by ns-2. In Table I, we compare the average window size,
maximum window size that experience losses; note that this the average queue length and the loss probability for the model
contribution exists only for windows greater than . and the ns-2 simulator. Note that for lower , the average
The growth rate of is obtained as the limit of the window size of class 2 elephants is smaller; at the same time,
usual growth rate of . The de- the average window size for class 1 flows increases, so that the
crease rate of is simply . average window size of all the 16 TCP elephants is roughly
constant and equal to 20. A model without window size clip-
C. Experiments With RED Buffers ping, like for example the one in [3]–[5], is capable of correctly
In this subsection, we discuss some numerical results refer- estimating the average window size of the 16 elephants, but
ring to the mathematical model in (11). Before proceeding, we fails in capturing the differences among classes with different
notice that all the results shown in this paper were obtained by maximum window size values. The comparison clearly shows
numerically solving the model. For this purpose, we applied that the fluid model is quite accurate.
standard discretization techniques; in particular, a first-order fi-
nite differences method for the sources equations, and a fourth- D. Considering Fast Recovery
order Runge–Kutta method for the queue equations, as better Newer versions of TCP (such as NewReno—see RFCs 2581
explained in Appendix E. and 3782) avoid halving the window more than once for RTT,
AJMONE MARSAN et al.: USING PARTIAL DIFFERENTIAL EQUATIONS TO MODEL TCP MICE AND ELEPHANTS IN LARGE IP NETWORKS 1293

even in the case of multiple losses. To model this fact, we divide

the population of TCP flows whose congestion window is
at time in two classes: class comprises all sources that expe-
rienced losses during the last RTT, while class is composed
by remaining sources,3 so that .
We can write

(13)

(14) Fig. 2. Fluid model and ns-2 simulator: queue size evolution for one bottleneck
link fed by a drop-tail buffer and traversed by TCP elephants.
A formal derivation of (13) and (14) is reported in Appendix B.
An intuitive explanation of the two equations can be provided the same buffer. In addition, during congestion epochs, losses
as follows. In the right-hand side of (13), the first two terms are not evenly distributed among TCP flows, but are more likely
account for the decrease rate of the number of elephants of class to affect TCP sources with larger window size. In this context,
whose window is at time , due to: (i) sources in class it is necessary to distinguish among sources with different in-
experiencing losses and moving to class , (ii) sources in stantaneous window size, while at the same time accounting for
class increasing their window. The third term refers to the the effects of the TCP fast recovery mechanism, which prevents
sources moving to class from class after experiencing a TCP sources from halving their window several times within
RTT without losses. In the right-hand side of (14), the first term one round trip time.
accounts for the growth rate of the number of elephants of class The level of detail in the description of the TCP sources dy-
whose window is at time , due to sources in class namics adopted in this paper allows an easy description of the
experiencing losses. The second and third terms account for the time-dependent behavior of the packet loss probability:
decrease rate due to: 1) sources moving to class from class
after a RTT without losses and 2) sources in class increasing (15)
their window.
More general fluid equations describing TCP elephants and that is, the loss probability equals (the
accounting for the TCP threshold mechanisms and for time-outs relative difference between the instantaneous arrival rate and the
are reported in [13]. service rate) only when the buffer is full, being the capacity
of buffer , and the indicator function.
E. Modeling Drop-Tail Buffers A different approach is used in [1] and [2] to describe the
As we have already mentioned, a fluid model for the descrip- dynamics of the average window size for TCP flows traversing
tion of RED AQM schemes was originally proposed in [3]. RED a network with drop-tail buffers. In those papers, the loss indi-
matches quite well the fluid modeling approach, since in RED cator rate is obtained by applying queueing theory results which
buffers the loss probability is a smooth function of the queue are not “internal” to the fluid model. That approach is probably
length averaged over a rather long time window. The case of difficult to generalize to networks including both drop-tail and
drop-tail buffers is instead much more difficult to describe with AQM buffers.
fluid models, since in this case the loss probability is a discon- F. Experiments With Drop-Tail Buffers
tinuous function of the instantaneous queue size.
Many studies have shown that the behavior of networks car- In this subsection, we briefly comment some numerical re-
rying TCP traffic is pulsing: congestion epochs in which some sults obtained with our modeling approach in the case of drop-
buffers are overloaded (and overflow) are interleaved to periods tail buffers.
of time in which traffic is lighter, buffers are not saturated, and First, we consider the case of a single bottleneck link (with
no loss is experienced. Light traffic periods are the result of data rate Mb/s, propagation delay 30 ms), traversed by
losses at the previous congestion epochs, that force TCP sources just one class of 30 TCP elephants, with maximum window size
to reduce their emission rate. As a consequence, the loss pro- 64 packets; the maximum buffer size is set to 1000. The curves
cesses experienced by TCP flows traversing drop-tail buffers are in Fig. 2 show the queue size evolution over time. Our model
quite bursty. This burstiness induces a high degree of correlation captures the well-known oscillating behavior of TCP, which was
(synchronization) among the dynamics of TCP sources sharing observed in simulation experiments as well as measurements
[11], [12].
3For the sake of simplicity, the equations in this section and in the rest of
The results of ns-2 simulations are reported in Fig. 2 and
the paper do not consider the effect of the maximum window size. However, in
all numerical results that are presented in this paper the effect of the maximum Table II for comparison, and again show that the fluid model
window size is always accounted for. is accurate.
1294 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 13, NO. 6, DECEMBER 2005

TABLE II
AVERAGE WINDOW SIZE (AWS), AVERAGE QUEUE LENGTH (AQL) AND
AVERAGE LOSS PROBABILITY (ALP) FOR THE SETUP OF SECTION IV-F

Fig. 4. Overlap of TCP0 curves from the fluid model (Fig. 3) and ns-2.

phase that is executed when the TCP connection is opened. For

this reason, in order to model TCP mice, we model the initial
slow-start phase up to the first loss or to the first hit of the max-
imum window size, and then we assume that flows stay in con-
gestion avoidance for the rest of the connection lifetime.
Fig. 3. Fluid model: window size evolution for three long-lived TCP flows Let be the number of flows in slow-start with
with interfering UDP traffic.
window size and residual workload at time . Analo-
gously, refers to flows in congestion avoidance. We
The second scenario we consider is a network topology com- can write
prising two links, the first fed by a RED buffer, the second fed by
a drop-tail buffer.4 The links are crossed by five classes of ele-
phants. Two classes of TCP flows are single-hop (TCP0 crosses
the first link, TCP1 crosses the second one), while the other one
(TCP2) crosses both links; the two links are also crossed by two
interfering classes of CBR UDP flows (UDP3 crosses the RED
buffer, UDP4 crosses the drop-tail buffer). UDP3 is on in the (16)
time interval [10, 30] s, UDP4 in the time interval [20, 40] s:
when the UDP flows are on, they consume about 40% of the
bandwidth of their link. Fig. 3 shows plots of the window size
evolution for the three TCP flow classes. When UDP3 starts, the
window size of the two TCP flow classes sharing the same link
decreases; when also UDP4 starts, the window size of TCP1 de- (17)
creases, and again that of TCP2 goes down, in favor of TCP0.
The window size of TCP0 and TCP2 increases when UDP3 A formal proof of these equations is given in Appendix C. An
ends, while those of TCP1 and TCP2 increase when UDP4 ends. intuitive explanation is as follows. In (16), the first two terms
The results of ns-2 simulations for the same setup once more on the right-hand side account for the decrease rate of
show that the fluid model is quite accurate: for instance, in Fig. 4 due to: 1) sources increasing their rate (first term) and 2) sources
we overlap the curves of the model and the ns-2 simulator for terminating because of null residual workload (second term).
the TCP0 elephants. Complete ns-2 results can be found in [13]. The last three terms account for the growth rate of .
These results prove that our model can cope with both con- The third term takes into account those sources with previous
trolled (TCP) and uncontrolled (UDP) long-lived flows, and is residual workload slightly greater than , assuming at time a
capable of predicting the TCP transient effects due to the pres- ; the weight of this term is because the work-
ence of on-off interfering sources. load is reduced with an average rate equal to units of fluid
per RTT. The fourth term represents those sources in congestion
V. MODELING TCP MICE avoidance with window between and and residual work-
We now come to the very important issue of modeling TCP load that experience a loss. They are added to
mice, whose dynamics are mostly, if not completely, due to because their window is halved (and becomes ) and their
the slow-start algorithm, and in particular to the first slow-start residual workload goes back to , as the lost unit of fluid must
be retransmitted. Finally, the fifth term represents an increase
4It is worth observing that all previous applications of fluid models to packet
similar to the fourth term, applied to sources in slow-start: these
networks always considered either RED buffers, or drop-tail buffers, but the two
types of buffers were never mixed, since the fluid models could not support this sources, with window size between 1 and and residual work-
feature. load , experience a loss and consequently move to a state
AJMONE MARSAN et al.: USING PARTIAL DIFFERENTIAL EQUATIONS TO MODEL TCP MICE AND ELEPHANTS IN LARGE IP NETWORKS 1295

in which they are in congestion avoidance, their window is description of the network dynamics by means of stochastic
and their residual workload goes back to . models, such as continuous-time or discrete-time Markov
Equation (17) is very similar to (16), since the evolution of chains and queueing models.
with respect to the residual workload (second and Deterministic fluid models have been proven to represent cor-
third terms) is the same, and the first term differs only for the rectly the asymptotic behavior of TCP when the number of ac-
fact that the window growth is in this case exponential rather tive flows (elephants) tends to infinity [8].
than linear. Moreover, the fourth term refers to sources moving Indeed, when considering scenarios with only elephants, ran-
into congestion avoidance because of a loss [similarly to the fifth domness, which is completely lost in fluid models, plays a minor
term of (16)]. Finally, being the mice arrival rate, the last role, because queues tend to be heavily congested, and the loss
term accounts for newly activated TCP mice. rate is basically determined by the load offered by TCP connec-
Note that the representation of the TCP window dynamics tions in excess of the bottleneck capacity.
over the space allows us to distinguish among TCP mice Instead, fluid models are not suitable to analyze network sce-
with different instantaneous window sizes, thus providing the narios in which the capacity of the links is not saturated. In par-
correct level of detail for the analysis of this type of TCP flows. ticular, they completely fail to predict the behavior of a network
Indeed, TCP mice open in slow-start, with window 1, and then loaded only by TCP mice: if links are underloaded (i.e., their av-
their window evolves according to (16) and (17). erage utilization is smaller than one), fluid models predict that
The model of TCP mice can be simplified by assuming flow buffers are constantly empty. This is not what we observe and
lengths to be exponentially distributed, with average . Thanks measure in packet networks; the discrepancy is essentially due
to the memoryless property of the exponential distribution, we to the fact that, in underload conditions, the stochastic nature
can write of the input traffic plays a fundamental role that cannot be ne-
glected.
Randomness, indeed, impacts the system behavior at many
levels:
1) at flow level, since the arrival process of TCP flows ex-
hibits a nonnegligible burstiness, which causes the short-
term offered load at the queues to randomly vary over
time, thus leading to sporadic periods of congestion;
(18) 2) at packet level, since the arrival process of packets at
queues exhibits a bursty behavior, thus causing sporadic
buffer overloads also during periods in which the average
utilization factor is smaller than 1.
This implies that, when analyzing the behavior of under-
loaded networks, the complete determinism of fluid models is
not satisfactory: we must stop short of reducing the network
(19) operations to the deterministic evolution of average parameters,
keeping in the model some of the stochastic characteristics of
where is the average loss probability experienced by the the network behavior. This can be done by using stochastic
flow, during its total active period. The formal derivation of the differential equations, rather than deterministic differential
second term is reported in Appendix D. We can approximate equations, to describe the system dynamics, and then solving
by using the same approach proposed in [3] and [4] to them (for example) with a Monte Carlo technique.
evaluate the average loss probability in a RED queue; we obtain In practice, with respect to the model we just presented, we
can:
(20) 1) use a Poisson counter with average to describe the
mice arrival rate, rather than a deterministic rate, so that
instead of (19) we have a stochastic partial differential
with the instantaneous loss probability, defined in (6), and equation.
the average window size at time . 2) use a nonhomogeneous Poisson process to describe the
We wish to stress the fact that (18)–(20) provide quite a pow- completion of TCP connections; the average at time of
erful tool for an efficient representation of TCP mice, since a such process is represented by the second term of (18); by
wide range of distributions (including those incorporating long- so doing, also (18) becomes a stochastic partial differen-
tail distributions) can be approximated with an arbitrary degree tial equation. Note that this suggests that, due to retrans-
of accuracy by a mixture of exponential distributions [14]. missions, the completion time of connections increases as
their loss probability grows.
A. Randomness in Fluid Models 3) use a Poisson point process (possibly with batch arrivals)
The fluid models that we have presented so far provide a de- to describe the workload emitted by TCP sources, rather
terministic description of the network behavior, thus departing than a continuous deterministic fluid process, but keeping
from the common approach of attempting a probabilistic the average rate .
1296 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 13, NO. 6, DECEMBER 2005

TABLE III
ARRIVAL RATES (AR), AVERAGE COMPLETION TIMES (ACT), AVERAGE
WINDOW SIZE (AWS), AVERAGE QUEUE LENGTH (AQL) AND AVERAGE LOSS
PROBABILITIES (ALP) FOR THE EXPERIMENTS OF SECTION V-B.

TABLE IV
PARAMETERS OF THE HYPER-EXPONENTIAL DISTRIBUTION
Fig. 5. Average window size evolution for elephants competing with mice on a APPROXIMATING THE PARETO DISTRIBUTION
single bottleneck link; the average window size decreases when the mice arrival
rate grows from 100 to 400 connections/s.

Of course, this is only one possible approach to account for

randomness when studying the behavior of TCP mice; we do not
claim any optimality of this approach, and a deeper investigation
is needed about the possible ways of coping with randomness,
without losing the property of independence of the model com-
plexity with respect to the number of flows.
quence, we now consider fluid models employing the stochastic
extensions described in Section V-A.
B. Experiments With Mice Consider a single bottleneck link fed by a drop-tail buffer,
In this subsection, we discuss results for network scenarios with capacity equal to 1000 packets. The link data rate is
comprising TCP mice. First, we consider a case in which both 1.0 Gb/s, while the propagation delay between TCP sources and
mice and elephants coexist. Second, we investigate the impact buffer is 30 ms. In order to reproduce a TCP traffic load close to
of the source emission model in a scenario where only mice are what has been observed on the Internet, flow sizes are distributed
active. Third, we study the impact of the flow size distribution. according to a Pareto distribution with shape parameter equal to
Finally, we investigate the invariance properties of the network 1.2 and scale parameter equal to 4.
when mice are present. Using the algorithm proposed in [14], we approximated the
1) Results With Both Elephants and Mice: These first re- Pareto distribution with a hyper-exponential distribution of the
sults refer to a single bottleneck link fed by a drop-tail buffer. ninth order, whose parameters are reported in Table IV. The re-
The buffer size is equal to 1000 packets, the link capacity is sulting average flow length is 20.32 packets. Correspondingly,
Mb/s, the propagation delay between TCP sources nine classes of TCP mice are considered in our model. The max-
and buffer is 30 ms. Twenty TCP elephants are active, with imum window size is set to 64 packets for all TCP sources. Ex-
maximum window size 64 packets, and coexist with TCP mice, periments with loads equal to 0.6, 0.8 and 0.9 were run; how-
whose length is geometrically distributed with mean 20 seg- ever, for the sake of brevity, we report here only the results for
ments. The TCP mice arrival rate is set equal to 100, 200, and load equal to 0.9.
400 connections per second. The presence of elephants is cru- Fig. 6 compares the queue lengths distributions obtained with
cial in order to saturate the link bandwidth, because they con- ns-2, and with the stochastic fluid model.
sume the capacity that is not used by mice. Indeed, in Fig. 5 While in the model the flow arrival and completion processes
and Table III we can see that the average window size for ele- have been randomized according to a nonhomogeneous Poisson
phants decreases when the arrival rate of mice increases. In the process (see Section V-A ), different approaches have been con-
same table, we also report the average completion time (ACT) sidered to model the traffic emitted by sources in a small interval
of mice, obtained from the average number of active mice, by :
applying Little’s theorem. Poisson: the emitted traffic is a Poisson process with time-
Table III and Fig. 5 also report the results of ns-2 simulations varying rate;
for the same setup, for comparison: the fluid model can be ob- Det-B: the emitted traffic is a batch Poisson process with
served to be quite accurate in this case too. time-varying rate and constant batch size, equal to the in-
2) Results With Mice Only: Impact of the Emission stantaneous average TCP mice window size;
Model: If elephants are removed from the network, deter- Exp-B: the emitted traffic is a batch Poisson process with
ministic fluid models do not provide useful information about time-varying rate and exponential batch size, whose mean is
the network behavior, as explained in Section V-A. As a conse- equal to the instantaneous average TCP mice window size;
AJMONE MARSAN et al.: USING PARTIAL DIFFERENTIAL EQUATIONS TO MODEL TCP MICE AND ELEPHANTS IN LARGE IP NETWORKS 1297

TABLE V
AVERAGE LOSS PROBABILITY (ALP), AVERAGE QUEUE LENGTH (AQL) AND
AVERAGE COMPLETION TIMES (ACT) IN SECONDS OF THE NINE CLASSES OF
MICE FOR THE SETUP OF SECTION V-B

TABLE VI
PARAMETERS OF THE THREE FLOW LENGTH DISTRIBUTIONS

Fig. 6. Queue length distribution for single drop-tail bottleneck, varying the
random process modeling the workload emitted by the TCP sources; comparison
with ns-2 simulator.

Win-B: the emitted traffic is a batch Poisson process with

time-varying rate, in which the batch size distribution is an optimistic prediction of completion times. The Win-B model
equal to the instantaneous TCP mice window size distribu- moderately overestimates the average queue length and loss
tion. probability, as pointed out in [15]. However, for very short
The Poisson approach corresponds to the most traditional and flows, completion time predictions obtained with the Win-B
simple choice; however, as shown in [15], the traffic emitted model are slightly optimistic; this is mainly due to the fact that
by a population of TCP sources, along with the well-known an idealized TCP behavior (in particular, without timeouts) is
effects of long-range correlation, which are essentially due to considered in the model.
the slow fluctuations in the number of active flows (as shown 3) Results With Mice Only: Impact of the Flow Size: We
later), also exhibits some short-term burstiness, which is inti- now discuss the ability of our model to capture the impact on
mately related to the TCP window mechanism, and must be the network behavior of the flow size variance.
considered by an accurate model. For this reason, following the We consider three different scenarios, in which flow lengths
approach proposed in [15], we modeled the traffic emitted by are distributed according to either an exponential distribution
TCP sources as a Poisson process with batch arrivals, where (“Distr.1”), or hyper-exponentials of the second order (“Distr.2”
the batch size is adapted to the current windows size of TCP and “Distr.3”). For all three scenarios, we keep the average flow
sources. According to the Det-B and Exp-B approaches, the size equal to 20.32 (this is the average flow size used in the pre-
batch size distribution is adapted to the current TCP window vious subsection), and we vary the standard deviation . De-
size distribution by matching just the first moment; instead, with tailed parameters of our experiments are reported in Table VI.
the Win-B approach, a complete match between the batch size Table VII shows a comparison between the results obtained
and the window size distributions is possible. As a consequence, with either the Win-B model or ns-2. As in previous experi-
we expect that the Win-B approach outperforms both Det-B and ments, the model moderately overestimates both the average
Exp-B. Indeed, Fig. 6 confirms our expectations. If we use a loss probability and the average queue length. The discrepan-
Poisson process to model the instants in which packets (or, more cies in the average completion times between model and ns-2
precisely, units of fluid) are emitted by TCP sources, the re- remain within 10%.
sults generated by the fluid model cannot match the results ob- Fig. 7, which reports the queue length distributions obtained
tained with the ns-2 simulator. Instead, the performance predic- by the model in the three scenarios, emphasizes the significant
tions obtained with the fluid model become quite accurate when dependency of the queue behavior on the flow size variance.
the workload emitted by TCP sources is taken to be a Poisson This dependency is mainly due to the complex interactions be-
process with batch arrivals. The best fitting (confirmed also by tween the packet-level and flow-level dynamics which are due
several other experiments, not reported here for lack of space) to the TCP protocol.
is obtained for batch size distribution equal to the instantaneous 4) Results With Mice Only: Impact of the Link Ca-
TCP mice window size distribution (case Win-B). Note that our pacity: Finally, we discuss the effect on performance of
proposed class of fluid models naturally provides the informa- the link capacity. The objective of this last study of networks
tion about the window size distribution. loaded with TCP mice only is to verify whether the performance
Table V reports the average loss probability, the average of networks which differ for a multiplicative factor in capacities
queue length, and the average completion time for each class of show some type of invariance, like in the case of elephants.
TCP mice, obtained either with ns-2, or with the Poisson and More precisely, we wish to determine whether the queue
Win-B models. The Poisson model significantly underestimates length distribution exhibits any insensitivity with respect to
the average queue length and loss probability, thus producing the bottleneck link capacity, for the same value of the traffic
1298 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 13, NO. 6, DECEMBER 2005

TABLE VII
AVERAGE LOSS PROBABILITY (ALP), AVERAGE QUEUE LENGTH (AQL) AND
AVERAGE COMPLETION TIMES (ACT) IN SECONDS OF THE DIFFERENT
CLASSES OF MICE FOR THE SETUP OF SECTION V-B, HAVING INTRODUCED
RANDOM ELEMENTS

Fig. 8. Queue size distribution for single drop-tail bottleneck, varying the
bottleneck capacity.

the flow-level dynamics on the network performance tend to be-

come negligible, and the packet-level behavior resembles that of
a single server queue loaded by a stationary Poisson (or batched
Poisson) process, for which the queue length distribution is in-
dependent of the server capacity.
To confirm this intuition, we solved the fluid model by elim-
inating the randomness at the flow level (i.e., in the flow arrival
and departure processes), and we observed that the dependency
on the capacity disappears.
Fig. 7. Queue size distribution for single drop-tail bottleneck, varying the flow We would like to remark, however, that the flow length dis-
length distribution. tribution plays a major role in determining the system capacity
above which the queue length distribution no longer depends
intensity. This curiosity is motivated by the fact that in many on the system capacity—the invariance phenomenon appears at
classical queueing models (e.g., the M/M/1 queue, possibly higher data rates when the variance of the flow length distribu-
with batch arrivals) the queue length distribution depends only tion increases.
on the average load, not on the server speed.
We consider the third scenario (“Distr.3”) of the previous ex- VI. CONCLUSION
periment, we fix the traffic load at 0.9, and we study four dif-
In this paper, we have proposed a new fluid model approach
ferent networks, in which the bottleneck capacity is equal to
for the investigation of the performance of IP networks loaded
10 Mb/s, 100 Mb/s, 1 Gb/s, and 10 Gb/s, respectively.
by TCP mice and elephants (as well as UDP flows). Our ap-
The results of the fluid model, depicted in Fig. 8, show that,
proach exploits partial differential equations, thus permitting
in general, the queue length distribution exhibits a dependency
the description of distributions, instead of averages, hence
on the link capacity. The packet-level behavior, indeed, strongly
achieving better accuracy in the results with respect to previ-
depends on flow-level dynamics, which cause a slowly varying
ously proposed fluid modeling approaches.
modulation of the arrival rate at the packet level. The flow-level
The performance estimates obtained with our fluid models
dynamics, however, do not scale up with the capacity of the
have been compared against ns-2 simulations in the cases in
system, since the random variable which represent the number
which the latter are feasible, proving both the accuracy and the
of active flows has a coefficient of variation which decreases
scalability of the proposed modeling approach.
as we increase the system capacity (consider, for example, the
In case of underloaded networks populated only by TCP
Poisson distribution of the number of active flows proposed in
mice, we have pointed out a fundamental limitation in the
[16]).
deterministic approach to describe the network dynamics,
Nevertheless, when the capacity of the system becomes very
and we have suggested and discussed different solutions to
large (in the considered example, greater than 1 Gb/s) the de-
introduce randomness in fluid models in order to obtain reliable
pendence of the queue distribution on capacity tends to vanish,
predictions of the system behavior.
and the queueing behavior becomes indeed independent from
the link capacity. This phenomenon was confirmed by ns-2 sim-
ulations. APPENDIX A
This behavior is mainly due to the fact that when the capacity PROOF OF (9)—BASIC SOURCES
becomes very large, the coefficient of variation of the number We wish to estimate the evolution of ; we define
of active flows becomes small. As a consequence, the effects of as the probability density of the
AJMONE MARSAN et al.: USING PARTIAL DIFFERENTIAL EQUATIONS TO MODEL TCP MICE AND ELEPHANTS IN LARGE IP NETWORKS 1299

window distribution at time . Consider a small enough such Now consider the population of sources which will leave class
that . Let be the number of sources with because an RTT is elapsed. We assume an exponential distri-
window at time , but with window at time . bution of the departure time of each source from class , with
All the sources which do not experience any loss indication average . Hence, the number of sources moving from class
during the interval increase their window with rate to class will be , by observing
. Among these sources, includes only the ones that the fraction of sources that already left class by the end of
with initial window , since they will exceed is . Now observe that the
by time . If we assume that the loss indication process can , defined as the number of sources moving from class to
be approximated locally (i.e, in the small interval ) class and exceeding window , will include sources counted
with a Poisson process with rate , the probability that no in both and . These source can be derived by ,
losses are experienced during is ( ); since . Now we are able to add all the
then contributions:

(24)
(21) By recalling (23), we can compute

Now let be the number of sources with at

time , but with window at time . include only
the sources: 1) with window in the range at
time , and 2) receiving a loss indication in the interval
. Note that the probability of receiving multiple loss indica-
tions is , hence, negligible. Hence, whose limit is

(22) In other words, is negligible with respect to .

Hence, from (24) we find (14):

Since , we can find

(9):

We can now estimate

APPENDIX B
where are the sources in class exceeding window by
PROOF OF (13) AND (14)—SOURCES WITH FAST
the time interval and the sources moving from class
RECOVERY MECHANISMS
to class , due to losses. It holds that
The proof is similar to the previous one. Let
and . Consider the
sources of class moving to class during the interval
; among these, will have a window and will
contribute to increase . Analogously to (22):

Analogously to , . It
can be shown that is negligible with respect to .
Hence, we can obtain (13):
The number of sources of class exceeding by time
is, analogously to (21):

(23)
1300 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 13, NO. 6, DECEMBER 2005

APPENDIX C ODEs. To solve the source dynamics PDEs of our model,

PROOFS OF (16) AND (17)—SOURCES WITH FINITE FLOWS we used a finite differences methodology. Consider a general
The only terms which need a formal proof are the ones which integro-differential equation of the form
model the workload evolution. is the number of sources
which enter during a time interval of size because (25)
their workload has just decreased. is given by all the sources
with window between 1 and , and residual workload between where and are continuous function on
and , being the instantaneous emission (for example in (9), and
rate of sources with window . Formally ).
First we sample , and onto a bidi-
mensional discrete lattice, defining ,
and ; then we approximate
the partial derivative:

Similarly, we approximate
Finally

At last, we approximate
To account for the sources which stop their activity during the
time interval of size , it is enough to set .

APPENDIX D
PROOFS OF (18) AND (19)—SOURCES WITH FINITE FLOWS with being the numerical approximation of
EXPONENTIALLY DISTRIBUTED .
In conclusion, we obtain for the PDE the numerical recursion
Regarding (18) and (19), we prove formally only those terms
accounting the variation of the population due to the
variation of the sources residual workload. Consider a time in-
terval of size and a source which does not experience any loss
with window . The probability that this source stops within The whole set of differential equations, which defines the
the interval, i.e., its residual life time is less than , is equal fluid model for the considered network, is solved according to
to , thanks to the the following procedure. At each time iteration , the param-
memoryless property. Then, the contribution of the sources stop- eters and are evaluated. Note that
ping is at time step we assume that all queues are empty, thus
there are no losses and the RTTs account only for fixed propaga-
tion delays. Then the equations of the sources’ dynamics (PDE)
are solved, obtaining as output the amount of fluid that sources
The final contribution is given by multiplying the previous for- inject into the network. Next, the network queues’ dynamics
mula for ( ), corresponding to the number of sources (ODE) are solved using as input the amount of fluid injected
not experiencing any losses (those experiencing losses have al- by sources. The state variables of the queues are then used to
ready been considered in other terms). update the RTTs and loss rates perceived by the sources. We
notice that the proposed scheme is very similar to the one used
APPENDIX E in [5].
NUMERICAL SOLUTIONS OF FLUID MODELS
REFERENCES
The solutions of our fluid models are obtained by solving nu-
[1] F. Baccelli and D. Hong, “Interaction of TCP flows as billiards,” in Proc.
merically the set of differential equations which define the fluid IEEE INFOCOM, San Francisco, CA, 2003, pp. 895–905.
model. Two different approaches have been employed to solve [2] , “Flow level simulation of large IP networks,” in Proc. IEEE IN-
the set of ordinary differential equations (ODEs) describing net- FOCOM, San Francisco, CA, 2003, pp. 1911–1921.
[3] S. Misra, W. B. Gong, and D. Towsley, “Fluid-based analysis of a
work queues dynamics and the set of partial differential equa- network of AQM routers supporting TCP flows with an application to
tions (PDEs) representing the source window dynamics. RED,” in Proc. ACM SIGCOMM, Stockholm, Sweden, Aug. 2000, pp.
Similarly to [5], we solve numerically the ODEs of our 151–160.
[4] C. V. Hollot, V. Misra, D. Towsley, and W. B. Gong, “On designing
model using a fourth-order Runge–Kutta methodology. The improved controllers for AQM routers supporting TCP flows,” in Proc.
Runge–Kutta algorithm is a widely used method to solve IEEE INFOCOM, Anchorage, AK, 2001, pp. 1726–1734.
AJMONE MARSAN et al.: USING PARTIAL DIFFERENTIAL EQUATIONS TO MODEL TCP MICE AND ELEPHANTS IN LARGE IP NETWORKS 1301

[5] Y. Liu, F. Lo Presti, V. Misra, and D. Towsley, “Fluid models and so- Michele Garetto (S’01–M’04) received the Dr.Ing.
lutions for large-scale IP networks,” in Proc. ACM SIGMETRICS, San degree in Telecommunication Engineering and the
Diego, CA, Jun. 2003, pp. 90–101. Ph.D. degree in electronic and telecommunication
[6] S. Deb, S. Shakkottai, and R. Srikant, “Stability and convergence of engineering, both from Politecnico di Torino, Italy,
TCP-like congestion controllers in a many-flows regime,” in Proc. IEEE in 2000 and 2004, respectively.
INFOCOM, San Francisco, CA, 2003, pp. 884–894. In 2002, he was a visiting scholar with the Net-
[7] S. Shakkottai and R. Srikant, “How good are deterministic fluid models works group of the University of Massachusetts,
of internet congestion control?,” in Proc. IEEE INFOCOM, New York, Amherst, and in 2004 he held a Postdoctoral position
Jun. 2002, pp. 497–505. at Rice University, Houston, TX. His research
[8] P. Tinnakornsrisuphap and A. Makowski, “Limit behavior of ECN/RED interests are in the field of performance evaluation
gateways under a large number of TCP flows,” in Proc. IEEE IN- of wired and wireless communication networks.
FOCOM, San Francisco, CA, 2003, pp. 873–883.
[9] F. Baccelli, D. R. McDonald, and J. Reynier, “A mean-field model for
multiple TCP connections through a buffer implementing RED,” Per-
form. Eval., vol. 49, no. 1/4, pp. 77–97, 2002. Paolo Giaccone (S’00–M’02) received the Dr.Ing.
[10] S. Floyd and V. Jacobson, “Random early detection gateways for con- and Ph.D. degrees in telecommunications engi-
gestion avoidance,” IEEE/ACM Trans. Netw., vol. 1, no. 4, pp. 397–413, neering from Politecnico di Torino, Italy, in 1998
Aug. 1993. and 2001, respectively.
[11] V. Jacobson, “Congestion avoidance and control,” in Proc. ACM SIG- He is Assistant Professor in the Electronics De-
COMM, Vancouver, Canada, Sep. 1988, pp. 314–329. partment, Politecnico di Torino. During the summer
[12] L. Zhang and D. Clark, “Oscillating behavior of network traffic: A Case 1998, he visited the High Speed Networks Research
study simulation,” Internetworking: Research and Experience, vol. 1, Group at Lucent Technology, Holmdel, NJ. During
no. 2, pp. 101–112, 1990. 2000–2001 and during summer 2002, he visited the
[13] M. Ajmone Marsan, M. Garetto, P. Giaccone, E. Leonardi, E. Schi- Electrical Engineering Department, Stanford Univer-
attarella, and A. Tarello. Using partial differential equations to model sity. he held a Postdoctoral position at Politecnico di
TCP mice and elephants in large IP network. [Online]. Available: Torino between 2001 and 2002, and during summer 2002 at Stanford Univer-
http://www.tlc-networks.polito.it/database/ricer.htm sity. His main area of interest is the design of scheduling policies for high-per-
[14] A. Feldmann and W. Whitt, “Fitting mixtures of exponentials to long-tail formance routers.
distributions to analyze network performance models,” in Proc. IEEE
INFOCOM, Kobe, Japan, 1997, pp. 1096–1104.
[15] M. Garetto and D. Towsley, “Modeling, simulation and measurements Emilio Leonardi (S’94–M’99) received the Dr.Ing.
of queueing delay under long-tail internet traffic,” in Proc. ACM SIG- degree in electronics engineering and the Ph.D. de-
METRICS, San Diego, CA, Jun. 2003, pp. 47–57. gree in telecommunications engineering, both from
[16] S. B. Fredj, T. Bonald, A. Proutiere, G. Régnié, and J. Roberts, “Statis- Politecnico di Torino, Italy, in 1991 and 1995, respec-
tical bandwidth sharing: A study of congestion at flow level,” in Proc. tively.
ACM SIGCOMM, San Diego, CA, 2001, pp. 111–122. He is currently an Associate Professor at the Di-
partimento di Elettronica, Politecnico di Torino. In
1995, he visited the Computer Science Department,
University of California, Los Angeles (UCLA);
in summer 1999, he joined the High Speed Net-
works Research Group at Bell Laboratories/Lucent
Marco Ajmone Marsan (F’99) holds degrees Technologies, Holmdel, NJ; in summer 2001, the Electrical Engineering
in electronic engineering from the Politecnico di Department of the Stanford University; and finally in summer 2003, the IP
Torino, Torino, Italy, and the University of California Group at Sprint, Advanced Technologies Laboratories, Burlingame, CA. His
at Los Angeles (UCLA). In 2002, he was awarded research interests are in the fields of performance evaluation of communication
an Honoris Causa degree in telecommunication net- networks, switching architectures, and all-optical networks.
works from the Budapest University of Technology
and Economics, Budapest, Hungary.
He is a Full Professor at the Electronics De-
Enrico Schiattarella (S’01) received the Dr.Ing.
partment of Politecnico di Torino, in Italy. Since
degree in electrical engineering from Politecnico di
September 2002, he has been the Director of the
Torino, Italy, in 2002, where he is currently pursuing
Institute for Electronics, Information Engineering
the Ph.D. degree.
and Telecommunications of the National Research Council. From November
From October 2001 to June 2002, he visited Cisco
1975 to October 1987, he was at the Electronics Department of Politecnico di
Systems, in summer 2003, Alcatel R&I, Marcoussis,
Torino, first as a Researcher, then as an Associate Professor. From November
France, and from March to November 2004, the
1987 to October 1990, he was a Full Professor at the Computer Science
Server Interconnect Fabrics group at IBM Zurich
Department of the University of Milan, Italy. During the summers of 1980 and
Research Lab, Switzerland. His main research inter-
1981, he was with the Research in Distributed Processing Group, Computer
ests are in high-performance switching and routing.
Science Department, UCLA. During summer 1998, he was an Erskine Fellow
at the Computer Science Department of the University of Canterbury, New
Zealand. He has coauthored over 300 journal and conference papers in the
areas of telecommunications and computer science, as well as the two books
Performance Models of Multiprocessor Systems (MIT Press), and Modeling Alessandro Tarello (S’03) received the M.Sc.
with Generalized Stochastic Petri Nets (Wiley). He has been the principal degree in communication engineering from Politec-
investigator in national and international research projects in the field of nico di Torino, Torino, Italy, in 2002, where he is
telecommunication networks. His current interests are in the fields of perfor- currently pursuing the Ph.D. degree.
mance evaluation of communication networks and their protocols. From January to December 2004, he visited the
Dr. Ajmone Marsan received the Best Paper Award at the Third International Laboratory for Information and Decision Systems,
Conference on Distributed Computing Systems, Miami, FL, in 1982. He is a MIT, Cambridge, MA.
corresponding member of the Academy of Sciences of Torino. He participates in
a number of editorial boards of international journals, including the IEEE/ACM
TRANSACTIONS ON NETWORKING and Computer Networks.

View publication stats