.. .

.~.
J. Phys. A Math. Gen. 27 (1994) L4174421. Printed in the UK

LETTER TO THE EDITOR

Uf noise in computer network traflic

Istvin Csabai
Department of Atomic Physics. ESW& University, Budapest Pus!& U. 5-7. H-1088, Hungary

Received 22 April 1994

Abstract A huge number of compum connected together form an international network

(Intemet). As a mult of their communication, uains of data packages travel through this
network: due to the limited bmsfer rate and the behaviour of data source agents, the aaRc
fluctuates. Our measurements show that the power spectrum of the roundaip time tetwem two
points on Intemet is llf-like over a broad range. Models of collective phenomena, such as
highway Vamc models could be appropriate for d-bing this pehaviour.

The main goal of this paper is to present our first results on Internet traffic measurements.
Beside this, we would also like to give a rough picture of Internet, and to call to researchers’
attention the close analogy between the basics of data traffic on Internet and vehicle traffic
on highways.
The international network of computers (Internet) has gradually reached a size where
methods of statistical physics, particularly of complex systems, can play a role. Up to
October 1993 more than 2 x lo6 hosts have been connected to Internet [I], and this number
seems to double in slightly more than a year. The topology of the network connecting
these machines is not a regular one, however, it can be regarded as a hierarchical, tree-like
stiucture, although some loops can also occur. Hosts use this network to send and get data
we use in our everyday work computer mail, data transfer (FTP), newsgroups, remote logins,
etc. At the vortices of the network there are special switches, called gateways, routers,
bridges and nameservers. They help .to route the data packets to their destinations. These
devices have significant autonomy inn finding the optimal way from source to destination,
breaking up the messages into smaller datagams, buffering the packets, etc. The interaction
of nodes can produce collective phenomena which cannot be foreseen from an examination
of individual parts of the system-statistical physics is often used in the understanding of
such complex systems [2].
Let us consider a connection between two points on Internet, for example, I would like
to transfer a file from a distant FTP site to my computer. We will not discuss the routing
process and we suppose that the route connecting the two machines is not changing during
this time, although this is not always true due to dynamical optimization of the connection
speed [3]. This connection is not galvanic as in an old telephone system, where wires
are directly connected at the crossbars. Messages go from gateway to gateway (figure 1).
Since the lines connecting the gateways have different speeds and they are often not capable
of transfering the data at the rate it is arriving’at the gateway, the latter has to buffer the
information. The capacity of the buffers is not infinite, so they can overtlow and the gateway
can get congested, so that the connection slows down, and data packets are ‘dropped‘.

0305-4470/94/120417+05$19.50 0 1994 IOP Publishing Ltd L417

L418 Letter to the Editor

mnneslforuIo oiher pans 01 lntemet

Figure 1. Schematic p i a m of a connection via Inmet. Tnnsfereddaa jumps from *way

to gateway, and on wait there if the output rate is not large enough.

To avoid the collapse of Internet several congestion control mechanisms have been
suggested [3-51, such as source quench, random drop, fair queueing, etc. We will shortly
discuss the most widely used source quench gateway protocol and the slow-start end-system
congestion control policy [4]. If a gateway has to discard a datagram because of lack of
buffer space, it sends the data back to its source, a so-called sourcequench message. A
destination host can send similar messages, if it is not fast enough to process the arriving
data. This message is a request to the source to cut back the rate at which it is sending
data. The strategy of the source host is the slow-start mechanism. It gradually increases its
demand until a source-quench message arrives. In response to this the source host decreases
the load and then begins an exploratory increase again. The hope is that this cycle will
keep the total demand around the optimal level.
To investigate this we have performed measurements on the speed of the network
between two points. We have measured the so-called round-trip time (RTT) between a
workstation at Eotvos University, Hungary (hercules.elte.hu)and a distant FTP site (funet&
in Finland. The two machines can be regarded as distant, not geographically, but in network
metrics. There are 15 gateways between the two hosts, and this number is not usually larger
for a host in a physically more distant place, e.g. in Australia. We used the ping command
of the Unix operating system, which sends a short datagram every second to the given
destination, and reports the return time measured in milliseconds. Our method was able
to collect m s shorter than 2000 ms. This RTT value can be regarded as the reciprocal of
the actual transfer speed, measured in datagramskecond, so we get a time series of inverse
speed values. The data were collected each second during a two week period. The total
number of RTTs was 821 383.
The changing of Rfi in time can be seen in figure 2. The data used in this plot is
averaged over a 10 minute time period because of the resolution of the graphics. Since
there is only one hour difference in time zone between the two endpoints, one can see the
daily and even the weekly periodicity of the network speed.
We have tried some simple methods,to analyse this time series. The embedding method
[7] shows that there is no simple low-dimensional (up to seventh dimension) chaotic attractor
behind the behaviour. To avoid possible errors due to missing datagram RTTs from our
measurement (22 percent of datagrams had not returned), we have hied different methods
[SI to calculate the power spectrum. All the methods gave nearly the same curve. Apart
from the daily periodicity, the power spectrum of data shows l/f-like behaviour in the
whole time domain, figure 3. The fit for data (without the daily periodicity peak) gives a
slope of -1.15.
The description given in the previous paragraphs: packets running along the edges of a
 Letter to the Editor L419
I I
I ’
! I !

Sat Sun Mon TUB Wed Thu Fri Sat Sun Mon Tue Wed Thu

Figure 2. The chaoging of the mund-trip time behvm hvo points of totemet from the 2nd m
the 14th of October 1993. Vertical lines separate the days f” each other. Daily and weekly
periodicity can be easily recognized.

complex graph with crosspoints can .evoke the picture of a road system, with cars stopping
at crossings and moving towards their destination. Measurements of highway traffic [9]
also gives I/f spectra. Recently, several models [lo, 111 of highway traffic were proposed.
This analogy was used to model density waves in the flow of granular media [12] and it
can also be used for computer networks, since the basics of packet sending, slowing and
acceleration, are very similar to the behaviour of cars in these simple models.
The prototype for a one lane highway 1111 is a one-dimensional cellular automaton.
Each cell can have zero (no car) or some finite value U (car with speed U). A vehicle with
+ +
velocity U is either accelerated (U + U I), when there are at least U 1 empty cells in
front of it, or it is slowed down (U + Ax) when there are only Ax empty cells in front of
it. After this velocity update, the car is advanced U positions in one direction. The direction
of movement is the same for all cars, and periodic boundary conditions are used in the
model.
Let us now consider a heavily loaded computer network. We will discard all the side
branches, and concentrate on the two-point connection described. Since we only want to
model the internal traffic, not the behaviour of sources, the different nodes: gateways,
sources and destinations are n’ot distinguished. With the dictionary given below we can
easily map computer network traffic to a highway traffic model:
speed = number of bytes sent per second
displacement (integral of speed over time) = the total amount of data sent
distance to next car = free space in next buffer.
+
So, node i sends ut bytes per second to node i 1, and the total number of bytes sent
+
successfully is x i . In one timestep this means that xi + xi U;. Data goes from node to
L420 Letter to the Editor
10

0.1

c
2 0.01
3
2m
0
.

5x o.ool

0.0001

19-05

18-06
, ,,I ,

18-06 18-05 0.0001 0.001 0.01 0.1 1

frequency (I/$

Figure 3. The power s p e m m of fmdata. The fit (without the daily peak al the lefl side) gives
a slope O f -1.15.

node. Each of them has buffer size b, and b: of them are initially empty. We suppose, that
this buffer is large compared to the outcoming data flow, so it is never empty. The sender
always tries to increase its speed (up to a maximum):
U, -+ ui +I. (1)
If there is no space in the buffer of node i + 1, since more data is received than sent
X~+I + bi+l0 <.xi +vi + 1. (2)
Source-quench messages reduce the speed of the sender to the number of empty spaces in
the buffer i 1: +
uj + xj+l + b!+] -xi . (3)
This is exactly the same as the highway model mentioned above. Of course, to get
better models one has to modify the above assumptions, distinguish between the behaviour
of sources and gateways, handle buffers more realistically, make a network rather than
singlelane model, etc.
Developing such models, more appropriate for Intemet, including its structure, the
behaviour of the data sources [131 and the different routing and,congestion control methods
is the subject of our future work. It could also be interesting to measure, for example, the
occupancy versus flow diagram [IO], which shows very good accordance between highway
models and real measurements.
Designers of Intemet devices usually take into account only the individual agents of
the network, but no collective effects. The models of highway traffic clearly show that the
formation of a traffic jam on an overloaded road is a collective phenomenon, not a direct
 Letter to the Editor L421
effect of the behaviour of a single driver. We hope that future models-apart from the
theoretical interest-an help designers to produce more suitable network devices.

Acknowledgments

The author is indebted to I M Jhosi, A Czir6k, T Geszti and T Vicsek for helpful discussions
and careful reading of the manuscript The work was supported partly by the Hungarian
Academy of Sciences (OTKA F4437).

References
[I] Lottor M 1993 Intemd Domnin Suwey InterNlC Information Services
121 Hogg T and Hubermann B A 1987 Phys. Rep. 56 227
131 Barden R and Postel J 1987 Requiremenls for Inremet Galeways (RFC-ZW9J USC Information Science
Institute
[41 Postel I 1981 Infernet Control Message Proloco1 (RFC-792) USC I n f o d o n Science Institute
[5] Nagle J 1984 Congestion Conml in IPLTCP lntemehvorks (RFC-8%) Ford Aerospace
[6] Mankin A and Ramalaishnan K 1991 Galeway Congestion Control Survey (RFC-ZZ54J Digital Equipment
Corpom+m
[7] Packard N H,Crutchfield J P, Fanner J D and Shaw R S 1980 Phys. Rev. Len. 45 712
[SI Press W H.Teukolsky S A, Veaerling W T and Flannery B P 1992N u m e r i d Recipes (Cambridge: Cambridge
Universir, Press)
191 Musha T and Higuchi H 19761qpm. 1.Appl. Phys. 15 1271
[IO] Nagel K and SchreckenbergM 1992 J. Physique I 2 2221
[I I] Nagel K and H e m " H I 1993 Physic= 199A 254
[I21 Leibig M 1994 Phys. Rev. E 49 184
[I31 Caceres R, Danzig P B, Iamin S and Mitlel D J 1991 Chnmferistics of Wide-muTCPl7P Conversaions
ACM SIGCOMM