Content uploaded by Wolfgang Mühlbauer
Author content
All content in this area was uploaded by Wolfgang Mühlbauer on Aug 05, 2015
Content may be subject to copyright.
Content uploaded by Steve Uhlig
Author content
All content in this area was uploaded by Steve Uhlig
Content may be subject to copyright.
IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 28, NO. 8, OCTEMBER 2010 1
Evolution of Internet Address Space Deaggregation:
Myths and Reality
Luca Cittadini, Wolfgang Mühlbauer, Steve Uhlig, Randy Bush, Pierre François, Olaf Maennel
Abstract—Internet routing table size growth and BGP update
churn are two prominent Internet scaling issues. There is
widespread belief in a high and fast growing number of ASs
that deaggregate prefixes, e.g., due to multi-homing and for the
purpose of traffic engineering [1]. Moreover, researchers often
blame specific classes of ASs for generating a disproportionate
amount of BGP updates. Our primary objective is to challenge
such widespread assumptions (“myths”) and not solely to confirm
previous findings [1]–[3]. Surprisingly, we find severe discrepan-
cies between existing myths and reality. According to our results,
there is no trend towards more aggressive prefix deaggregation or
traffic engineering over time. With respect to update dynamics,
we observe that deaggregated prefixes generally do not generate
a disproportionate number of BGP updates, with respect to their
share of the BGP routing table. On the other side, we observe
much more widespread traffic engineering in the form of AS path
prepending and scoped advertisements compared to previous
studies [1]. Overall, our work gives a far more positive picture
compared to the alarming discourses typically heard [1], [2], [4]:
The impact of “bad guys” on routing table size growth and BGP
churn has not changed for the worse in recent years. Rather, it
increases at the same pace as the Internet itself.
Index Terms—Routing table growth, update churn, address
deaggregation, traffic engineering
I. INTRODUCTION
THE RESEARCH community is now accustomed to
alarmist discourse regarding the lack of scalability of
the current routing system. By now, full BGP routing tables
contain more 300,000 prefixes and continue to grow super-
linearly [3]. This goes along with a steady increase in the rate
of changes to the routing and forwarding tables. Having to
handle big tables which need frequent re-computations, routers
waste more and more resources to maintain their routing tables
rather than to forward packets. Considerable efforts have been
made to quantify the evolution of routing table size and BGP
update rates [2], [4], to characterize the BGP routing table
growth [3] or to study the dynamics that can be observed in
Internet routing [5], [6].
Manuscript received ; revised .
Luca Cittadini is with the Department of Computer Science and Automa-
tion, Roma Tre University, Rome, Italy (e-mail: luca.cittadini@gmail.com).
Wolfgang Mühlbauer is with ETH Zürich, Zürich, Switzerland (e-mail:
wolfgang.muehlbauer@tik.ee.ethz.ch).
Steve Uhlig is with Technische Universität Berlin/Deutsche Telekom Lab-
oratories, Berlin, Germany (e-mail: steve@net.t-labs.tu-berlin.de).
Randy Bush is with Internet Initiative Japan, Tokyo, Japan (e-mail:
randy@psg.com).
Pierre François is with Université catholique de Louvain, Louvain-la-
Neuve, Belgium (e-mail: pierre.francois@uclouvain.be).
Olaf Maennel is with Loughborough University, Loughborough, UK (e-
mail: olaf@maennel.net).
Digital Object Identifier 10.1109/JSAC.2010.1010xx.
Address space fragmentation, i.e., the use of a large num-
ber of small, potentially overlapping prefixes in the routing
system, is widely seen as one major driving force behind
the inflation of routing tables and high BGP update rates.
Fragmentation of the IPv4 address space can be caused by
either (i) the Regional Internet Registries (RIR), who allocate
small and unconnected blocks of IPv4 addresses to Internet
Service Providers (ISP), or (ii) ISPs who use the Border
Gateway Protocol (BGP) to inject multiple more specific
prefixes instead of, or in addition to, their assigned prefix
blocks.
For the latter case, the current routing system poses an
inevitable conflict: While individual domains inject more
specifics for what they perceive as legitimate reasons, e.g.,
to implement multi-homing or to better control how traffic
enters their own network1, other Autonomous Systems (ASs)
pay a fee, e.g., in terms of increasing convergence times, for
having to deal with an “inflated” routing table. We believe
that a deeper understanding of the root causes of the Internet’s
routing scaling problems is urgently needed, in particular in
the light of recent interest in re-designing the Internet routing
architecture using “clean-slate” approaches [7].
We, in this paper, seek to shed light on the evolution of
address space deaggregation and of update dynamics. For
this purpose, we rely on BGP routing tables and update
traces from RIPE [8] and RouteViews [9], as well as on
the official allocation files as available at the RIRs [10]. An
important contribution of our work is to extract information
about traffic engineering practices from observed routing data
and correlate this information with the deaggregation that we
can observe for prefixes in BGP routing tables. Overall, our
primary interest is not just to confirm previous findings [1]–
[3]. Rather we challenge wide-spread assumptions that are
frequently made whenever researchers discuss the root causes
of the current scaling problems. Amongst others, such “myths”
are:
(i) There exists a large number of “bad guys”, i.e., ASs
that strongly contribute to routing table growth and high
update churn via aggressive use of prefix deaggregation.
(ii) Recently, there has been a strong shift towards a more
widespread use of prefix deaggregation. This can be
explained by an increasing number of multi-homed ASs
that apply traffic engineering.
(iii) With respect to BGP dynamics, there are differences
between various types of ASs. For example, BGP update
1Note that there exist more reasons why ASs may deaggregate their address
space. For example, ASs have announced sub-prefixes (e.g., /24) of their
assigned prefix block (e.g., /16) to protect themselves against prefix hijacking.
0733-8716/10/$25.00 c
2010 IEEE
2 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 28, NO. 8, OCTEMBER 2010
patterns differ depending on the business role of an AS,
e.g., transit provider vs. access network or stub vs. tier-1.
Our results reveal discrepancies between such myths and
reality. Surprisingly, there is no trend towards more aggressive
prefix deaggregation or traffic engineering over time. Most
of the prefixes advertised in BGP routing tables actually
match the address block that has been allocated by the RIRs.
Yet, there exists a roughly constant number of (“bad”) guys,
which (strongly) deaggregate their address space for possibly
legitimate reasons, e.g. insufficient support for traffic engineer-
ing [1] in the current Internet. With respect to BGP updates,
we cannot claim that a specific class of prefixes or origin ASs
misbehave more than others. Similar to routing table size, we
cannot identify any trends, e.g., towards deaggregated prefixes
causing a lot of churn and thus frequent re-computations of
routing tables. All in all, our analysis shows that the impact of
“bad guys” on routing table size growth and BGP churn has
not changed for the worse in recent years: rather, it increases
at the same pace as the Internet growth itself.
The remainder of this paper is structured as follows: We
start with a study of the allocations made by the RIRs
in Section II. Then, Section III turns to deaggregation as
observed only from static BGP data. In Section IV we
compare allocated with observed IPv4 addresses, before we
introduce in Section V a classification scheme to estimate the
prevalence of traffic engineering and correlate this information
with our results on prefix deaggregation from the preceding
sections. Finally, Section VI explores whether specific classes
of prefixes or ASs exhibit significant misbehavior in terms of
BGP dynamics. This allows us to understand whether or not
high updates rates are a direct consequence of an increasingly
fragmented IPv4 address space. We summarize and conclude
in Section VIII.
II. ALLOCATION OF ADDRESS SPACE
The current IP addressing architecture requires central
coordination to ensure that different networks use unique
non-overlapping IP prefixes. To this end, IP addresses are
allocated by the Internet Assigned Numbers Authority (IANA)
from pools of unused address space and delegated to the
appropriate Regional Internet Registries (RIR). An Internet
Service Provider (ISPs) that requests IP address space from
these registries2is called Local Internet Registry (LIR) [11].
By the end of 2008 some 75% of the total usable IPv4
address space have been allocated. Current estimates for
address space exhaustion are somewhere in 2011 for IANA
address space and 2012 for RIRs [12]. The goal of this section
is to investigate how the allocation policy adopted by RIRs has
accommodated the topological growth of the Internet and to
what degree it has promoted the inflation of routing tables.
The data sources we use for our analysis are the official
allocation files [10] provided by the RIRs. These files sum-
marize the current state of allocations and the assignment of
Internet number resources. Figure 1 presents the evolution
of the distribution of allocated blocks. The stacked area plot
shows for each year between 1982 and 2008 the number of
address blocks allocated up to and including that year.
2In the APNIC and LACNIC region IP address allocations are mainly
coordinated by National Internet Registries (NIRs)
Fig. 1. Breakdown of prefix blocks allocated up to a certain year.
Overall, allocations are dominated by /24 and /16 blocks,
especially during the classful era of the Internet before 1994.
/24 allocations still represent more than 40% of all allocated
blocks today. Since 1994, the number of /24 and /16 blocks
has been rather stable, while intermediate sized blocks (/19,
/20 and /23) have been slowly increasing. This was made
possible by the introduction of Classless Inter-Domain Routing
(CIDR) [13] in 1993. Indeed, registries started to allocate
CIDR blocks around 1994/1995 and, with CIDR, could al-
locate blocks that actually matched the expected use of the
address space [14]: On the one hand, there was the fear that
the class B space (/16) would soon run out, since it was
consumed at a massive rate. On the other hand, some ISPs
were afraid of too many small allocations (/24) that would
tremendously inflate their routing tables and induce high costs
for new router hardware. As a compromise, RIRs started to
allocate intermediate size blocks, e.g., /19.
We point out that shorter prefix lengths correspond to
significant fractions of the IPv4 address space although they
only account for a negligible number of allocations: at the
end of 2008, 50% of the total IPv4 address space has been
allocated in blocks of length between /8and/15.
In summary, the historical evolution of IPv4 address space
allocation indicates a shift towards intermediate sized pre-
fixblocks(/19-/23) since 1994. We conjecture that RIRs
switched to a more parsimonious allocation policy to accom-
modate Internet’s growth. Various allocation policies which
are being discussed and adopted in the RIRs [15] envision
much smaller sized blocks, e.g., /24 to /29 in an attempt
to allow a long period of entry for newcomers despite the
depletion of the main free pool of IPv4 address space. This is
a change in the very conservative policies which attempted to
minimize routing table impact, but are believed to have created
barriers to entry.
III. OBSERVED ADDRESS SPACE FRAGMENTATION
We now turn to actually announced prefixes and observed
deaggregation, an “avoidable” cause for the explosion of
routing table entries. We investigate if patterns of deaggre-
gation have changed recently (Section III-B) and if ASs of
different business types follow different strategies with respect
CITTADINI et al.: EVOLUTION OF INTERNET ADDRESS SPACE DEAGGREGATION: MYTHS AND REALITY 3
routing table
10.1/16 top
10.1.1/24 deaggregated
10.1.2/24 delegated
10.2/16 lonely
AS 10
AS 20
AS 30
10.1/16
10.1.1/24
10.1.2/24
10.2/16
route
collector
p2p
c2p
Fig. 2. BGP prefix classification.
to deaggregation of assigned prefix blocks (Section III-C). To
this end, we classify prefixes observed in BGP table dumps
according to the inclusion relationships between prefixes, i.e.,
which one is more or less specific than another (Section III-A).
A. Prefix Classification
To study the advertised IPv4 address space, we rely on BGP
table dumps from RIPE [8] and RouteViews [9]. Based on the
presence of overlapping prefix blocks, we assign every prefix
in the routing table of our observation points to one of the
following classes:
•Lonely:aprefix that does not overlap with any other
prefix.
•Top:aprefix that covers one or more smaller prefix
blocks, but is not itself covered by a less specific.
•Deaggregated:aprefixthatiscoveredbyalessspecific
prefix, and this less specific is originated by the same AS
as the deaggregated prefix.
•Delegated:aprefixthatiscoveredbyalessspecific, and
this less specific is not originated by the same AS as the
delegated prefix.
For illustration purposes, let us assume that the complete
Internet consists of only three ASs, see Figure 2. There is a
customer-to-provider (c2p) link between AS 10 and AS 20,
i.e., AS 20 is a customer of AS 10. Moreover, there is a peer-
to-peer (p2p) link between AS 10 and AS 30. Therefore, the
following prefixes are observed at the observation point in
AS 20: 10.1/16 (origin: AS 10), 10.1.1/24 (origin: AS 10),
10.1.2/24 (origin: AS 20) and 10.2/16 (origin: AS 30).
Based on this information, we classify 10.2/16 as lonely,
10.1/16 as top, 10.1.1/24 as deaggregated and 10.1.2/24
as delegated. We point out that delegated corresponds to a
provider aggregatable (PA) addressing model: A customer AS
is reachable via its provider or, to put it differently, the AS path
observed for a delegated prefix is going through the provider
from which the address space is sub-allocated.
One important difference between our classification and the
one presented in [1] is that we distinguish the case where the
AS that originates a more specificprefix is not the same as the
AS that originated the covering prefix. In that case, we talk
about delegation. This difference is important if one wants to
realistically quantify the portion of BGP routing tables that
Fig. 3. BGP prefix classification distribution.
could be shrinked by aggregating more specificprefixes. In
fact, we point out that delegated prefixes cannot be aggregated
into their less specific without impairing the ability of some
ASs to be reachable via multiple providers.
B. Prefix Deaggregation
We study multiple observation points for RIPE [8] and
RouteViews [9] data, but only find trivial differences with
respect to our prefix classification of Section III-A. In the
following discussion, we rely on an observation point from
RouteViews inside Level 3 (AS 3356). Since this router is
part of the default-free zone (DFZ) of the Internet, it provides
a representative picture of all globally-routable prefixes.
Figure 3 shows the evolution of these four prefix classes
from 2001 to 2008. About 40% of all prefixes are of the lonely
type. Top prefixes represent a stable 5% of the prefixes in the
BGP routing tables. Delegated prefixes account for 35% of the
prefixes in 2001, declining to a bit more than 20% in 2008.
The decreasing fraction of delegated prefixes is matched by an
increasing fraction of deaggregated prefixes, which increase
from less than 20% in 2001 to more than 30% in 2008.
Clearly, it is possible to shrink routing tables by aggregat-
ing or filtering deaggregated and delegated prefixes. Lonely
prefixes, on the other hand, are not aggregatable based only
on the content of the BGP routing table. Consistent with other
studies [2], [4], we find that, in theory, the current BGP routing
tables could be reduced by a factor of slightly more than 2 if
delegated and deaggregated were aggregated into top prefixes,
leaving only non-overlapping prefixes in the routing tables.
However, note that delegated prefixes announced by multi-
homed ASs cannot be aggregated without losing the ability to
load balance incoming traffic among multiple providers.
From Figure 3 we learn that the combined fraction of dele-
gated and deaggregated prefixes has remained approximately
constant over the years. However, there is a shift from dele-
gated to deaggregated prefixes. We speculate that this trend
reflects the increasing popularity of provider-independent (PI)
addresses, which avoid provider lock-in and renumbering
issues when changing the provider.
4 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 28, NO. 8, OCTEMBER 2010
TAB LE I
BREAKDOWN OF PREFIX CLASSES WITHIN AS TYPES AS DEFINED IN [16].
AS class Prefix type #Prefixes %ofprefixes
in class
Enterprise
Customers (EC)
lonely 57,567 44.7%
delegated 35,819 27.8%
deaggregated 29,459 22.9%
top 5,893 4.6%
Content/Access/-
Hosting Provider
(CAHPs)
lonely 5,884 54.2%
delegated 1,260 11.6%
deaggregated 2,700 24.9%
top 1,008 9.3%
Small Transit
Provider (STP)
lonely 28,260 38.3%
delegated 13,206 17.9%
deaggregated 27,131 36.8%
top 5,147 7%
Large Transit
Provider (LTP)
lonely 5,456 50.2%
delegated 1,695 15.6%
deaggregated 2,505 23.1%
top 1,211 11.1%
C. Business Type of Originating ASs
According to [16] not all parts of the Internet grow at
the same rate. It is widely agreed that the Internet has a
tiered structure [17] which reflects AS business relationships,
e.g., [17]–[19]. A few tier-1 ISPs form the core.Alarger
number of transit providers buy service from other providers,
including tier-1 providers, and provide connectivity to other
ASs. Stub ASs get their connectivity from transit or tier-1
providers and form the edge. Most of the growth of the AS-
level topology is due to networks at the edge [16]. Accord-
ingly, one might expect that deaggregation is also mostly due
to edge networks. To verify this, we now look closely at the
origin ASs where prefixes of different classes are injected into
the routing system.
We rely on the classification of ASs according to their
business type taken from [16]. Based on manual classifica-
tion, an initial training set and on machine-learning tech-
niques, Dhamdhere et al. distinguish between four different
types of ASs: Enterprise Customers (EC), Small Transit
Providers (STP), Large Transit Providers (LTP) and Con-
tent/Access/Hosting Providers (CAHP). Most of the ASs
(92%) are found to be EC ASs. For more details about this
classification, we refer to Section 4 of [16]. Generally, we will
assume that LTP and STP ASs form the “core” of the Internet
while CAHPS and ECs make up the “edge”. Yet, we point out
that this mapping is not strict in the sense that for example
certain parts of large LTP/STP networks are actually part of
the “edge”, e.g., DSL end-customers within large tier-1 ASs.
For each classified prefix (see Section III-A), we check the
business type of its originating ASs according to the categories
of Dhamdhere et al. [16]. Table I shows the breakdown of our
four types of prefixes into the four business classes of ASs.
The first column gives the AS class as defined in [16], the
second the prefixclassasdefined above, the third column
the number of prefixes originated by that class among those
originated by the ASs of the given type.
In Table I, we first notice that the majority of prefixes are
originated by EC and STP networks. Content/Access/Hosting
providers and large transit providers do not advertise a large
number of prefixes. Thus, most prefixes are advertised by
networks that are not in the “core.”
Irrespective of the business type of ASs, there is always a
large fraction of lonely prefixes, then delegated and deaggre-
gated follow, and finally top prefixes represent the smallest
fraction of prefixes. Still there are differences. For example,
STPs show a bit more deaggregation than other AS types; EC
networks have slightly more delegated prefixes than the other
AS types. CAHPs and LTPs do not have so many deaggregated
prefixes. This smaller number is only due to the fact that
these AS types advertise less prefixes. However, overall the
proportions between our four prefix types are similar for every
AS type.
This finding is important as it shows that there are no
fundamental differences in how ASs of different business types
announce prefixes in the Internet. In particular, it disproves the
widespread belief that edge networks, such as EC or CAHP
domains, tend to deaggregate more than other ASs. In this
respect, the predicted growth of the Internet at the edge [16]
does not worsen address space deaggregation as is widely
thought.
IV. ALLOCATED VS.OBSERVED PREFIXES
In the preceding two sections we have studied the frag-
mentation of allocated address space and the deaggregation of
announced prefixes as separate phenomena. In this section we
aim to understand the combined effect of these two aspects
on the routing tables in today’s Internet. Consequently, we
now investigate to what degree individual ASs deaggregate
the allocations they obtained from the registries, possibly
announcing multiple prefixes for each allocated address block.
Section IV-A defines categories that characterize the use of
allocated address blocks while Section IV-B presents the
results. Then, Section IV-C studies the ratio between the
number of announced prefixes and the number of allocated
address blocks for each AS,
A. Use of allocated blocks – classification
We again rely on the official allocation files [10], already
used in Section II. In addition, we obtain routing table
snapshots for AS 3356 for every November 1st between
the year 2001 and 20083. For each allocated address block,
we determine all the prefixes from the routing tables which
overlap with that allocated block. If the routing table contains
at least one such prefix, we classify the allocated block into
one of the following five categories:
•Only root: The complete allocated address block (called
“root prefix”) is announced and nothing else.
•root/MS-complete: The root prefix and at least two sub-
prefixes are announced. The set of all sub-prefixes spans
the whole root prefix.
•root/MS-incomplete: The root prefix and at least one sub-
prefix is announced. Together, the set of announced sub-
prefixes does not cover the root prefix.
3This is the same data source as used in Section III, but we only rely on
the observation point in AS 3356.
CITTADINI et al.: EVOLUTION OF INTERNET ADDRESS SPACE DEAGGREGATION: MYTHS AND REALITY 5
2001 2002 2003 2004 2005 2006 2007 2008
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Year
Fraction o
f
allocated/seen blocks
only root
root/MS−complete
root/MS−incomplete
no root/MS−complete
no root/MS−incomplete
Fig. 4. Use of address blocks allocated by year.
•no root/MS-complete: The root prefix is not announced.
However, there are at least two sub-prefixes which to-
gether cover the complete root prefix.
•no root/MS-incomplete: The root prefix is not announced.
There is at least one sub-prefix. Taking all sub-prefixes
together, they do not cover the complete root prefix.
B. Is Deaggregation Popularity Increasing?
To study the use of allocated blocks we consider the time
between November of a specific year and November of the
previous year: for each address block that has been allocated
during that time, we compare against BGP routing tables
and classify the allocated block according to the categories
described in Section IV-A. Figure 4 shows the breakdown of
newly allocated blocks in fractions.
We observe that the majority (more than 60%) of address
blocks that have been newly allocated between 2001 and
2008 fall into the “only root” class. Then the “no root/MS-
incomplete” class follows. All the curves remain roughly con-
stant over time. Apparently, there are no significant changes in
how newly allocated blocks are used by the people requesting
them. Moreover, Figure 4 even reveals that, in recent years,
we observe a slightly decreasing popularity of address space
deaggregation. In fact, the percentage of “only root” alloca-
tions is slightly increasing since 2001. Here we point out that
“only root” address blocks are almost exclusively announced
as “lonely prefixes” when comparing the classification of
allocated blocks (Section IV-A) with the classification of pre-
fixes (Section III-A). Overall, from a general perspective, we
cannot confirm a trend towards increasing fragmentation, with
more and more sub-prefixes being announced for individual
allocated blocks.
C. “Bad” Guys
Results in the previous section show that most allocations
are announced exactly as received by RIRs. This must not
mislead us into thinking that address block deaggregation is
not a problem: even if the percentage of deaggregated blocks is
small, the corresponding number of prefixes might be relevant.
In order to take this into account, we define the deaggre-
gation factor of an AS to be the ratio between the number
of announced prefixes and the number of allocated address
blocks. We investigate if there exist ASs where the number
of announced prefixes significantly exceeds the number of
address blocks that have been allocated to it. Although we call
such ASs “bad guys”, it is not our objective to blame such
ASs for the scaling problems of the routing system. After all,
there may be legitimate reasons for them to deaggregate their
available address space (see Section V).
Rather, our goal is to quantify how much “bad guys”
contribute to the routing load in today’s Internet and to check
if the number and impact of “bad guys” has gone for the worse
or not over time.
Figure 5 shows the deaggregation factor for Enterprise Cus-
tomers (EC) and Large Transit Providers (LTP), respectively.
For a point (x,y)on the curve, the xvalue on the logarithmic x-
axis indicates the percentage of ASs that have a deaggregation
factor of more than y. As shown in Figure 5(a), some EC
domains do deaggregate their allocated space by a factor of
up to one hundred. About 1% of EC ASs split each allocation
into more than 10 more specifics on average. More than
10% of all ASs deaggregate their allocations (i.e., they have
deaggregation factor greater than 1), with no notable trends
since 2001.
Figure 5(b) shows the completely different behavior of
LTP ASs: most of them advertise more than one prefix
per allocation on average, but their deaggregation factor
never exceeds 10. We run the same analysis on Small Tran-
sit Providers (STPs) and on Content, Access and Hosting
Providers (CAHPs). We find that STPs behave similarly to
LTPs, while CAHPs behave more like ECs.
According to our findings, there exist only a few “bad guys”
amongst EC and CAHP edge networks which deaggregate
their allocated address space a lot. Yet, we point out that this
situation has been the case since 2001, and probably even
earlier. A limited fraction of “bad guys” has existed since
quite some time and their total fraction has not significantly
increased despite the natural growth of the Internet at the
edge [16]. Overall, the vast majority of edge ASs refrain from
deaggregation, with more than 80% of EC domains in the
Internet advertising their allocations exactly as received by
routing registries.
Figure 5 helps to identify heavy deaggregators, namely
our “bad guys”. But what is their impact on routing table
size? To this end, Figure 6 illustrates the contribution of the
top deaggregators in terms of the prefixes they contribute to
the routing tables in today’s Internet. We sort all ASs by
decreasing deaggregation factor (x-axis) and display on the
y-axis the cumulative fraction of prefixes for which the top
6 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 28, NO. 8, OCTEMBER 2010
(a) Enterprise Customers (b) Large Transit Providers
Fig. 5. Deaggregation factor of ASs, differentiated by business type.
Fig. 6. Percentage of the IP prefixes injected by heavy deaggregators. A
data point (x,y)means that the x% top deaggregator ASs are responsible for
injecting y% of the prefixes in the routing table.
x% deaggregator ASs are responsible4.
We observe a skewed distribution: The 1% top deaggrega-
tors inject almost 10% of the prefixes in the Internet. ASs
that deaggregate their address space, which account for more
than 10% of ASs, are responsible for almost half of the
routing table size. Overall, such top deaggregators make up a
considerable fraction of the routing table entries. This justifies
the alarmist discourses about prefix deaggregation. However,
we point out that the fraction of prefixes announced by heavy
deaggregators has not strongly increased since 2001.
Our analysis contradicts the widespread belief that huge
prefix deaggregation is more dominant today than it has
been in the past: over time, address block deaggregation is
not becoming more popular (Figures 4 and 5) and heavy
deaggregators are not inflating routing tables at a growing
pace (Figure 6). We conjecture that the growth of the Internet
topology is actually the main factor that is the cause of the
widespread belief that the routing system is going worse than
4The y-axis does not span the 100% of prefixes since the deaggregation
factor metric does not capture delegated prefixes.
in the past.
It may be considered unacceptable that a limited fraction
of all ASs inflates the routing tables, damaging the whole
Internet. If we are to mitigate the impact of such ASs,
we believe that the Internet community should discuss the
reasons why these ASs are relying on deaggregation, and find
alternative ways for those ASs to do what they want without
such an impact on the routing system.
V. D EAGGREGATION AND TRAFFI C ENGINEERING
So far, we have been mainly concerned with understand-
ing the extent to which prefix deaggregation is applied and
whether deaggregation has become more popular or more
intense recently. Many researchers consider traffic engineering
as the main driver that causes network operators to split
available address blocks into multiple prefixes. For example, in
order to influence where traffic enters its network, an ISP can
announce different sets of more specificprefixes to different
upstreams.
In this section, we explore how prevalent such traffic
engineering actually is. An important contribution of our
work is to infer information about traffic engineering practices
from observed routing data and to correlate this information
with observed deaggregation, as discussed in the preceding
sections. Contrary to previous sections, we also consider AS
path information that is announced together with prefixes. In
Section V-A we propose a classification scheme that we use in
Section V-B to estimate the prevalence of traffic engineering
in the current Internet.
A. BGP Traffic Engineering – Classification
While traffic engineering in its general sense is about how
to optimize the performance of a network, this section focuses
on mechanisms that BGP provides to control route selection
and thus trafficflow to certain prefixes. In this regard, BGP
routing data allow us to identify two different types of traffic
engineering activities: we refer to them as AS-path prepending
and scoped advertisements.
AS-path prepending adds consecutive repetitions of the
same AS number in the AS-path attribute. The goal is to make
CITTADINI et al.: EVOLUTION OF INTERNET ADDRESS SPACE DEAGGREGATION: MYTHS AND REALITY 7
BGP announcements less appealing for the BGP decision
process, which prefers routes with shorter AS-path length.
AS path prepending provides some degree of control over
inbound traffic, since an AS Acan induce a peering AS B
to prefer other shorter routes over the prepended one that is
announced by Ato B. We only consider prepending performed
by the last two distinct AS hops on the AS path. Many
origin ASs either perform prepending themselves, or ask their
upstream providers to do it on their behalf through BGP
communities [20]. Prepending made by other ASs on the path
is unlikely to be related to traffic engineering activities at the
origin AS.
Another traffic engineering technique are scoped adver-
tisements. Here ASs selectively announce distinct prefixes to
different upstream providers. For example, some prefixes are
not advertised to selected upstreams, such that traffic destined
to those prefixes cannot be received via those upstreams. For
more details about BGP-based traffic engineering, see [21].
We are interested in understanding which ASs exhibit a
behavior that can be related to some form of BGP traffic
engineering. To this end, we investigate on a per-prefix basis to
what degree the two traffic engineering techniques described
above are applied in the current Internet. The data source we
rely on are the same BGP routing tables as used in Section IV,
but we consider all RouteViews RIBs, not only the view from
AS 3356. Thus, our analysis is not restricted to only two
vantage points as in [1]. Yet, visibility provided by observed
BGP data is inherently limited [22]. To cope with this, we
rely on a threshold that considers that some paths may not be
visible from the data. Limited visibility is especially relevant
for scoped advertisements, so our threshold requires that we
see scoping across at least 80% of the prefixes of an AS
before concluding that scoping is actually used. Overall, our
analysis provides a lower bound on the amount of actual traffic
engineering performed in the Internet.
The classification scheme that we propose is along three
dimensions. In addition to scoped advertisements and AS-
path prepending, we check if an AS has a single upstream5
or multiple upstream providers according to the AS path
information from our BGP routing information. Altogether,
this results in the following six categories:
(i) No TE - single upstream: no AS-path prepending nor
scoped advertisements are identified and the AS has a
single visible upstream.
(ii) No TE - multiple upstreams: no AS-path prepending nor
scoped advertisements are identified, and the AS has
multiple visible upstreams.
(iii) Prepending - single upstream: some of the prefixes origi-
nated by the AS are prepended, no scoped advertisements
are identified, and the AS has a single visible upstream.
(iv) Prepending - multiple upstreams: some of the prefixes
originated by the AS are prepended, no scoped adver-
tisements are identified, and the AS has multiple visible
upstreams.
(v) Prepending and scoped advertisements - multiple up-
streams: some of the prefixes originated by the AS are
prepended and scoped advertisements are identified.
5Note that the BGP data used can miss some upstream providers.
(a) All origin ASs
(b) Deaggregating origin ASs
Fig. 7. Quantifying BGP traffic engineering in observed routing data.
(vi) Scoped advertisements - multiple upstreams:noAS-path
prepending is identified and scoped advertisements are
identified.
B. Prevalence of BGP Traffic Engineering
Figure 7 quantifies the extent of traffic engineering practices
we can identify from our BGP data. Each origin AS is assigned
to one of the six traffic engineering classes from above. While
Figure 7(a) considers the evolution over time for all origin
ASs, Figure 7(b) only takes into account those ASs that
advertise deaggregated prefixes (according to the classification
in Section III-A).
As shown by Figure 7(a), approximately 20% of origin ASs
appear to be single homed. Despite some 80% of multi-homed
origin ASs, 50% of all origin ASs are multi-homed and do not
show evidence of applying any of the BGP traffic engineering
techniques introduced above. On the contrary, 30% of all ASs
fall into the categories related to traffic engineering, suggesting
that BGP-based traffic engineering plays a non-negligible role
in the Internet.
Contrary to [1] we take into account the evolution since
2001. Here, we observe that the proportion of origin ASs that
apply BGP deaggregation has only slightly increased from
11.3% in 2001 to 14.6% in 2008. Besides, we observe more
8 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 28, NO. 8, OCTEMBER 2010
ASs with single upstreams and without traffic engineering,
possibly reflecting the growth of the Internet at the edge [16].
Overall, our findings contradict the widespread belief that the
importance of BGP traffic engineering has massively increased
and threatens to collapse the current routing system.
Given the results of the previous sections, we investigate
whether there is any correlation between prefix deaggrega-
tion and traffic engineering. For example, is it true that
deaggregated prefixes are deaggregated for reasons of traffic
engineering, e.g., to influence where traffic enters the network?
The approach we take is to study how much traffic engineer-
ing is related to the different classes of prefixes defined in
Section III-A. For this purpose, we generate in Figure 7(b) a
similar plot to Figure 7(a), but only consider those prefixes
that are classified as deaggregated in Section III-A.
Comparing the two plots of Figure 7, it becomes evident
that there is a correlation between prefix deaggregation and
traffic engineering. Around 80% of the ASs for which we see
deaggregated prefixes in our routing tables fall into the traffic
engineering categories. This percentage is less than 40% for
2008 if we consider all ASs irrespective of the type of prefixes
that they originate, see Figure 7(a).
Our analysis is not limited to deaggregated prefixes only.
The 40% of lonely prefixes (see Section III-A) are apparently
not affected by BGP traffic engineering as defined in this
section. As expected, we do not observe scoped advertisements
for delegated prefixes, only prepending. This supports our
classification as scoped advertisements are expected to be
strongly related to deaggregation.
To sum up, given the differences in the two plots of
Figure 7, we conclude that traffic engineering is a significant
driver for address space deaggregation. However, based on
the evolution since 2001, we cannot identify any sudden shift
towards a more aggressive and widespread use of prefix deag-
gregation in order to achieve traffic engineering. This suggests
that it is not an increasing demand for traffic engineering that
inflates routing tables: rather, inflation is a consequence of the
topological growth of the Internet combined with the lack of
support for traffic engineering in the current routing system.
VI. BGP DYNAMICS
Large routing tables in the Internet are not only dangerous
per se, but also because they increase the likelihood that the
best routes frequently need to be recomputed. For this reason,
we now investigate the dynamics of routing information. Past
work has claimed the existence of highly unstable edge net-
works [6], [23]. Therefore, we aim at understanding whether
or not there are specific classes of ASs or prefixes that generate
a disproportionate number of BGP messages. To this end,
we start in Section VI-A by studying the relationship of the
business class of originating ASs and update dynamics. In a
similar way, we explore whether specificprefix classes (see
Section III) can be held responsible for a significant fraction of
BGP update churn (Section VI-B). Finally, in Section VI-C we
focus our analysis on subsets of prefixes that are responsible
for a high number of observed BGP updates and summarize
our findings in Section VI-D.
Observation points
Absolute number of updates
500000 1500000 2500000
500000 1500000 2500000
Small Transit Providers
Large Transit Providers
Enterprise Customers
Content/Access/Hosting Providers
Fig. 8. BGP updates by the four AS classes of [16].
A. BGP update rate by business type of originating AS
Based on the classification of ASs by business type such
as LTP, STP, CAHP, and EC (see Section III and [16]), we
compute the number of updates seen for all prefixes of each
class at a given observation point. The data we use are taken
from RIPE collectors [8]
rrc00
,
rrc01
and
rrc03
and contain
routing updates recorded at 180 observation points in more
than 142 ASs. We analyze BGP updates for the first week
in November of the years 2001 to 2008, but only present the
results for 2008, as the findings for other years are very similar.
Figure 8 is generated as follows: We sort the observation
points by the number of prefixes they receive along the xaxis.
For each observation point, we classify BGP prefix updates
into four categories (EC, STP, CAHP and LTP) according to
the business type of the AS which originates the BGP update.
The y-value represents the total number of updates collected
at an observation point.
Figure 8 reveals that a large fraction of the updates (for
most observation points, more than 60%) are due to prefixes
originated by enterprise networks (EC). This number is lower
for STP (around 30%). CAHP and LTP networks are respon-
sible for a small share, around 5% each. Unfortunately, not
all AS types are equally represented in publicly available data
sets. Most observation points are inside CAHP ASs, only 3
observation points are inside enterprise networks. However,
further analysis confirms that, irrespective of the location
of the observation point, most observed BGP updates are
originated by enterprise networks and small transit providers.
Since Internet domains that are either EC or CAHP belong
to the edge of the Internet, people may argue that the edge
of the Internet should be blamed for the scaling issues with
respect to BGP dynamics. However, we believe that such
statements over-simplify the root cause of the high BGP
update rates. After all, EC and CAHP networks together make
up 95% of the total number of ASs in today’s Internet and
this number is continuously increasing. Obviously, prefixes
originated by these two types of ASs contribute considerably
CITTADINI et al.: EVOLUTION OF INTERNET ADDRESS SPACE DEAGGREGATION: MYTHS AND REALITY 9
Observation points
Fraction of updates
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
lonely
top
delegated
de−aggregated
Fig. 9. Fraction of BGP updates by prefix type.
to BGP update churn due to their sheer number, but not
necessarily due to the fact that such prefixes are more likely
to experience changes in routing policies or are subject to
more aggressive traffic engineering practices. With respect to
the topological growth in terms of ASs at the edge, core-edge
separation approaches [24] for the future Internet, if they come
to pass, can be hoped to contain the increase in update rates.
B. BGP update rates by prefixclass
We now rely on the same BGP update trace as in Sec-
tion VI-A, but study the breakdown of the number of BGP
updates according to the four prefix classes introduced in
Section III: lonely,delegated,deaggregated and top. Our goal
is to understand to what extent specificprefix classes, e.g.,
delegated prefixes used by stub networks, are more frequently
involved in update churn than others.
Figure 9 illustrates the breakdown according to our four
prefix types in a manner similar to Figure 8: The observation
points are aligned along the xaxis, while the y-value represents
updates collected at an observation point. In order to directly
compare BGP updates with the data shown in Figure 3,
we normalize the number of updates for each type by the
total number of updates seen at the same observation point.
To avoid biases in the breakdown, we filtered out those
observation points which do not advertise the full BGP routing
table.
We find that slightly less than 50% of BGP updates occur
for lonely prefixes, while delegated and deaggregated get
roughly 20% and 27%, respectively. Overall, the fraction of
BGP updates for each prefix class is strongly related to the
fraction of prefixes in each class (see Figure 3). Thus, no
class is responsible for a disproportionate number of updates.
Contrary to our expectations, deaggregated prefixes do not
generate more than their share (see Figure 3) of BGP updates.
At a macroscopic level, deaggregated prefixes behave similarly
as the other types of prefixes.
However, we note that, if prefix aggregation were enforced,
for example by aggregating some of the delegated and deag-
gregated prefixes into their least specifics (top prefixes), BGP
updates could be reduced by up to 50%. Recall from Sec-
tion III that prefix aggregation would also shrink the routing
table by a similar factor. One way to perform aggregation
of deaggregated or delegated prefixes is to tag such routes
with BGP communities [20], [25] that tell the upstream ASs
whether the prefix can or should be aggregated into a less
specific entry of the BGP routing table. Given the widespread
usage of BGP communities, the impact of such a technique
can be large. However, its success depends also very much
on the usage of route tagging by edge networks which may
be unwilling to do so. And we must take into account that
aggregation is basically just not done in the commercial
Internet.
C. Prefixes with high update rates
We now turn to a more detailed study of those prefixes
for which we can observe a large number of updates within
a certain time period. Geoff Huston has repeatedly pointed
out that a few prefixes are responsible for a large fraction
of BGP updates [2]. To confirm this, we rely on the same
BGP update data set as in Section VI-A, and compute the
number of updates observed for each prefix. For each prefix,
we consider the median number of updates observed across
our observation points, in order to avoid biases due to outlying
observation points.
Similarly to the results in [2], we find that 98% of prefixes
undergo less than one hundred updates during the considered
one month period, while a tiny fraction of prefixes are respon-
sible for a huge number of updates.
In the previous section we observed that, when looking at
all prefixes, there is no specific class of ASs that announce a
disproportionate number of prefixes. An interesting question
is whether this property holds also for the subset of prefixes
that account for a large number of routing updates. Again, we
rely on the median number of observed BGP updates across
our observation points for each prefix. We rank the prefixes by
decreasing values of the median number of observed updates,
and select the subsets corresponding to the top x%ofthe
ranked prefixes, for x∈{0.1,0.25,0.5,1,2.5,5,10,25,50}.For
each subset, we break observed prefix updates down into
classes, according to the deaggregation type of the prefix(see
Section III-A).
In Figure 10 values at x=100 represent the breakdown
of BGP updates into deaggregation classes when considering
all prefixes. Values at x=1 represent the breakdown when
only considering the 1% of prefixes that generated the highest
number of updates.
According to Figure 10, even when restricting the analysis
to the prefixes with the highest churn, there are only negligible
differences among different deaggregation classes. Moreover,
the fraction of BGP updates each class is responsible for is
highly correlated with the portion of the routing table that
class accounts for (see Figure 3). In particular, we point out
10 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 28, NO. 8, OCTEMBER 2010
0.0 0.2 0.4 0.6 0.8 1.0
Subsets of the x% prefixes accounting for most updates
Fraction of updates
0.1 0.25 0.5 1 2.5 5 10 25 50 100
deaggregated
delegated
lonely
top
Fig. 10. Prefixes ranked by frequency of updates — Fraction of updates by
class. Each point on the x-axis corresponds to the subset of the x%prefixes for
which most updates are observed. The y-axis shows the fraction of observed
updates by class.
that deaggregated prefixes do not generate a disproportionate
number of BGP updates, even when we consider the small
subset of prefixes that are responsible for a huge amount of
routing updates.
To check that the way prefixes are classified does not
bias the results, we repeated the analysis of this section
with the classification introduced in Section IV-A. We do
not show these results due to space limitations, but found
no statistically significant difference in the behavior of the
different prefix classes. Only the “root/MS-complete” class
has lower worst-case behavior compared to other classes.
Otherwise, the distributions of the number of BGP updates
per prefix class are extremely similar.
D. Summary
Our analysis of BGP updates confirms the findings from
previous sections: we cannot claim that a specificclassof
prefixes misbehaves more than others. In particular, deaggre-
gated prefixes cannot be blamed for putting more stress than
their fair share on the routing system.
VII. RELATED WORK
A large number of research projects [26]–[29] which are
concerned with routing architectures for the future Internet,
motivate their work with alarming numbers of the growth of
routing table size or rates of routing updates. In general such
numbers can be obtained from various sources, including [2],
[4], [8], [9], [30]. Probably, the most prominent source of
information is the Potaroo web page [2], [4]. Based on results
from [2], Huston concluded that the Internet grows “at a rate
which has, at a minimum, doubled in size each year” [31].
In the past, many publications studied the evolution of
the Internet topology, mainly with a focus on the AS-level
structure, e.g., [16], [22], [32]. One general insight, relevant
for this work, is that the Internet is predominantly growing at
the “edge”, i.e. the number of enterprise, access, or content
networks is increasing faster than the core.
Moreover, efforts have been made to quantify the growth
of routing tables, to understand the underlying causes, and
to make predictions for the future. Bu et al. [3] explore the
degree to which factors such as multi-homing or address
fragmentation contribute to routing table size. While their
study does not include data on either allocations or routing
dynamics, others [33] explore how IPv4 address allocation
practices affect the BGP table growth. Meng et al. [1] take
the step of correlating information about allocated blocks with
what is actually announced in BGP routing tables. Compared
to our classification of Section III-A and Section IV-A, their
analysis is more coarse-grained, since [1] solely distinguishes
between covered and covering prefixes. Finally, work has
been done to explore BGP routing dynamics, e.g., [6], and
to identify prefixes that significantly contribute to routing
churn [23]. Despite the importance of interdomain traffic
engineering, to the best of our knowledge we are the first
to quantify the popularity of AS-path prepending and scoped
advertisements based on observed routing data.
According to recent findings [34] certain proportions be-
tween different measures are invariant in spite of the ongoing
growth of the Internet. For example, an estimation of the lower
bound on the number of active public IP addresses appears to
be consistent with a square law relationship with both the
number routing table entries and the number of ASs. The
findings of this paper are similar in that certain patterns in
the Internet, such as usage of allocated blocks, have not and
do not significantly change in the long term.
VIII. CONCLUSION
There is widespread belief in a high and fast growing
number of ASs that deaggregate prefixes, e.g., for the purpose
of traffic engineering [1]. Moreover, researchers often blame
specific classes of ASs for generating a disproportionate
amount of BGP updates.
Our primary objective is to challenge such widespread
assumptions (“myths”) and not just to confirm previous
findings [1]–[3]. Surprisingly, we find severe discrepancies
between existing myths and reality. According to our results,
there is no trend towards more aggressive prefix deaggrega-
tion or traffic engineering over time. With respect to update
dynamics, deaggregated prefixes generally do not generate a
disproportionate number of BGP updates with respect to their
share of the BGP routing table.
We stress that our work should not be used as an argument
to deny that there exists a problem with address space deag-
gregation. Instead, what our results reveal is that the global
impact of “bad guys” on Internet address space deaggregation
and BGP churn has not changed for the worse in recent years.
These problems have been present for a long time and have
not become more prevalent recently. This is a far more positive
picture of the behavior of IPv4 address space and its dynamics
compared to the alarming discourses typically heard [1], [2],
[4].
CITTADINI et al.: EVOLUTION OF INTERNET ADDRESS SPACE DEAGGREGATION: MYTHS AND REALITY 11
Our findings indicate that the growth of routing table and
BGP dynamics is due to the growing edge of the Internet. In
this light, core and edge separation, as advocated by clean-
slate approaches [7], could be an approach worth exploring to
improve routing dynamics.
ACKNOWLEDGMENT
The research results presented herein are partly funded
by a grant from Deutsche Telekom Laboratories, from Tril-
ogy (http://www.trilogy-project.eu), a research project (ICT-
216372) partially funded by the European Community under
its Seventh Framework Programme, as well as from G-lab
(http://www.german-lab.de), a research project (support code
01 BK 0805) funded by the Federal Ministry of Education
and Research of the Federal Republic of Germany. The views
expressed here are those of the author(s) only. The European
Commission is not liable for any use that may be made of the
information in this document.
REFERENCES
[1] X. Meng, Z. Xu, B. Zhang, G. Huston, S. Lu, and L. Zhang, “IPv4
Address Allocation and the BGP Routing Table Evolution,” ACM CCR,
vol. 35, no. 1, pp. 71–80, 2005.
[2] G. Huston, “BGP Routing Table Analysis Reports,”
http://bgp.potaroo.net/.
[3] T. Bu, L. Gao, and D. Towsley, “On Characterizing BGP Routing Table
Growth,” Computer Networks, vol. 45, no. 1, pp. 45–54, 2004.
[4] T. Bates, P. Smith, and G. Huston, “CIDR Report,” http://www.
cidr-report.org/as2.0/.
[5] C. Labovitz, G. Malan, and F. Jahanian, “Origins of Internet Routing
Instability,” in Proc. IEEE INFOCOM, 1999.
[6] J. Li, M. Guidero, Z. Wu, E. Purpus, and T. Ehrenkranz, “BGP Routing
Dynamics Revisited,” ACM CCR, vol. 37, no. 2, pp. 5–16, 2007.
[7] A. Feldmann, “Internet Clean-Slate Design: What and Why?,” ACM
CCR, vol. 37, no. 3, pp. 59–64, 2007.
[8] “RIPE Routing Information Service,” http://www.ripe.net/ris/.
[9] “University of Oregon Route Views Project,” http://www.routeviews.
org/.
[10] “Regional Internet Registries allocations statistics files,” ftp://ftp.ripe.
net/pub/stats.
[11] K. Hubbard, M. Kosters, D. Conrad, D. Karrenberg, and J. Postel,
“Internet Registry IP Allocation Guidelines,” IETF RFC2050, 2000.
[12] T. Hain, “A Pragmatic Report on IPv4 Address Space Consumption,”
The Internet Protocol Journal, vol. 8, no. 3, 2005.
[13] Y. Rekhter and T. Li, “An Architecture for IP Address Allocation with
CIDR,” IETF RFC1518, September 1993.
[14] V. Fuller, T. Li, J. Yu, and K. Varadhan, “Classless Inter-Domain
Routing (CIDR): an Address Assignment and Aggregation Strategy,”
IETF RFC1519, September 1993.
[15] P. Smith, J. Martin, and R. Bush, “Use of Final /8,” http://www.apnic.
net/policy/proposals/prop-062- v002.html.
[16] A. Dhamdhere and C. Dovrolis, “Ten Years in the Evolution of the
Internet Ecosystem,” in Proc. ACM IMC, 2008, pp. 183–196.
[17] L. Subramanian, S. Agarwal, J. Rexford, and R. Katz, “Characterizing
the Internet Hierarchy from Multiple Vantage Points,” in Proc. IEEE
INFOCOM, 2002.
[18] F. Wang and L. Gao, “Inferring and Characterizing Internet Routing
Policies,” in Proc. ACM IMC, 2003.
[19] G. Di Battista, M. Patrignani, and M. Pizzonia, “Computing the Types
of the Relationships Between Autonomous Systems,” in Proc. IEEE
INFOCOM, 2003.
[20] B. Donnet and O. Bonaventure, “On BGP Communities,” SIGCOMM
Comput. Commun. Rev., vol. 38, no. 2, 2008.
[21] B. Quoitin, S. Uhlig, C. Pelsser, L. Swinnen, and O.Bonaventure,
“Interdomain Traffic Engineering with BGP,” IEEE Commun. Mag.,
2003.
[22] R. Oliveira, D. Pei, W. Willinger, B. Zhang, and L. Zhang, “In Search
of the Elusive Ground Truth: The Internet’s AS-level Connectivity
Structure,” in Proc. ACM SIGMETRICS, 2008.
[23] B. Zhang R. Oliveira, R. Izhak-Ratzin and L. Zhang, “Measurement of
Highly Active Prefixes in BGP,” in Proc. IEEE GLOBECOM, 2005.
[24] D. Jen, M. Meisel, H. Yan, D. Massey, L. Wang, B. Zhang, and
L. Zhang, “Towards A New Internet Routing Architecture: Arguments
for Separating Edges from Transit Core,” in Proc. ACM Workshop on
Hot Topics in Networks, 2008.
[25] R. Chandra, P. Traina, and T. Li, “BGP Communities Attribute,” IETF
RFC1997, August 1996.
[26] EU Framework Programme 7 Integrated Project, “Trilogy,” http://www.
nets-find.net.
[27] NSF, “NeTS Initiative, Future Internet Design,” http://www.nets-find.
net.
[28] “Global Environment for Network Innovations,” http://www.geni.net/.
[29] “Internet Research Task Force – Routing Research Group,” http://tools.
ietf.org/group/irtf/trac/wiki/RoutingResearchGroup.
[30] “Cooperation for Internet Data Analysis (CAIDA),” http://caida.org/
research/topology.
[31] G. Huston, “Analyzing the Internet BGP Routing Table,” The Internet
Protocol Journal, vol. 4, no. 1, 2002.
[32] R. Oliviera, B. Zhang, and L. Zhang, “Observing the Evolution of
Internet AS Topology,” in Proc. ACM SIGCOMM, 2007.
[33] Z. Xu, X. Meng, L. Zhang, S. Lu, and C. Wittbordt, “Impact of IPv4
Address Space Allocation Practice on BGP Routing Table Growth,”
in Proc. IEEE 18th Annual Workshop on Computer Communications
(CCW), 2003.
[34] B. Carpenter, “Observed Relationships between Size Measures of the
Internet,” ACM CCR, vol. 39, no. 2, pp. 5–12, 2009.
Luca Cittadini received his master degree from
Roma Tre University in 2006, and a PhD degree in
Computer Science and Automation from the same
institution in 2010. During his PhD he was a teach-
ing assistant in the computer network research lab.
His research activity is primarily focused on inter-
domain routing, including theoretical analysis of the
protocol as well as active and passive measurements.
Wolfgang Mühlbauer is a senior researcher in
the Communication Systems Group at ETH Zürich,
Switzerland since November 2009. He received a
diploma degree from the Technische Universität
München, Germany, in 2005, and a PhD degree
in Computer Science (Informatik) from Technische
Universität Berlin, Germany, in 2009. Between 2006
and 2009 he was a research assistant and teach-
ing assistant in the network architectures groups,
first at Technische Universität München and then
at Technische Universität Berlin/Deutsche Telekom
Laboratories. His current research interests include inter-domain routing,
active and passive network measurements, and the design of network testbeds
and of distributed systems.
Steve Uhlig received the PhD degree in Applied
Sciences from the University of Louvain, Louvain-
la-neuve, Belgium, in 2004.
He is currently a Senior Research Scientist at
Deutsche Telekom Laboratories (T-labs) and Tech-
nische Universität Berlin, Germany. Prior to this,
he was at Delft University of Technology (2006-
2008) and was a Postdoctoral Fellow of the Belgian
National Fund for Scientific Research (2004-2006).
His main research interests are focused on the
large-scale behavior of the Internet and network
measurements. His current research topics include traffic characterization,
interdomain routing, traffic engineering, DNS usage, geolocation, and network
topology modeling.
Dr. Uhlig serves as an editor for Computer Communications journal. He
was the recipient of the annual IBM Belgium/F.N.R.S. Computer Science
Prize 2005.
12 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 28, NO. 8, OCTEMBER 2010
Randy Bush is a Senior Researcher and Network
Operator at Internet Initiative Japan, the first com-
mercial IPv6 deployment in the world. He special-
izes in IPv6 deployment, network security, proto-
cols, and network measurement especially routing.
Randy has been in computing for 45 years, and
has a few decades of Internet operations experience.
He was the engineering founder of Verio, which
is now NTT/Verio. He has been heavily involved
in transferring Internet technologies to developing
economies for over 20 years.
He was a chair of the IETF WG on the DNS for a decade and served as
a member of the IESG, as co-chair of the IETF Operations and Management
Area for six years. Randy was the first Chair of the NANOG Steering
Committee, a co-founder of AfNOG, on the founding Board of Directors
of ARIN, helped start AfriNIC, and has participated in APNIC, RIPE, et alia
since each was founded.
Pierre François is a Postdoctoral Fellow of
the ”Fonds national de la recherche scientifique”
(FNRS), Belgium. He obtained his Ph. D. from the
Université catholique de Louvain (UCL) in 2007,
Belgium. His Ph. D. thesis was supported by Cisco
Systems. He obtained his MS degree in computer
science from the University of Namur, Belgium, in
2003. He received the IEEE Infocom best paper
award in 2007. His research interests include intra-
and interdomain routing.
Olaf Maennel is a lecturer at Loughborough Univer-
sity in the United Kingdom since September 2009.
Before that he was at Deutsche Telekom Labora-
tories and Technische Universität Berlin, Germany.
From 2005 until 2008 he was a senior researcher
at the School of Mathematical Science of the Uni-
versity of Adelaide in South Australia, working
with Matthew Roughan on robustness of Internet
Routing. Dr. Maennel obtained his Ph.D. (Dr. rer.
nat.) in computer science in 2005 from the Technical
University of Munich, Germany, and his Diploma
(master) in 2002 from the Saarland University, Saarbruecken, Germany. His
research interests are routing, active measurements, next generation internet
technology, as well as configuration management. He is also serving in the
Internet Engineering Task Force (IETF).