Computer Science Technical Report Archive

USC Computer Science Technical Reports, no. 958 (2015)

(USC DC Other)

USC Computer Science Technical Reports, no. 958 (2015)

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content InvestigatingInterdomainRoutingPoliciesintheWild
Ruwaifa Anwar
Stony Brook University
Haseeb Niaz
Stony Brook University
David Choffnes
Northeastern University
Ítalo Cunha
Universidade Federal de
Minas Gerais
Phillipa Gill
Stony Brook University
Ethan Katz-Bassett
University of Southern
California
Abstract
Models of Internet routing are critical for studies of Internet
security, reliability and evolution, which often rely on sim-
ulations of the Internet’s routing system. Accurate models
are difficult to build and suffer from a dearth of ground truth
data, as ISPs often treat their connectivity and routing poli-
cies as trade secrets. In this environment, researchers rely on
a number of simplifying assumptions and models proposed
over a decade ago, which are widely criticized for their in-
ability to capture routing policies employed in practice.
In this study we put Internet topologies and models under
the microscope to understand where they fail to capture real
routing behavior. We measure data plane paths from thou-
sands of vantage points, located in eyeball networks around
the globe, and find that between 14-35% of routing decisions
are not explained by existing models. We then investigate
these cases, and identify root causes such as selective pre-
fix announcement, misclassification of undersea cables, and
geographic constraints. Our work highlights the need for
models that address such cases, and motivates the need for
further investigation of evolving Internet connectivity.
1. INTRODUCTION
Research on existing and new protocols on the Internet
is challenging because key aspects of the network topol-
ogy is hidden from public view by interdomain routing pro-
tocols, and deploying new protocols at Internet scale re-
quires convincing large numbers of autonomous networks
to participate. As a result, networking researchers rely
on assumptions, models, and simulations to evaluate new
protocols [12, 24], network reliability [19, 34], and secu-
rity [3, 15, 22].
Our existing models of interdomain routing [10], how-
ever, have important limitations. They are built and validated
on the same incomplete topology datasets, typically routes
observed via route monitors such as RouteViews and RIS.
These vantage points expose a large fraction of paths from
global research & education networks (GREN) and core net-
works, but they are incomplete in two keys ways. First, they
expose few paths to and from eyeball and content networks.
Second, they do not expose less preferred paths that would
be used if the most preferred next-hop AS were not available.
As a result, they do not capture partial peering, more com-
plex routing policies based on traffic engineering, or load
balancing and the rich peering mesh which exists near the
edge of the network [31].
While limitations of our existing models are well
known [25, 27, 31]–and are even being addressed in recent
work [14]–we lack a solid understanding of how much these
limitations impact our ability to accurately model the inter-
domain routing system. Recent work has attempted to ad-
dress this issue by observing destination-based routing vio-
lations in control plane data [26] and by surveying a popula-
tion of network operators about their policies [11], however,
these approaches are limited in terms of scale and their abil-
ity to observe behavior at the network edge.
In this paper, we take a systematic approach to under-
standing how our models of routing policies hold in prac-
tice. To accomplish this, we leverage a combination of data-
plane measurements covering the network edge (“eyeball
networks”) and control-plane experiments which allow us to
directly measure relative preference of routing options. We
create a methodology that takes into account numerous po-
tential causes of violations to our assumptions including sib-
ling ASes, complex AS relationships, prefix-specific routing
policies, and the impact of geography. We use this method-
ology to investigate the prevalence of each of these sources
of error in AS-level paths observed via measurements of the
data and control planes.
With these measurements, we revisit generally held as-
sumptions and models of Internet routing. Our goal isnot to
measure a complete Internet topology; rather, we seek to im-
prove our understanding of routing decisions made by ASes
when routing their traffic. Towards this goal we make the
following observations for our measured paths:
Hybrid and partial transit relationships (e.g., those ex-
plored in [14]) contribute a surprisingly small amount to
unexpected routing decisions.
Per-prefix routing policies explain 10-20% of unexpected
routing decisions, where an AS chooses a longer or more
expensive path than our model predicts.
We find that some large content providers like Akamai
and Netflix are destinations for a large fraction (21% and
17%, respectively) of unexpected routing decisions.
1
Routing decisions vary based on geography. We find
paths that traverse multiple continents deviate from our
models more, owing to undersea cable ASes which are
not accounted for in our models of AS relationships, and
a tendency for ASes to prefer non-international paths
when endpoints are in the same country.
Our results highlight areas where more investigation
would yield the largest payoff in terms of improving our ac-
curacy when modeling AS relationships and routing poli-
cies. We also identify key areas, specifically investigat-
ing prefix-specific routing policies, where additional vantage
points and looking glass servers could improve the fidelity of
our AS topology data.
2. MODELING INTERDOMAIN ROUTING
The now standard model of routing policies was devel-
oped by Gao and Rexford [9, 10] based on seminal work by
Griffin, Sheppard, and Wilfong [16] and Huston [17, 18]. In
this model, ASes connect to each other based on business re-
lationships: (1) customer-provider, where the customer pays
the provider and (2) peer-to-peer, where the ASes engage in
settlement-free peering and exchange traffic at no cost. This
model gives the following view of local preferences and ex-
port policies, based on the economic considerations of ASes:
Local Preferences. An AS will prefer routes through a
neighboring customer, then routes through a neighboring
peer, and then routes through a provider.
Export Policy. A customer route may be exported to all
neighboring ASes. A peer or provider route may only be
exported to customers.
This model is sometimes augmented with the assumption
that ASes only consider the next hop AS on the path when
making their routing decisions. This simplifies analysis and
makes debugging more tractable [19]. Simulation studies
also often restrict path selection to the shortest among all
paths satisfying Local Preference to induce unique routing
decisions [12, 13].
While the above model and variations of it have been used
in many studies (e.g., [3, 12, 15, 20, 34]), it is well known
that this model fails to capture many aspects of the interdo-
main routing system [25, 27, 31]. These aspects include AS
relationships that vary based on the geographic region [14]
or destination prefix, and traffic engineering via hot-potato
routing or load balancing.
Prior work has used traceroute measurements and BGP
data to address some of these issues (e.g., [25, 27]); how-
ever, these measurements only offer a glimpse into ASes’
routing preferences. Namely, they expose only the set of
paths that are in use at the time of measurements. Further,
these datasets have poor or no coverage of paths used by
edge networks serving residential users [7]. On a smaller
scale, network operators were surveyed about their routing
policies to better understand how our models correspond to
practice [11], but the scale and representativeness of a survey
approach makes generalizing these observations infeasible.
3. METHODOLOGY
We aim to understand the gap between interdomain rout-
ing models and empirically observed behavior on the In-
ternet. Our methodology combines two measurement tech-
niques to gain better visibility into interdomain routing poli-
cies. First, we measure paths between edge networks and
content providers to understand routing on paths that carry
the bulk of the Internet’s traffic [32]. Second, we perform
BGP announcements to explore less preferred paths and di-
rectly measure relative preference among next-hop ASes.
3.1 Data-plane measurements
It is well known that a disproportionately large amount
of Internet traffic originates from a few popular content
providers [21, 32] towards large populations of end users.
However, there is little empirical data about the paths this
traffic takes [21]. We target our data plane measurements
to cover these paths. Note that it is not our goal to explain
routing decisions for the entire Internet. Rather, we focus on
the more tractable task of measuring a subset of important
Internet paths (those carrying most traffic) from a diverse set
of vantage points, and putting those paths under the micro-
scope to understand how and why they differ from predicted
paths based on routing models.
Selecting content providers. We consider a list of the top
applications from Sandvine [32] and top Web sites from
Quantcast [29] and arrive at a list of 34 DNS names rep-
resenting 14 large content providers.
0
100
200
300
400
500
600
700
800
Stub AS Small ISP Large ISP Tier1
Number of Vantage Points
Figure 1: Distribution of the RIPE Atlas probes used in
this study.
Vantage points (VPs). We leverage the RIPE Atlas platform
which provides a large collection of probes located around
the world for our traceroute measurements. RIPE Atlas has
broad global coverage, but is known to have a dispropor-
tionate fraction of probes in Europe. To avoid a bias towards
European ASes, we determine how many probes we would
like to use and evenly divide the number of probes across
all continents. We then start choosing probes from coun-
tries within each continent in a round robin manner focusing
on distributing probes in different ASes within each coun-
try. We do this for each continent until we have allocated the
target number of probes. Figure 1 summarizes the location
of these probes in terms of AS type using the categorization
method of Oliveira et al. [28]. The bulk of the probes
are located near the network edge in stub and small ISP net-
works.
To measure paths to content providers, each RIPE Atlas
node performs a DNS lookup for each of the 34 content DNS
2
names, and then performs a traceroute to the resolved IP. In
our experiments we use 1,998 RIPE Atlas probes,
1
located
in 633 ASes, distributed according to our sampling method-
ology. Combined, these probes perform 28,051 traceroutes
to 218 destination ASes. The number of destination ASes is
large relative the number of content providers because large
numbers of content servers are hosted outside the provider’s
network (e.g., inside ISPs).
From traceroutes to routing decisions. We convert the
traceroute-based IP-level paths into AS paths using the
method described by Chen et al. [7]. Since interdomain rout-
ing is destination based, we can observe routing decisions
for all ASes along the path to a given destination. We thus
observe routing decisions for a total of 746 ASes.
3.2 Control-plane measurements
Data-plane measurements observe only the most preferred
route for an AS toward a destination. We use PEERING [33]
to expose alternate, less preferred routes and to attempt to
reverse engineer BGP decisions.
PEERING operates an ASN and owns IP address space that
we can announce via several upstream providers. PEERING
allows us to manipulate BGP announcements of its IP pre-
fixes and observe how ASes on the path react. We used
PEERING to announce prefixes using providers at six US uni-
versities (GaTech, Clemson, Southern California, Northeast-
ern, Stony Brook, and Cornell) and one university in Brazil.
We change announcements at most once per 90 minutes to
allow for route convergence and avoid route flap dampening.
We use traceroutes from PlanetLab and RIPE Atlas, as well
as BGP feeds from RouteViews and RIPE RIS, to monitor
paths toward PEERING prefixes.
Discovering alternate routes. We start announcing an IP
prefix from all PEERING locations in an ‘anycast’ announce-
ment. At each round, we observe the preferred route at a
target AST and the next-hop neighborN thatT is using to
route toward our prefix. We then poison N, i.e., add N’s
AS number to the path [4, 8], to trigger BGP loop preven-
tion atN and causeN to no longer have a path to our prefix
(and stop announcing a route toT ). This forcesT to choose
a different route, through a different neighbor N
0
. We re-
peat this process in consecutive rounds, poisoning the newly-
discovered neighbor, to identify all neighbors and routes T
can use toward our prefixes. When we observe different
routes at the target AST (through different neighbors) from
multiple vantage points (e.g., due to different routing pref-
erences at different geographic locations), we run the algo-
rithm for each vantage point separately. We can potentially
execute this algorithm for each AS in the topology as the tar-
get AS. A similar experiment was performed by Colitti [8];
here, we use the same mechanism with a more diverse set of
providers and with a different goal.
1
We targeted 2,000 probes but two did not return any data and had
to be discarded.
BGP poisoning does not work when BGP loop prevention
is disabled. It may also not work when ASes filter poisoned
announcements. These problems may prevent us from see-
ing all available neighbors and routes at the target AS. We
discuss these limitations in Sec. 4.4.
Reverse engineering BGP decisions. We first announce an
IP prefix from one PEERING location (called the magnet),
wait five minutes to allow for route convergence, then an-
nounce (anycast) the same IP prefix from all other PEERING
locations. After we anycast the prefix, An AS may change to
a new route with higher LocalPref, shorter AS-path length,
or better intradomain tie-breakers. If an AS keeps using the
route toward the magnet after we anycast the prefix, the AS
may be using route age as a tie-breaker (the last tie-breaker
before router ID).
If the AS did not choose the route to the magnet, we
check if the chosen route has a higher LocalPref or shorter
AS length. If none of these checks are satisfied, we con-
clude that the BGP decision was made at an intermediate
tie-breaker that considers the AS’s intradomain topology.
We repeat this process using each PEERING location as the
magnet. We also check whether the route chosen after we
anycast the prefix has a lower LocalPref or equal LocalPref
but longer AS-path length; which is a violation of the Gao-
Rexford model. The route to the magnet may become un-
available when a downstream AS receives and choses a more
preferred route; in these cases we consider the downstream
AS’s decision.
Data set. We performed a total of 188 distinct prefix an-
nouncements to infer preferences for all 360 target ASes
we observe on paths toward PEERING. We observe 739
inter-AS links, 45 (6.1%) of which are not in CAIDA’s AS-
relationship database.
3.3 Comparison with existing models
We compare paths observed in our our data- and control-
plane measurements with CAIDA’s topology of inferred
inter-AS relationships. We aggregate 5 topologies (Oct 14 to
Feb 15) inferred using the method presented by Luckieetal.
[23]. We aggregate these snapshots of the AS level topol-
ogy to mitigate the impact of transient link failures on the
topology used in our analysis. When inferences conflicted,
we took the majority poll of inferred relationships while as-
signing higher weight to more recent inferences. We use this
topology to compute all paths that satisfy the Gao-Rexford
(GR) local preference model described in Sec. 2.
We compare the measured paths with all paths satisfying
the GR model of local preference computed using CAIDA’s
inferred relationships. We consider two properties: (1)
whether the measured path satisfies the GR model of local
preference, and (2) whether the measured path has the same
length as the shortest paths satisfying the GR model of local
preference. Based on this we classify routing relationships
as either obeying GR local preference; i.e., using the neigh-
3
bor with the Best Relationship type (Best), routing based on
shortest path (Short), or a combination of the two.
For control-plane measurements to discover alternate
routes, we consider the order in which the target AS T
chooses paths. Again, we consider two properties: (1)
whether the relationship betweenT and the next-hop on the
first path is equal or better than the relationship with the
next-hop on the second path, and (2) whether the first path
is shorter or equal in length as the second path. We simi-
larly label the observed decisions which obey property (1)
as Best, and those that obey (2) as Short.
In both cases, the sets should be treated as disjoint, with
ASes that obey both Best and Short path policies appearing
only in the Best/Short category. Observations which follow
neither of these properties are considered inconsistent with
existing models (i.e., violations).
4. HOW OFTEN DO MODELS HOLD?
We now consider how empirically observed AS paths
compare with those predicted by models using AS relation-
ships inferred in [23]. We then investigate how often devia-
tions can be explained by known sources of inaccuracies.
Encouragingly, we find that a majority of routing deci-
sions (65%) are correctly inferred by the commonly used
Best/Short model; however, a significant fraction (35%) are
not. Figure 2 characterizes the observed routing decisions
based on whether the path chosen is Best or Short. We find
only a small number of cases (8%) where decisions can nei-
ther be explained by Best or Shortest path selection. In the
following sections, we explore the reasons behind these de-
cisions that differ from model-based predictions.
.
0
20
40
60
80
100
Simple Complex Sibs SPA-1 SPA-2 All-1 All-2
Fraction of Decisions
Best/Short
NonBest/Short
Best/Long
NonBest/Long
Figure 2: Breakdown of routing decisions observed tak-
ing into account the complex relationships.
4.1 Complex routing relationships
A well known limitation of existing routing policy mod-
els is the simplification of relationships into either customer-
provider or settlement-free peering relationships. Recent
work by Giotsasetal. aims to address this limitation by aug-
menting existing relationship inferences with cases of hybrid
relationships (i.e., ASes whose arrangements vary based on
location) and partial transit relationships (i.e., ASes who will
behave as providers, but only for a subset of prefixes) [14].
Figure 2 (Complex) shows the breakdown of routing deci-
sions observed taking into account these complex relation-
ships. Interestingly, we find that taking these relationships
has nearly no impact on the classification in our dataset.
4.2 Sibling ASes
The mapping between AS numbers and organizations is
not one-to-one [5]. Many organizations manage multiple AS
numbers, either for geographic regions (e.g., Verizon with
ASNs 701, 702, and 703) or due to mergers (e.g., Level 3
(AS 3356) and Global Crossing (AS 3549)).
We take a similar approach to Cai et al. [5] to identify AS
siblings, but our approach differs in two key ways. First, we
focus only on e-mail addresses in whois data, which pre-
vious work identified as the field with best precision and
recall [5]. Second, we use DNS SOA records to identify
different e-mail domains that belong to the same organiza-
tion. For example, dish.com and dishaccess.tv share the
dishnetwork.com authoritative domain. We also remove
groups where the e-mail address is hosted by a popular e-
mail provider (e.g., hotmail.com), or regional Internet reg-
istry (e.g., ripe.net). This results in a total of 94 sibling
groups identified in our traceroute data set.
For every non GR decision that an AS makes, we check
whether the AS chose a path via a sibling. If the path is a
via a sibling, we mark this decision as satisfying the Best
condition. Figure 2 (Sibs) shows the result of this change—
3% more decisions are classified as Best/Short.
4.3 Prefix-specific policies
Interdomain routing is often abstracted to the level of a
destination AS. However, in practice routing is based done
on IP prefixes which may be subject to different export poli-
cies by their originating AS (e.g., forwarding prefixes host-
ing enterprise-class services to a more expensive provider).
While Giotsas et al. consider partial transit [14], which is a
type of prefix-specific policy, they do not explicitly consider
per-prefix policies as implemented by origin ASes.
We use two criteria to identify origin-based prefix specific
policies based on correlation with BGP data obtained from
Routeviews/RIPE [2, 30]. Given an origin AS (O), neighbor
N and prefixP : Criteria 1 do not assume the edgeNO
exists for prefixP unless we observeO announcingP toN
in the BGP data. Criteria 2 is similar to Criteria 1, except
that we require that we observe at least one prefix announced
fromO toN before applying Criteria 1. The first criteria can
be seen as being more aggressive whereas the second aims
to ensure that our observation is actually caused by selective
prefix announcement and not poor visibility.
Figure 2 (SPA-1, SPA-2) shows the breakdown of routing
decisions using Criteria 1 and 2 above, respectively. We find
that prefix-specific policies account for a significant fraction
(10-19%) of unexpected routing decisions.
Validation. In order to validate cases of prefix-specific poli-
cies, we try to find a Looking Glass server hosted by the
neighboring AS. There were a total of 630 cases of prefix-
specific policies involving 149 unique neighboring ASes.
4
We were able to find looking glass servers in 28 of the neigh-
boring ASes. Using these looking glass servers we manually
verify 100 cases of prefix-specific policies and confirm that
applying Criteria 1 was correct 78% of the time.
4.4 Active BGP Measurements
Our control-plane experiments allow us to check how of-
ten our models match real routing choices and how many
routing decisions they capture.
Alternate routes. Here analyze AS routing choices when
we use PEERING to discover alternate, less preferred routes.
We compare the sequence of routes chosen by target ASes
with CAIDA’s AS-relationships database. Out of the 360
ASes we targeted, 310 (86.1%) chose routes following both
Best and Shortest (as defined in Sec. 3.3); 29 (8.0%) chose
routes following Best only; 18 (5.0%) following Shortest
only; and 3 (0.8%) did not follow either properties. We now
discuss the 3 observations that did not satisfy either property
to illustrate limitations of current models.
One violation occurs for a European network E that
chooses to route via OpenPeering (AS20562)–a partial-
transit relationship validated using whois. After poisoning
OpenPeering,E chooses a route through another settlement-
free peer-to-peer relationship with AMPATH (AS20080) at
AMS-IX. We list this as a violation because CAIDA iden-
tifies OpenPeering as a provider for E. Interestingly, the
second route is the suffix of the first route (i.e., the route
through OpenPeering also reaches PEERING through AM-
PATH at AMS-IX), indicating the first route includes an un-
necessary detour. Peering relationships are not only com-
plex, one settlement-free (or paid) peering relationship may
be preferred over another. Models with finer granularity for
ranking neighbors of an AS may resolve these issues.
Another violation occurs at a US university U. The uni-
versity first chooses a route through Internet2 (AS11537)
toward one of the PEERING locations in the US. After
we poison Internet2, U chooses a route through AMPATH
(AS20080) toward the PEERING location in Brazil. We
list this as a violation because CAIDA identifies Internet2
as a provider and AMPATH as a settlement-free peer of
U. Our last observed violation is similar, where a Euro-
pean network first chooses a route through Switch (AS559,
identified as a provider) and then chooses a route through
NCSA (AS10764, identified as a settlement-free peer) to
reach PEERING after we poison Switch. These violations in-
dicate that identifying links used as back-up might improve
our routing models.
Reverse engineering BGP decisions. We now turn to our
second control-plane experiment, where we use anycast to
explore considerations such as older path on routing deci-
sions. Table 1 shows the root cause behind BGP routing
decisions. Although most decisions are made based on re-
lationship and path length, more than 15% of decisions are
made based on intradomain factors and route age, which are
not considered in current models.
CAUSE BGP FEEDS TRACEROUTES
Higher LocalPref 435 (46.0%) 228 (42.4%)
Shorter path 152 (16.0%) 158 (29.4%)
Intradomain 155 (16.4%) 84 (15.6%)
Route age (magnet) 24 (2.5%) 9 (1.6%)
GR violations 179 (18.9%) 58 (10.8%)
Total 945 (100%) 537 (100%)
Table 1: Reverse engineering of the BGP decision pro-
cess.
Limitations. We note our results for control-plane experi-
ments cover a small fraction of the Internet and are probably
biased toward academic and research networks. Our control-
plane techniques, however, are general and could be used by
other networks to cover different portions of the Internet. We
believe better coverage and visibility would result in discov-
ering more violations. To this end, we are working to extend
the PEERING platform as well as talking with RIPE about
using RIPE Atlas to monitor routes to PEERING prefixes.
5. SOURCES OF VIOLATIONS
In this section we investigate which source and destina-
tion ASes account for most of the routing decisions which
deviate from our model. Figure 3 (a) and (b) shows the cu-
mulative fraction of routing decisions which violate either
the Best or Short condition (i.e., the AS chooses a path that
is longer or more expensive than we would expect). If viola-
tions were evenly distributed across ASes, the curves would
fix y = x; otherwise, some ASes are responsible for a dis-
proportionately larger (or smaller) fraction of violations. We
find this effect is present in both plots, but more prominently
for destination ASes. We focus on the latter.
Destination ASes owned by Akamai account for 21% of
violations. Of these, Cogent (AS174) is the most common
source, responsible for 3.4% of violations. These Cogent-
Akamai violations tend to occur when Cogent prefers a peer-
to-peer path through a Tier-1 AS over a longer customer
route. Netflix’s AS is the destination on 17% or paths with
violations. Of these, nearly a quarter (24%) are due to a stale
inter-AS link in CAIDA’s topology, which included a direct
link between AS3549 and Netflix that no longer exists. For
source ASes, the distribution is less skewed. Cogent and
Time Warner are the top two sources, responsible for 4.1%
and 2.2% of violations, respectively.
6. IMPACT OF GEOGRAPHY
We next consider the role of geography on routing deci-
sions. First, we isolated traceroutes that stay within a con-
tinent (Continental traceroutes), i.e., all hops stay inside a
given continent based on geolocating router IP addresses.
We use the geolocation data from [6], which offers good cov-
erage of infrastructure IPs such as routers. Figure 4 shows
the breakdown of decisions in the continental traceroutes
(45% of those in our dataset). The fraction of decisions
5
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
Cumulative fraction of decisions
Source ASes ranked by fraction of decisions
Best+NonShort
NonBest+Short
NonBest+NonShort
(a) Distribution of violations across source ASes.
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
Cumulative fraction of decisions
Destination ASes ranked by fraction of decisions
Best+NonShort
NonBest+Short
NonBest+NonShort
(b) Distributions of violations across destination ASes.
Figure 3: CDF plot of the fraction of violations (x-axis)
explained by source and destinations ASes (y-axis). Vi-
olations observed in our dataset are skewed significantly
toward Akamai and Netflix. The skew for source ASes is
less prominent.
explained by GR for continental traceroutes is significantly
greater than for transcontinental ones.
0
20
40
60
80
100
AF NA EU SA AS Cont Non Cont
Fraction of Decisions
Best/Short
NonBest/Short
Best/Long
NonBest/Long
Figure 4: Breakdown of traceroutes that stay within a
continent
Domestic paths. Next we focused on traceroutes where we
infer that the entire traceroute stayed within a single coun-
try, but there is a bettermultinationalBest/Shortpath (in the
CAIDA data), which we define to be a path with at least
one AS registered (via whois data) in a country outside the
source and destination AS’s country. We find that more than
40% of non-Best/Short decisions can be explained by avoid-
ing alternative multinational paths. Table 2 details the non-
Best/Short decisions explained by ASes preferring domestic
routes.
Undersea cables. Undersea cable ASes are a criti-
cal component of Internet topologies that previous work
overlooks. While some cables are jointly owned by large
ISPs, e.g., Pan-American Crossing, Americas-II (owned by
AT&T, Sprint, and many others), we observed that others,
e.g., EAC- C2C (PACNET), are operated by independent
Continent Non-Best/Short Decisions explained
Asia 40.1%
Africa 62.5%
N. America 10.9%
Oceana 62.9%
S. America 66.6%
Table 2: Summary of Non-Best/Short decisions ex-
plained by ASes preferring intra-country routes.
Violation type Pct. of decisions explained
Non-Best & Short 3%
Best & Long 6.5%
Non-Best & Long 4.5%
Table 3: Fraction of decisions of each type that can be
attributed to undersea cables.
organizations using their own allocated ASNs and IP pre-
fixes. Because these cable operators only provide point-to-
point transit along the cables (i.e., they do not originate traf-
fic and peer in locations proportional to cable landings), they
resemble high-latency, high-cost IXPs and thus confuse ex-
isting AS relationship models. As such, we need techniques
to identify cable ASes and correct their relationships in in-
ferred topologies.
We use a list of undersea cables from the TeleGeography
Submarine Cable Map [1] to identify ASes for undersea ca-
ble operators. Overall, cable-ASes appear on less than 2% of
paths but most of the decisions (51%) involving cable-ASes
caused deviations from Best/Short paths. Table 3 shows
fraction of each type of decision explained by undersea cable
ASes.
7. CONCLUSION
In this work, we investigated how interdomain paths pre-
dicted by state-of-the-art routing models differ from empir-
ically observed routes. We found that while a majority of
paths in our dataset agree with models, more than a third
do not. We explained a significant fraction of these differ-
ences due to factors such as sibling ASes, selective prefix
announcements and undersea cables. Further, we investi-
gated how the models hold up when comparing with ground-
truth routing preferences revealed using PEERING announce-
ments, and identified AS behavior that is not included in
existing models. As part of future work, we are continu-
ing to investigate cases of routing decisions that violate to-
day’s models, and we aim to incorporate our findings into
new models of Internet routing.
Acknowledgments
Research in this report was funded by NSF grant CNS-
1423659 and a Comcast TechFund award.
8. REFERENCES
[1] TeleGeography Submarine Cable Map.
http://www.submarinecablemap.com/.
[2] University of Oregon Route Views Project.
http://www.routeviews.org/.
[3] H. Ballani, P. Francis, and X. Zhang. A study of prefix hijacking and
interception in the Internet. InSIGCOMM, 2007.
6
[4] R. Bush, O. Maennel, M. Roughan, and S. Uhlig. Internet optometry:
Assessing broken glasses in internet reachability. InACMIMC, 2009.
[5] X. Cai, J. Heidemann, B. Krishnamurthy, and W. Willinger. Towards
an AS-to-organization map. InUSENIXATC, 2010.
[6] B. Chandrasekaran, M. Bai, M. Schoenfield, A. Berger, N. Caruso,
G. Economou, S. Gilliss, B. Maggs, K. Moses, D. Duff, K. Ngâ
˘
A˘ a,
E. G. Sirer, R. Weberâ
˘
A˘ a, and B. Wong. Alidade: Ip geolocation
without active probing.DepartmentofComputerScience,Duke
University,TechnicalReport, 2015.
[7] K. Chen, D. Choffnes, R. Potharaju, Y . Chen, F. Bustamante, D. Pei,
and Y . Zhao. Where the sidewalk ends: Extending the internet AS
graph using traceroutes from P2P users. InCoNEXT’09, 2009.
[8] L. Colitti.InternetTopologyDiscoveryUsingActiveProbing. Ph.D.
thesis, University di Roma Tre, 2006.
[9] L. Gao, T. Griffin, and J. Rexford. Inherently safe backup routing
with BGP.IEEEINFOCOM, 2001.
[10] L. Gao and J. Rexford. Stable Internet routing without global
coordination.Trans.Netw., 2001.
[11] P. Gill, S. Goldberg, and M. Schapira. A survey of interdomain
routing policies.ACMCCR, 2014.
[12] P. Gill, M. Schapira, and S. Goldberg. Let the market drive
deployment: A strategy for transistioning to BGP security.
SIGCOMM’11, 2011.
[13] P. Gill, M. Schapira, and S. Goldberg. Modeling on quicksand:
dealing with the scarcity of ground truth in interdomain routing data.
SIGCOMMComput.Commun.Rev., 42(1):40–46, Jan. 2012.
[14] V . Giotsas, M. Luckie, B. Huffier, and K. Claffy. Inferring Complex
AS Relationships. InACMIMC, November 2014.
[15] S. Goldberg, M. Schapira, P. Hummon, and J. Rexford. How secure
are secure interdomain routing protocols? InSIGCOMM’10, 2010.
[16] T. Griffin, F. B. Shepherd, and G. Wilfong. The stable paths problem
and interdomain routing.Trans.Netw., 2002.
[17] G. Huston. Peering and settlements - Part I.TheInternetProtocol
Journal(Cisco), 2(1), March 1999.
[18] G. Huston. Peering and settlements - Part II.TheInternetProtocol
Journal(Cisco), 2(2), June 1999.
[19] U. Javed, I. Cunha, D. R. Choffnes, E. Katz-Bassett, T. Anderson,
and A. Krishnamurthy. Poiroot: Investigating the root cause of
interdomain path changes. InSIGCOMM, 2013.
[20] J. Karlin, S. Forrest, and J. Rexford. Nation-state routing:
Censorship, wiretapping, and BGP.CoRR, 2009.
[21] C. Labovitz, S. Iekel-Johnson, D. McPherson, J. Oberheide, and
F. Jahanian. Internet inter-domain traffic. InSIGCOMM’10, 2010.
[22] M. Lad, D. Massey, D. Pei, Y . Wu, B. Zhang, and L. Zhang. PHAS:
A prefix hijack alert system. InProc.USENIXSecuritySymposium,
2006.
[23] M. Luckie, B. Huffaker, A. Dhamdhere, and V . Giotsas. AS
relationships, customer cones, and validation. InACMInternet
MeasurementConference, 2013.
[24] R. Lychev, S. Goldberg, and M. Schapira. Is the juice worth the
squeeze? BGP security in partial deployment. InSIGCOMM’13,
2013.
[25] H. Madhyastha, E. Katz-Bassett, T. Anderson, A. Krishnamurthy,
and A. Venkataramani. iPlane Nano: Path prediction for peer-to-peer
applications. InUsenixNSDI, 2009.
[26] R. Mazloum, M. Buob, J. Auge, B. Baynat, D. Rossi, and
T. Friedman. Violation of Interdomain Routing Assumptions. In
PassiveandActiveMeasurementConference, March 2014.
[27] W. Mühlbauer, A. Feldmann, O. Maennel, M. Roughan, and
S. Uhlig. Building an AS-topology model that captures route
diversity. InSIGCOMM, 2006.
[28] R. Oliveira, D. Pei, W. Willinger, B. Zhang, and L. Zhang.
Quantifying the completeness of the observed internet AS-level
structure.UCLAComputerScienceDepartment-TechicalReport
TR-080026-2008, Sept 2008.
[29] Quantcast.http://www.quantcast.com.
[30] RIPE Network Coordination Center. RIPE Routing Information
Service.
http://www.ripe.net/data-tools/stats/ris/routing-information-service.
[31] M. Roughan, W. Willinger, O. Maennel, D. Perouli, and R. Bush. 10
lessons from 10 years of measuring and modeling the Internet’s
autonomous systems.JSAC, 2011.
[32] Sandvine. Fall 2012 global internet phenomena, 2012.
[33] B. Schlinker, K. Zarifis, I. Cunha, N. Feamster, and E. Katz-Bassett.
PEERING: An AS for Us. InProc.ACMHotNets, Los Angeles, CA,
October 2014.
[34] J. Wu, Y . Zhang, Z. M. Mao, and K. Shin. Internet routing resilience
to failures: Analysis and implications. InCoNEXT, 2007.
7

Linked assets

Computer Science Technical Report Archive

Conceptually similar

PDF

USC Computer Science Technical Reports, no. 934 (2013)

PDF

USC Computer Science Technical Reports, no. 957 (2015)

PDF

USC Computer Science Technical Reports, no. 961 (2015)

PDF

USC Computer Science Technical Reports, no. 971 (2017)

PDF

USC Computer Science Technical Reports, no. 655 (1997)

PDF

USC Computer Science Technical Reports, no. 935 (2013)

PDF

USC Computer Science Technical Reports, no. 944 (2014)

PDF

USC Computer Science Technical Reports, no. 840 (2005)

PDF

USC Computer Science Technical Reports, no. 603 (1995)

PDF

USC Computer Science Technical Reports, no. 673 (1998)

PDF

USC Computer Science Technical Reports, no. 674 (1998)

PDF

USC Computer Science Technical Reports, no. 900 (2008)

PDF

USC Computer Science Technical Reports, no. 631 (1996)

PDF

USC Computer Science Technical Reports, no. 963 (2015)

PDF

USC Computer Science Technical Reports, no. 727 (2000)

PDF

USC Computer Science Technical Reports, no. 972 (2017)

PDF

USC Computer Science Technical Reports, no. 657 (1997)

PDF

USC Computer Science Technical Reports, no. 765 (2002)

PDF

USC Computer Science Technical Reports, no. 789 (2003)

PDF

USC Computer Science Technical Reports, no. 959 (2015)

Description

Ruwaifa Anwar, Haseeb Niaz, David Choffnes, Italo Cunha, Phillipa Gill, and Ethan Katz-Bassett. "Investigating interdomain routing policies in the wild." Computer Science Technical Reports (Los Angeles, California, USA: University of Southern California. Department of Computer Science) no. 958 (2015).

Asset Metadata

Creator Anwar, Ruwaifa (author), Choffnes, David (author), Cunha, Italo (author), Gill, Phillipa (author), Katz-Bassett, Ethan (author), Niaz, Haseeb (author)

Core Title USC Computer Science Technical Reports, no. 958 (2015)

Alternative Title Investigating interdomain routing policies in the wild (title)

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)

Tag OAI-PMH Harvest

Format 7 pages (extent), technical reports (aat)

Language English

Unique identifier UC16270907

Identifier 15-958 Investigating Interdomain Routing Policies in the Wild (filename)

Legacy Identifier usc-cstr-15-958

Format 7 pages (extent),technical reports (aat)

Rights Department of Computer Science (University of Southern California) and the author(s).

Internet Media Type application/pdf

Source 20180426-rozan-cstechreports-shoaf (batch), Computer Science Technical Report Archive (collection), University of Southern California. Department of Computer Science. Technical Reports (series)

Access Conditions The author(s) retain rights to their work according to U.S. copyright law. Electronic access is being provided by the USC Libraries, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Repository Email csdept@usc.edu

Inherited Values

Title Computer Science Technical Report Archive

Description Archive of computer science technical reports published by the USC Department of Computer Science from 1991 - 2017.

Coverage Temporal 1991/2017

Repository Email csdept@usc.edu

Repository Name USC Viterbi School of Engineering Department of Computer Science

Repository Location Department of Computer Science. USC Viterbi School of Engineering. Los Angeles\, CA\, 90089

Publisher Department of Computer Science,USC Viterbi School of Engineering, University of Southern California, 3650 McClintock Avenue, Los Angeles, California, 90089, USA (publisher)