NANO: Network Access Neutrality Observatory
Mukarram Bin Tariq, Murtaza Motiwala, Nick Feamster
{mtariq,murtaza,feamster}@cc.gatech.edu
ABSTRACT
We present NANO, a system that establishes whether per-
formance degradations that services or clients experience
are caused by an ISP’s discriminatory policies. To distin-
guish discrimination from other causes of degradation (e.g.,
overload, misconfiguration, failure), NANO uses a statisti-
cal method to estimate causal effect. NANO aggregates pas-
sive measurements from end-hosts, stratifies the measure-
ments to account for possible confounding factors, and dis-
tinguishes when an ISP is discriminating against a particular
service or group of clients. Using simulationwe demonstrate
the promise of NANO for both detecting discrimination and
absolving an ISP when it is not discriminating.
1. Introduction
In late 2005, Ed Whitacre sparked an intense debate on
network neutrality when he decried content providers “us-
ing his pipes [for] free”. Network neutrality says that end
users must be in “control of content and applications that
they use on the Internet” [1], and that the ISPs respect that
right by remaining neutral in treating the traffic, irrespec-
tive of its content or application. This paper does not take a
stance in this debate, but instead studies a technical question:
Can users in access networks detect and quantify discrimi-
natory practices of an ISP against a particular group of users
or services? We define any practice by the ISP that degrades
performance or connectivity for a service as discrimination
or violation of network neutrality. We refer to such viola-
tions as discrimination. We argue that, regardless of whether
or not discrimination is ultimately deemed to be acceptable,
the network should be transparent; that is, users should be
able to ascertain the behavior of their access ISPs.
Unfortunately, because ISP discrimination can take many
forms, detecting it is difficult. Several ISPs have been in-
terfering with TCP connections for BitTorrent and other
peer-to-peer applications [3]. Other types of discrimination
may include blocking specific ports, throttling bandwidth, or
shaping traffic for specific services, or enforcing traffic quo-
tas. Existing detection mechanisms actively probe ISPs to
test for specific cases of discrimination: Glasnost [3], auto-
mates detection of spurious TCP reset packets, Beverly et
al. [8] present a study of port-blocking, and NVLens [10]
detects the use of packet-forwarding prioritization by ISPs
by examining the TOS bits in the ICMP time exceeded mes-
sages. The main drawback of these mechanisms is that each
is specific to one type of discrimination; thus, each form of
discrimination requires a new test. Worse yet, an ISP may
either block or prioritize active probes associated with these
tests, making it difficult to run them at all.
If end users could instead somehow detect discrimination
by observing the effect on service performance using pas-
sive, in-band methods, the detection mechanism would be
much more robust. Unlike active measurements, ISPs can-
not prioritize, block, or otherwise modify in-band measure-
ments. To achieve this robustness, we use a black-box ap-
proach: we make no assumptions about the mechanisms for
implementing discrimination and instead use statistical anal-
ysis primarily based on in situ service performance data to
quantify the causal relationship between an ISP’s policy and
the observed service degradation.
In this paper, we present the design for Network Access
Neutrality Observatory (NANO), a system that infers the ex-
tent to which an ISP’s policy causes performance degrada-
tions for a particular service. NANO relies on participat-
ing end-system clients that collect and report service per-
formance measurements for a service. Establishing such a
causal relationship is challenging because many confound-
ing factors (or variables) that are unrelated to ISP discrimi-
nation can also affect the performance of a particular service
or application. For example, a service may be slow (e.g.,
due to overload at a particular time-of-the-day). A service
might be poorly located relative to the customers of the ISP.
Similarly, a service may be fundamentally unsuitable for a
particular network (e.g., Internet connectivity is not suitable
for VoIP applications in many parts of the world).
A necessary condition for inferring a valid causal relation-
ship is to show that when all the other factors are equal, a
service performs poorly when accessed from an ISP com-
pared to another ISP. Themain challenge in designingNANO
is to create an environment where all other factors are in fact
equal. Creating such an environment requires (1) enumer-
ating the confounding factors; (2) establishing a “baseline”
level of performance where all factors besides the confound-
ing variables are equal. Unfortunately, the nature of many
confounding factors makes it difficult to create an environ-
ment on the real Internet where all other factors, except for
an ISP’s discriminative policy and service, would be equal.
Instead, to correctly infer the causal relationship, we must
adjust for the confounding factor by creating strata of clients
that have “similar” values for all factors except for their ac-
cess network. Our approach is based on the theory of causal
inference, which is applied extensively in other fields, in-
cluding epidemiology, economics, and sociology.
This paper is organized as follows. In Section 2 we
overview necessary background for establishing a causal
relationship between ISP policy and service performance
degradation and formalize the problem. In Section 3, we
describe the steps in the causal inference, the confound-
ing variables for the problem, and the NANO architecture
for collecting and processing the necessary data. Section 4
presents simulation-based results, and Section 5 concludes
with a discussion of open issues and a research agenda.

2. Background and Problem Formulation
In this section, we formalize the definitions and basic con-
cepts used for establishing ISP discrimination as the cause
of service degradation. We describe the concept of service,
service performance, ISP, and discrimination; the inference
of causal effect, how it relates to association and correla-
tion; and finally, approaches for quantifying causal effect.
We also formalize the application of causality to detecting
ISP discrimination.
2.1 Definitions
Service and Performance. A service is the “atomic unit”
of discrimination. An ISP may discriminate against traffic
for a particular service, e.g., Web search, traffic for a par-
ticular domain, or particular type of media, such as video.
Such traffic may be identifiable using the URL or the pro-
tocol. Similarly, ISPs may target specific applications, e.g.,
VoIP, or peer-to-peer file transfers. Performance, the out-
come variable, is specific to the service. For example, we
use server response time for HTTP requests, loss, and jit-
ter for VoIP traffic, and average throughput for peer-to-peer
traffic.
ISP and Discrimination. Discrimination against a service
is a function of ISP policy. The performance for a service
depends on both the properties of the ISP’s network, e.g.,
its location, as well as the policy of treating the traffic dif-
ferently. Thus, an objective evaluation of ISP discrimination
must adjust for the ISP’s network as a confounding factor. To
differentiate an ISP’s network from its discrimination policy,
we use the ISP brand or name as the causal variable referring
to the ISP’s discrimination policy. In the rest of the paper,
when we use ISP as the cause, we are referring to the ISP
policy or the brand with which the policy is associated.
We aim to detect whether a certain practice of an ISP re-
sults in poorer performance for a service compared to other
similar services or performance for the same service through
other ISPs. If an ISP’s policy of treating traffic differently
does not result in degradation of performance, we do not
consider it as discrimination.
2.2 Background for Causal Inference
Statistical methods offer tools for causal inference that
have been used in observational and experimental studies [6,
7]. NANO draws heavily on these techniques. In this sec-
tion, we review basic concepts and approaches for causal in-
ference, and how they relate to inferring ISP discrimination.
Causal Effect. The statement “X causes Y” means that if
there is a change in the value of variableX , then we expect a
change in value of variableY . We refer toX as the treatment
variable and Y as the outcome variable.
In the context of this paper, accessing a particular service
through an ISP is our treatment variable (X), and the ob-
served performance of a service (Y ) is our outcome variable.
Thus, treatment is a binary variable; X ∈ {0, 1}, X = 1
when we access the service through the ISP, and X = 0
when we do not (e.g., access the service through an alterna-
tive ISP). The value of outcome variable Y depends on the
performance metric and the service for which we are mea-
suring the performance.
The goal of causal inference is to estimate the effect of
the treatment variable (the ISP) on the outcome variable (the
service performance). Let’s define ground-truth value for
the outcome random variable as GX , so that G1 is the out-
come value for a client when X = 1, and G0 is the outcome
value when X = 0. We will refer to the outcome when not
using the ISP (X = 0) as the baseline—we can define base-
line in a number of ways, as we describe in more detail in
Section 3.1.2.
We can quantify the average causal effect of using an ISP
as the expected difference in the ground truth of service per-
formance between using the ISP and the baseline.
θ = E(G1) − E(G0) (1)
Note that to compute the causal effect, θ, we must observe
values of the outcome both under the treatment and without
the treatment.
Association vs. Causal Effect. In a typical in situ dataset,
each sample presents only the value of the outcome variable
either under the treatment, or under the lack of the treatment,
but not both; e.g., a dataset about users accessing a particular
service through one of the two possible ISPs, ISPa and ISPb,
will comprise data of the form where, for each client, we
have performance data for either ISPa or ISPb, but not both.
Such a dataset may thus be incomplete and therefore not suf-
ficient to compute the causal effect, as shown in Equation 1.
Instead, we can use such a dataset to compute correlation
or association. Let’s define association as simply the mea-
sure of observed effect on the outcome variable:
α = E(Y |X = 1) − E(Y |X = 0) (2)
It is well known that association is not a sufficient metric for
causal effect, and in general α 6= θ.
Example. Tables 1(a) and (b) illustrate the difference be-
tween association and causal effect using an example of
eight clients (a–h). The treatment variable X is binary; 1
if a user uses a particular ISP, and 0 otherwise. For simplic-
ity, the outcome (Y ) is also binary, 1 indicating that a client
observes good performance and 0 otherwise; both α and θ
are in the range [−1, 1].
Table 1(a) shows an in situ dataset. In this dataset, clients
a–d do not use the ISP in question and clients e–h use the
ISP. Note that for each sample, only one or the other out-
come is observable. The association value in this dataset is
α = −3/4. If we use the association value as an indicator
of causal effect, we would infer that using the ISP causes a
significant negative impact on the performance.
Table 1(b), on the other hand, presents the ground-truth
performance values, G0 and G1, as the performance when
not using the ISP and performance when using the ISP for
the same client, respectively. These values could be obtained
by either subjecting the client to the two cases, or through an
oracle. For this set of clients, the true average causal effect
θ = 1/8, which is quite small, implying that in reality, the
choice of ISP has no or little effect on the performance for
these clients. Although the in situ dataset is consistent with
the ground-truth, i.e., Y = GX , there is a clear discrepancy
between the observed association and the true causal effect.

(a) (b) (c)
Original Dataset Ground Truth (Oracle) Random Treatment
X Y G0 G1 X Y
a 0 1 1 1 1 1
b 0 1 1 1 0 1
c 0 1 1 1 1 1
d 0 1 1 1 0 1
e 1 0 0 0 1 0
f 1 0 0 0 0 0
g 1 0 0 0 0 0
h 1 1 0 1 1 1
α = −3/4 θ = 1/8 α = 0
Table 1: (a) Observed Association (α) in a passive dataset (b) True
causal effect (θ) in an example dataset: α 6= θ. (c) Association con-
verges to causal effect under random treatment assignment: α ≈ θ.
2.3 Approaches for Estimating Causal Effect
This section presents two techniques for estimating the
causal effect, θ. The first, random treatment, involves an ac-
tive experiment, where we randomly assign the treatment to
the clients and observe the association. The second, adjust-
ing for confounding variables, is a passive technique, where
we work with only an in situ dataset and estimate the overall
causal effect by aggregating the causal effect across several
small strata.
1. Random Treatment. Because the ground-truth values
(G0, G1) are not simultaneously observable, we cannot es-
timate the true causal effect (Eq. 1) from an in situ dataset
alone. Fortunately, if we assign the clients to the treatment in
a way that is independent of the outcome, then under certain
conditions, association is an unbiased estimator of causal ef-
fect. This property holds because when X is independent of
GX , then E(GX) = E(GX |X) = E(Y |X); see [9, pp. 254–
255] for a proof. In Table 1(c) we randomly assign a treat-
ment, 0 or 1, to the clients and see that association, α, con-
verges to the true causal effect, θ.
For association to converge to causal effect with random
treatment, all other variables in the system that have a causal
association with the outcome variable must remain the as
we change the treatment. In the case of the example above,
association will converge to true causal effect under random
treatment, if and only if the original ISP and the alternative
ISP are both similar except for their discrimination policy.
Random treatment is difficult to emulate in the Internet for
two reasons. First, it is difficult to make users switch to an
arbitrary ISP, because not all ISPs may offer services in all
geographical areas, the users may be contractually bound to
a particular ISP, and asking users to switch ISPs is incon-
venient for users. Second, if changing the ISP brand also
means that the users must access the content through a rad-
ically different network which could affect the service per-
formance, then we cannot use the mere difference of per-
formance seen from the two ISPs as indication of interfer-
ence: the association may not converge to causal effect un-
der these conditions because the independence condition is
not satisfied. This situation is called operational confound-
ing: changing the treatment inadvertently or unavoidably
changes a confounding variable.
2. Adjusting for Confounding Variables. Because it is
difficult to emulate random treatment on the real Internet and
control operational confounding, we need to find a way to
adjust for the effects of confounding variables. NANO uses
the well-known stratification technique for this purpose [6].
Confounding variables are the extraneous variables in the
inference process that are correlated with both the treatment
and the outcome variables. As a result, if we simply ob-
serve the association between the treatment and the outcome
variables, we cannot infer causation or lack of it, because
we cannot be certain whether the change is due to change
in the treatment variable or a change in one or more of the
confounding variables.
With stratification, all samples in a stratum are similar in
terms of values for the confounding variables. As a result,
X and GX are independent of the confounding variables
within the stratum, essentially creating conditions that re-
semble random treatment. Thus, the association value within
the stratum converges to causal effect, and we can use asso-
ciation as a metric of causal effect within a strata.
2a. Challenges. This approach presents several challenges.
First, we must enumerate the confounding variables and col-
lect sufficient data to help disambiguate the true causal ef-
fect from the confounding effects. Second, we must define
the stratum boundaries in a way that satisfies the above con-
ditions. Unfortunately, there is no automated way to enu-
merate all the confounding variables for a given problem;
instead, we must rely on domain knowledge. Section 3 ad-
dresses these challenges.
2b. Formulation. In the context of NANO, we have multi-
ple ISPs and services; we wish to calculate the causal effect
θi,j that estimates how much the performance of a service j,
denoted by Yj , changes when it is accessed through ISP i,
versus when it is not accessed through ISP i. Let Z denote
the set of confoundingvariables, and s a stratum as described
above. The causal effect θi,j is formulated as:
θi,j(s;x) = E(Yj |Xi = x, Z ∈ B(s)) (3)
θi,j(s) = θi,j(s; 1) − θi,j(s; 0) (4)
θi,j =
∑
s
θi,j(s) (5)
B(s) represents the range of values of confounding variables
in the stratum s. θi,j(s) represents the causal effect within
the stratum s. A key aspect is the term θi,j(s; 0) in Equa-
tion 4: it represents the baseline service performance, or the
service performance when the ISP is not used; we define this
concept in more detail in Section 3.1.2. Note that the units
for causal effect are same as for service performance, so we
can apply simple thresholds to detect discrimination.
2c. Sufficiency of Confounding Variables. Although there
is no simple or automatic way to enumerate all the confound-
ing variables for a problem, we can test whether a given list
is sufficient in the realm of a given dataset. To do so, we pre-
dict the value of the outcome variable using a non-parametric
regression function, f(), of the treatment variable, X , and
the confounding variables, Z , as yˆ = f(X ;Z). We then
compare the predicted value with the value of outcome vari-
able observed in the given dataset, y, using relative error,
|y − yˆ|/y. If X and Z are sufficient to define the outcome
Y , then the prediction error should be small.

3. NANO System Design
This section explains how NANO performs causal infer-
ence, enumerates the confounding variables required for this
inference, and describes the system architecture for collect-
ing and processing the relevant data.
3.1 Establishing the Causal Effect
Estimating the causal effect for a service degradation in-
volves three steps. First, we stratify the data. Next, we esti-
mate the extent of causal impact of possible ISP interference
within each stratum and across the board. Finally, we try to
infer the criteria that the ISP is using for discrimination.
3.1.1 Stratifying the data
To stratify the data, NANO creates bins (i.e., ranges of val-
ues) along the dimensions of each of the confounding vari-
ables, such that the value of the confounding variable within
the bin is (almost) constant. The bin size depends on the
nature of the confounding variable. As a general rule, we
create strata such that there is a bin for every unique value of
categorical variables; for the continuous variables, the bins
are sufficiently small, that the variable can be assumed to
have essentially a constant value within the stratum. For
example, for a confounding variable representing the client
browser, all the clients using a particular version and make
of the browser are in one stratum. Similarly, we create one
hour strata along the time-of-the-day variable.
We use simple correlation to test whether the treatment
variable and the outcome variable are independent of the
confounding variable within a stratum. We combine adja-
cent strata if the distribution of the outcome variable condi-
tioned on the treatment variable is identical in each of the
stratum; this reduces the total number of strata and the num-
ber of samples needed.
3.1.2 Establishing the baseline performance
A thorny aspect of Equation 4 is the term θi,j(s; 0), which
represents the baseline service performance, or the service
performance when the ISP is not used. This aspect raises
the question: What does it mean to not use ISP i to ac-
cess service j? It could mean using another ISP, k, but if
ISP k is also discriminating against service j, then θk,j(s; 1)
will not have the (neutral) ground-truth baseline value. To
address this problem, NANO takes θi,j(s; 0) as the average
service performance when not using ISP i, calculated as:
∑
k 6=i θk,j(s; 1)/(ns − 1), where ns > 2 is the number of
ISPs for which we have clients in stratum s.
An important implication of defining the baseline in this
way is that NANO is essentially comparing the performance
of a service through a particular ISP against the average
performance achieved through other ISPs, while adjusting
for the confounding effects. If all or most of the ISPs
across which NANO obtains measurements are discriminat-
ing against a service, it is not possible to detect such discrim-
ination using the above definition of baseline; in this case,
discrimination becomes the norm. In such cases, we might
consider using other definitions of discrimination, such as
the comparing against the best performance instead of the
average, or using a performance model of the service ob-
tained from laboratory experiments or mathematical analysis
as the baseline.
3.1.3 Inferring the discrimination criteria
NANO can infer the discrimination criteria that an ISP uses
by using simple decision-tree based classification methods.
For each stratum and service where NANO detects discrimi-
nation, NANO assigns a negative label, and for each stratum
and service where it does not detect discrimination, it assigns
a positive label. NANO then uses the values of the confound-
ing variables and the service identifier as the feature set and
uses the discrimination label as the target variable, and uses
a decision-tree algorithm to train the classifier.
The rules that the decision tree generates indicate the dis-
crimination criteria that the ISP uses, because the rules in-
dicate the boundaries of maximum information distance be-
tween discrimination and the lack of it.
3.2 Confounding Variables
Confounding variables are the extraneous factors in in-
ferring whether an ISP’s policy is discriminating against a
service; these variables correlate, either positively or neg-
atively, with both the ISP brands and service performance.
Because there is no automated way to enumerate these vari-
ables for particular problem, we must rely on domain knowl-
edge. In this section, we describe three categories of con-
founding variables and explain how they correlate with both
the ISP brands and the service performance. In Section 3.3,
we describe the specific variables that we collect to adjust
for these confounding variables.
Client-based. Client-side applications, as well as system
and network setup, can confound the inference. The particu-
lar application that a client uses for accessing a service might
affect the performance. For example, in the case of HTTP
services, certain Web sites may be optimized for a particu-
lar Web browser and perform poorly for others. Similarly,
certain Web browsers may be inherently different; for exam-
ple, at the time of this writing, Opera, Firefox, and Internet
Explorer use different number of simultaneous TCP connec-
tions, and only Opera uses HTTP pipelining by default. For
peer-to-peer traffic, various client software may experience
different performance. Similarly, the operating system and
the configuration of the client’s computer and local network,
as well as a client’s service contract, can affect the perfor-
mance that the client perceives for a service.
We believe that the above variables also correlate with ISP
brand, primarily because the ISP may serve particular com-
munities or localities. As an example, we expect that Mi-
crosoft’s Windows operating system may be more popular
among home users, while Unix variants may be more com-
mon in academic environments. Similarly, certain browsers
may be more popular among certain demographics and lo-
calities than other.
Network-based. Various properties of the Internet path,
such as location of the client or the ISP relative to the lo-
cation of the servers on the Internet, can cause performance
degradation for a service; such degradation is not discrim-
ination. Similarly, a path segment to a particular service

Figure 1: NANO System Architecture.
provider might not be sufficiently provisioned, which could
degrade service. If we wish to not treat these effects as dis-
crimination, we should adjust for the path properties.
Time-based. Service performance varies widely with time-
of-day due to changes in utilization. Further, the utilization
may affect both the ISPs and the service providers, thus con-
founding the inference.
3.3 Data Collection andAnalysis Architecture
An important aspect of NANO is collecting the necessary
data for facilitating inference. This section describes the cri-
teria for data collection, the features that we collect, and,
finally, the data collection mechanism.
1. Criteria. The criteria for data collection has two parts.
First, the feature should quantify the treatment variable, the
outcome variable, or the values of the confounding factors.
Second, the data should be unbiased.
The first criterion helps us determine a set of features for
which to collect data; this list is explained below. We can
collect many of these features through active or passive mon-
itoring. The second criterion, however, suggests that we
must take care that the measurements are not biased. As we
discussed in Section 1, ISPs may have the incentive to inter-
fere with identifiable active measurements to deter inference
of discrimination or improve their rankings. Similarly, we
believe that while we could use the data directly from service
providers as the “baseline” service performance, such infor-
mation could be biased in the favor of the service provider.
Therefore, to the extent possible, NANO relies on passive
measurements to determine the values of the features.
2. Mechanism. Given the nature of the confounding factors
and the desire to collect data passively, we believe the best
place to collect this data is usingmonitoring agents at clients.
Figure 1 shows the system architecture. The primary source
of data for NANO are client-side agents installed on com-
puters of voluntarily participating clients (NANO-Agents).
Each agent continuously monitors and reports the data to
the NANO servers. We are developing two versions of this
agent. The first is a Web-browser plug-in that can monitor
Web-related activities, and the second is a packet-level snif-
fer that can access more fine-grained information from the
client machines.
3. Dataset Features. The NANO-Agent collects three sets
of features for the confounding factors, corresponding to the
three classes of confounding variables (Section 3.2)
First, the NANO-Agent collects features that help identify
the client setup, including the operating system, basic sys-
tem configuration and resource utilization on the client ma-
chine. Second, NANO-Agents perform active measurements
to a corpus of diverse benchmark sites (PlanetLab nodes)
to establish the topological location of the clients and their
ISPs. These measurements include periodic short and long
transfers with the benchmark sites. These measurements are
similar in spirit to ones used by many Internet coordinate
systems [5]. NANO uses this information to establish the
topological properties of the ISP and stratify ISPs with sim-
ilar topological location to adjust path properties factor. Fi-
nally, all the data is time-stamped to allow adjustment for the
time-of-day factor.
To identify the treatment variable, i.e., ISP brand, we use
the IP address of the client and look it up againstwhois reg-
istry servers. To determine the performance for each service,
the NANO-Agent monitors and logs the information about
the ongoing traffic from the client machine for the services
that NANO monitors. The sniffer version of the agent logs
the network round-trip time (RTT) measurements to various
destinations for small and large packets. It also collects un-
sampled flow statistics for the ongoing flows to determine
the throughput, and also maintains the applications asso-
ciated with each flow. The latter is used to disambiguate
the performance differences that might be associated with
particular applications. The NANO-Agent tags this infor-
mation with a service identifier that it infers by inspecting
the packet payloads (e.g., by looking for regular-expression
google.com/search?q= in the HTTP request message
to identify search service), or, if possible, by looking at the
protocol and port numbers.
4. Simulation
To illustrate how NANO can detect ISP discrimination
against a particular service, we evaluate the technique in
simulation for a specific example. A rigorous validation of
the approachwill ultimately require a real deploymentwhere
there is less control over confounding variables; we discuss
this issue and various others in more detail in Section 5.
Our simulation setup comprises three ISPs, ISPA, ISPB ,
and ISPC , that provide connectivity for two services, S1 and
S2 for their respective clients. The clients use one of the two
types of applications, App1 and App2 to access the services
S1 and S2. Performance for the service S1 is sensitive to the
choice of application: clients perceive better performance
when using application App1, compared to using App2. The
performance of service S2 is agnostic to the choice of ap-
plication. The distribution of clients using the two types of
applications is different across the ISPs (e.g., due to demo-
graphics). In particular, we set the distribution of App1 to
be 60%, 10%, and 90%, across the three ISPs, respectively.
This distribution makes the client application a confounding
variable because its distribution correlates with both the ISP
and service performance.
We set up the experiment such that ISPB throttles the
bandwidth for all of its clients that access service S1. To
achieve this, we configure the request rates from the clients
and the available throttled bandwidth between the ISPs and
the services to achieve certain expected utilization levels. In

(a) Association
Service S1 Service S2
Baseline 7.68 2.67
ISPB 8.60 2.71
Association 0.92 0.04
(b) Causal Effect
Service S1 Service S2
Confounding Var. App1 App2 App1 App2
Baseline 9.90 2.77 2.61 2.59
ISPB 11.95 7.95 2.67 2.67
Causal Effect 2.0 5.18 0.06 0.12
Table 2: Simulation Results: Estimating the causal effect between
ISPB and S1. All the numbers are request completion times in seconds.
particular, we configure the links so that the expected uti-
lization on all links is about 40%, except for the traffic from
ISPB to service S1, which faces about 90% utilization.
We aim to detect discrimination by ISPB against the ser-
vice S1, and establish a causal relationship. We compute
the association and causal effect using Equation 2 and Equa-
tion 4, respectively. We use the average response time seen
through ISPA and ISPC , combined, as the baseline.
Table 2 presents the association and causal effect esti-
mated for this simulation. Table 2(a) shows that ISPB has
little to no association for either service S1 or S2. How-
ever, as we stratify on the application variable, we see in
Table 2(b) that ISPB has significant causal effect for both
applications for service S1, but there is still no causal effect
for service S2. This example shows that, for this case, NANO
can establish a causal relationship where one exists (ISPB
and service S1), and rule out where one does not exist, e.g.,
between ISPB and S2.
5. Summary and Research Agenda
We presented Network Access Neutrality Observatory
(NANO), a system for inferring whether an ISP is discrim-
inating against a particular service. We have examined only
basic criteria for discrimination in a simulation environment;
in future work, we will evaluate NANO’s effectiveness in the
wide area, for a wider range of possible discrimination ac-
tivities, and in adversarial settings where ISPs may attempt
to evade detection. In this section, we pose several questions
that are guiding our ongoing work.
How can NANO-agents be deployed? NANO relies on par-
ticipating clients to collect the data needed to perform causal
inference. PlanetLab and CPR [4] nodes are our initial de-
ployment candidates, but in the long run, we wish to incen-
tivize home users to deploy NANO-Agents. Because NANO
inference engine must exclude all extraneous factors, includ-
ing transient faults to establish ISP discrimination, NANO-
Agents can additionally act as a network troubleshooting and
diagnostics application that users might find useful.
How can NANO-agents protect user privacy while still
exposing enough client-side information to expose dis-
crimination? Because some of the measurements that
NANO-Agent collects can lead to invasion of user privacy,
NANO stores the data in a stratified form and does not main-
tain any client-identifiable data (e.g., client IP addresses or
search queries). Further, we are instrumenting NANO to
give users full control over the services that the a NANO-
agent can monitor. Finally, we are investigating ways of dis-
tributed inference, were we may be able to mitigate the need
for aggregating the data at a central repository for inference,
instead, the NANO server can act as a coordinator and the
clients infer the discrimination in a peer-to-peer fashion.
Is NANO general enough to detect diverse, evolving, and
adversarial discrimination policies? ISPs may continue
to evolve their policies for distinguishing between various
services. One such policy is imposing quotas for Internet
traffic from a client [2]; in the future, ISPs may extend this
policy by exempting traffic to certain partner sites from such
quotas, thereby creating a discriminatory environments. We
believe that NANO can quickly detect such new policies and
infer the criteria if we measure sufficient metrics about the
state of the network and service.
How much data is needed to perform inference? NANO
requires a collection of sample data inputs from each stra-
tum to be able to adjust for each confounding variable. The
greater the variance of the confounding variables, the more
data samples are needed to adjust for each of them and estab-
lishing confidence bounds. While statistics theory does help
determine the number of samples needed for each stratum,
each variable will be distributed differently across the set of
clients and ISPs, so it is difficult to determine how many
clients would need to run NANO-Agents to allow sufficient
confidence levels for inference. We expect to understand this
better as we deploy these agents.
Acknowledgements
We would like to thankMostafa Ammar at Georgia Tech and
Joseph Hellerstein at UC Berkeley for their valuable feed-
back that helped improve several aspects of our work. This
work is supported in part by NSF Awards CNS-0643974,
CNS-0721581, and CNS-0721559.
REFERENCES
[1] A Guide to Network Neutrality for Google Users.
http://www.google.com/help/netneutrality.html.
[2] Comcast Excessive Use Policy.
http://help.comcast.net/content/faq/
Frequently-Asked-Questions-about-Excessive-Use.
[3] Glasnost: Bringing Transparency to the Internet.
http://broadband.mpi-sws.mpg.de/transparency/.
[4] CPR: Campus-Wide Network Performance Monitoring and
Recovery. http://www.rnoc.gatech.edu/cpr/, 2006.
[5] F. Dabek, R. Cox, F. Kaashoek, and R. Morris. Vivaldi: A
decentralized network coordinate system. In Proc. ACM SIGCOMM,
Portland, OR, Aug. 2004.
[6] N. Jewell. Statistics for Epidemiology. Chapman & Hall/CRC, 2004.
[7] J. Pearl. Causality: Models, Reasoning, and Inference. Cambridge
University Press, 2000.
[8] S. B. Robert Beverly and A. Berger. The internet’s not a big truck:
Toward quantifying network neutrality. In Passive & Active
Measurement (PAM), Louvain-la-neuve, Belgium, Apr. 2007.
[9] L. Wasserman. All of Statistics: A Concise Course in Statistical
Inference. Springer, 2003.
[10] Y. Zhang, Z. M. Mao, and M. Zhang. Ascertaining the Reality of
Network Neutrality Violation in Backbone ISPs. In Proc. 7th ACM
Workshop on Hot Topics in Networks (Hotnets-VII), Calgary, Alberta.
Canada., Oct. 2008.