Ascertaining the Reality of Network Neutrality Violation in Backbone ISPs
by Ying Zhang, Z. Morley Mao and Ming Zhang.

This talk was presented by Ming Zhang from Microsoft Research. 

Net neutrality is a hot topic. The Internet was designed for a 
best-effort service model, but the emergence of new applications
with more demanding service requirements means that ISPs need to 
implement QOS measures as well as traffic shaping. This leads to 
traffic discrimination and way for ISPs to make more profit.
Large ISPs stand to gain the most from traffic discrimination, 
while content providers like Yahoo and Microsoft as well as smaller
ISPs are interested in network neutrality. There have been a number
of known cases of traffic discrimination already, for example, 
it was shown recently that Comcast throttles BitTorrent traffic, 
but few studies have looked at backbone networks. The authors have
developed a tool called NVLens that allows them to do this. 
NVLens uses probes with different TTLs to measure content based 
as well as previous-hop AS based differences in traffic rate at 
the ingress and egress of a backbone ISP. The methodology consists 
of two phases; the first phase involves mapping the topology of 
the backbone ISP, while the second phase measures loss rate as well
as latency in the probe traffic. The results shown were based on 3 
weeks of data collected on traffic of various types of applications,
and shows that traffic discrimination is indeed being carried out. 
There are still some open questions including when traffic 
discrimination is actually taking place (e.g. for high loads only),
as well as how to deal with traffic discrmination. The methodology 
used suffers from the inability to detect source based traffic
discrimination. 

Q&A:

(1) ISP reaction and detection of NVLens.
ISPs will not be able to detect NVLens, since the payload carries 
traffic that appears valid. The only way to detect the probes is 
by examining the TTL values, which can be expensive. Additionally, 
the author hopes more content providers will start doing this, so 
source based detection (currently NVLens is deployed in PlanetLab)
will not work either.

(2) Discrimination vs differentiation
The paper uses the term differentiation, and only ISPs know if it 
is discrimination for sure, since SLAs are confidential. The goal 
nevertheless is to bring transparency to end users and content 
providers so they can take the appropriate action.


NANO: Network Access Neutrality Observer
by Mukarram bin Tariq, Murtaza Motiwala and Nick Feamster

This talk was given by Mukarram Tariq from Georgia Tech University.

The discussion on network neutrality can be traced back to a 
statement by the AT&T CEO who said that ISP "pipes are not free".
There have been clear cases of traffic discrimination, for example,
in the case of Comcast which was throttling BitTorrent traffic.
When measuring degradation of a service, we should be able to state 
with certain amount of confidence that the degradation is caused by 
traffic discrimination and not by other factors, or confounding 
variables. In this paper, the authors use methods from epidemology
to study the causal relationships between ISPs and service performance.
Correlation does not equate to causation as there could be other 
confounding variables that should be taken into account. Instead, we 
should measure for causal effects using expected values of ground-truth
variables. The problem with this is that these variables differ for
different clients and ISPs. Instead, we can use the observed values
to measure association. When using a large enough random sample, 
association will converge to causation. However, it is not feasible
to change ISPs randomly. Instead, the solution is to stratify the 
observed samples along confounding variables, and with all things 
being equal, association implies causation. Verification of the 
methodology was done using simulations, and work is underway to 
deploy this method on PlanetLabs.


Q&A

(1) Errors and Regression Analysis
Using regression analysis, the errors should be small.
Correlations between confounding variables does not affect 
the error level of the analysis.


Q&A with authors of both talks

(1) NANO vs NVLens
NANO is used for a single protocol across many ISPs, while NVLens 
is used for many protocols on a single ISP backbone network. In NANO,
ICMP rate throttling in ISPs can be viewed as a confounding variable,
while NVLens suffers from the inability to distinguish ICMP rate 
throttling from true losses of the probe traffic. The argument for 
NVLens is that ISPs use the simplest method when discriminating traffic. 

(2) BitTorrent traffic and NANO
When there are many clients and many ISPs, like in the case of 
BitTorrent traffic, NANO is best equipped to detect traffic 
discrimination, as NANO treats the variability in client 
download/upload speeds as confounding variables.

(3) NANO simulation results
The NANO simulation results presented in the paper are based on 
HTTP traffic, but work is currently being done to deploy a BitTorrent
tracker on PlanetLabs to extend the results to real P2P traffic.

(4) ISP detection
Can ISPs detect probes by NVLens? Currently, NVLens makes no attempt
at hiding itself. The main characteristic is the TTL values, as the 
payload is made to look like real traffic. ISPs may decide to block 
traceroute, but this is highly unlikely, as traceroute is useful to 
their network administrators too. It would also be possible to extend
NVLens so that TTL is not needed, assuming access to both sender
and receiver.