NANO: Network Access Neutrality Observatory Mukarram Bin Tariq, Murtaza Motiwala, Nick Feamster, Mostafa Ammar Georgia Tech Tuesday, October 7, 2008 1
Net Neutrality Tuesday, October 7, 2008 2
Net Neutrality Tuesday, October 7, 2008 2
Net Neutrality Tuesday, October 7, 2008 2
Example: BitTorrent Blocking Tuesday, October 7, 2008 3
Example: BitTorrent Blocking http://broadband.mpi-sws.mpg.de/transparenccy Tuesday, October 7, 2008 3
Many Forms of Discrimination Throttling and prioritizing based on destination or service Target domains, applications, or content Tuesday, October 7, 2008 4
Many Forms of Discrimination Throttling and prioritizing based on destination or service Target domains, applications, or content Discriminatory peering Resist peering with certain content providers ... Tuesday, October 7, 2008 4
Problem Statement Identify whether a degradation in a service performance is caused by discrimination by an ISP Quantify the causal effect Tuesday, October 7, 2008 5
Problem Statement Identify whether a degradation in a service performance is caused by discrimination by an ISP Quantify the causal effect Existing techniques detect specific ISP methods TCP RST (Glasnost) ToS-bit based de-prioritization (NVLens) Tuesday, October 7, 2008 5
Problem Statement Identify whether a degradation in a service performance is caused by discrimination by an ISP Quantify the causal effect Existing techniques detect specific ISP methods TCP RST (Glasnost) ToS-bit based de-prioritization (NVLens) Goal: Establish a causal relationship in the general case, without assuming anything about the ISP’s methods Tuesday, October 7, 2008 5
Causality: An Analogy from Health • Epidemiology: study causal relationships between risk factors and health outcome • NANO: infer causal relationship between ISP and service performance Tuesday, October 7, 2008 6
Does Aspirin Make You Healthy? Tuesday, October 7, 2008 7
Does Aspirin Make You Healthy? Sample of patients Aspirin No Aspirin 40% 15% Healthy Positive correlation in 10% 35% Not Healthy health and treatment Tuesday, October 7, 2008 7
Does Aspirin Make You Healthy? Sample of patients Aspirin No Aspirin 40% 15% Healthy Positive correlation in 10% 35% Not Healthy health and treatment Aspirin Can we say that Aspirin ? causes better health? Health Tuesday, October 7, 2008 7
Does Aspirin Make You Healthy? Sample of patients Aspirin No Aspirin 40% 15% Healthy Positive correlation in 10% 35% Not Healthy health and treatment Sleep Aspirin Diet Can we say that Aspirin Duration ? causes better health? Other Gender Health Drugs Confounding Variables: correlate with both cause and outcome variables and confuse the causal inference Tuesday, October 7, 2008 7
Does an ISP Cause Service Degradation? Tuesday, October 7, 2008 8
Does an ISP Cause Service Degradation? Sample of client Comcast No Comcast performances BitTorrent 5 sec 2 sec Some correlation in Download Time ISP and service performance Tuesday, October 7, 2008 8
Does an ISP Cause Service Degradation? Sample of client Comcast No Comcast performances BitTorrent 5 sec 2 sec Some correlation in Download Time ISP and service Client Comcast ToD performance Setup ? Can we say that Comcast is BT Download Location Content discriminating? Time Many confounding variables can confuse the inference. Tuesday, October 7, 2008 8
Causation vs. Association (1) Causal Effect = E(Real Download time using Comcast) E(Real Download time not using Comcast) Tuesday, October 7, 2008 9
Causation vs. Association (1) Performance with the ISP Causal Effect = E(Real Download time using Comcast) E(Real Download time not using Comcast) Tuesday, October 7, 2008 9
Causation vs. Association (1) Performance with the ISP Causal Effect = E(Real Download time using Comcast) E(Real Download time not using Comcast) Baseline Performance Tuesday, October 7, 2008 9
Causation vs. Association (1) Performance with the ISP Causal Effect = E(Real Download time using Comcast) E(Real Download time not using Comcast) Baseline Performance θ = E ( G 1 ) − E ( G 0 ) G 1 , G 0 : Ground-truth values for performance (aka. Counter-factual values) Tuesday, October 7, 2008 9
Causation vs. Association (1) Performance with the ISP Causal Effect = E(Real Download time using Comcast) E(Real Download time not using Comcast) Baseline Performance θ = E ( G 1 ) − E ( G 0 ) G 1 , G 0 : Ground-truth values for performance (aka. Counter-factual values) Problem: Generally, we do not observe both ground truth values for the same clients. Consequently, in situ data sets are not sufficient to directly estimate causal effect. Tuesday, October 7, 2008 9
Causation vs. Association (2) We can observe association in an in situ data set. Tuesday, October 7, 2008 10
Causation vs. Association (2) We can observe association in an in situ data set. Association = E(Download time using Comcast) E(Download time not using Comcast) Tuesday, October 7, 2008 10
Causation vs. Association (2) We can observe association in an in situ data set. Observed Performance with the ISP Association = E(Download time using Comcast) E(Download time not using Comcast) Tuesday, October 7, 2008 10
Causation vs. Association (2) We can observe association in an in situ data set. Observed Performance with the ISP Association = E(Download time using Comcast) E(Download time not using Comcast) Observed Baseline Performance Tuesday, October 7, 2008 10
Causation vs. Association (2) We can observe association in an in situ data set. Observed Performance with the ISP Association = E(Download time using Comcast) E(Download time not using Comcast) Observed Baseline Performance α = E ( Y | X = 1) − E ( Y | X = 0) Tuesday, October 7, 2008 10
Causation vs. Association (2) We can observe association in an in situ data set. Observed Performance with the ISP Association = E(Download time using Comcast) E(Download time not using Comcast) Observed Baseline Performance α = E ( Y | X = 1) − E ( Y | X = 0) In general, . al α � = θ . Tuesday, October 7, 2008 10
Causation vs. Association (2) We can observe association in an in situ data set. Observed Performance with the ISP Association = E(Download time using Comcast) E(Download time not using Comcast) Observed Baseline Performance α = E ( Y | X = 1) − E ( Y | X = 0) In general, . al α � = θ . How to estimate causal effect ( ) ? θ . Tuesday, October 7, 2008 10
Estimating the Causal Effect Two common approaches a. Random Treatment b. Adjusting for Confounding Variables Tuesday, October 7, 2008 11
Random Treatment H !H !H H !H H H !H !H Tuesday, October 7, 2008 12
Random Treatment Given a population: H !H !H H !H H H !H !H Tuesday, October 7, 2008 12
Random Treatment Given a population: H !H !H H !H 1. Treat subjects with Aspirin randomly, H H !H !H irrespective of their health Aspirin Not Aspirin Treated Treated Tuesday, October 7, 2008 12
Random Treatment Given a population: H !H !H H !H 1. Treat subjects with Aspirin randomly, H H !H !H irrespective of their health Aspirin Not Aspirin Treated Treated 2. Observe new outcome and H H H !H H measure association H !H !H !H α = 0.8 - 0.25 = 0.55 Tuesday, October 7, 2008 12
Random Treatment Given a population: H !H !H H !H 1. Treat subjects with Aspirin randomly, H H !H !H irrespective of their health Aspirin Not Aspirin Treated Treated 2. Observe new outcome and H H H !H H measure association H !H !H !H α = 0.8 - 0.25 = 0.55 3. For large samples, association converges to causal effect if confounding variables do not change θ . Diet, other drugs, etc. should not change Tuesday, October 7, 2008 12
Random Treatment (How to apply to the ISP Case?) Tuesday, October 7, 2008 13
Random Treatment (How to apply to the ISP Case?) • Ask clients to change their ISP to an arbitrary one Tuesday, October 7, 2008 13
Random Treatment (How to apply to the ISP Case?) • Ask clients to change their ISP to an arbitrary one • Difficult to achieve on the Internet Changing ISP is cumbersome for the users Changing ISP may change other confounding variables, i.e., the ISP network Tuesday, October 7, 2008 13
Adjusting for Confounding Variables 1. List confounders e.g., gender ={ , } Tuesday, October 7, 2008 14
Adjusting for Confounding Variables H !H H H 1. List confounders !H H !H !H !H H H H !H H e.g., gender ={ , } !H H !H !H H !H !H H !H H !H !H 2. Collect a data set !H !H H !H H H H H H H An in situ data set Treatment: Baseline: Border( , ) Treated: No border Outcome: Healthy (H), Not Healthy (!H) Stratum: Type {Circle, Square} Tuesday, October 7, 2008 14
Recommend
More recommend
