NET NEUTRALITY

OUTLINE
- Types of differentiation
- Glasnost (Active measurement)
- NANO  (Passive measurement)
- Quality of Service
- Traffic shapping in enterprises/campuses

WHAT IS NET NEUTRALITY?
- "ISPs remain neutral to how they forward user traffic, irrespective of
  content, application or sender"

TRAFFIC DIFFERENTIATION
- An ISP treats packets of one flow differently than those of another
- How can traffic be differentiated
    1) Based on flow types
        - IP header - src/dst addrs
        - Transport header - port numbers
        - Packet payload - use deep packet inspection
    2) Independent of flow type
        - Time of day
        - Network load
        - User behavior - differentiate against users with heavy BW usage
    3) Traffic manipulation
        - Block packets; reset connection
        - Decrease priority
        - Drop packets - at fixed or variable rate
        - Modify TCP window size
- Do not know what mechanism is being used for discrimination
- Glasnot detects only based on flow type using transport hdrs and pkt payload

MEASUREMENT
- Challenges
    - Mechanism for discrimination may not be known
    - Baseline performance is not known
    - Many factors could give the appearance of differentiation
- What factors affect measurements?
    - End host OS
    - Network path (src, dst, route)
    - Cross-traffic
- Passive vs. Active
    - Passive - looks at traffic already being sent; problem is it introduces
      lots of other factors (OS on end hosts, different routes, etc.)
    - Active - sends traffic specifically for measurement; this approach is
      also used by tools like SpeedTest

GLASNOST (ACTIVE) 
- Send all flows to same end host - eliminates all variables except transient
  noise; also, in theory, your route could change in the middle of the test
- Use pairs of flows
    - Send first flow with payload/ports that expect to be differentiated
    - Send second flow with random payload/ports
- Flows are 20 seconds in duration - enough time to get to stable TCP rate
- Analysis
    - Use data from server and client
    - Repeat each test twice
    - Calculate difference between maximum throughput for each flow in pair
    - If difference is greater than 50%, assume differentiation (false
      positive rate is 0.7%, false negative rate is 1.7%)
- Dealing with noise
    - Running flows simultaneously is worse than running one after the other
    - Four patterns of cross-traffic
        - Consistently low - narrow range of throughput measurements
        - Mostly low, but occasionally high - only a few high outliers
        - Highly variable - no consistent pattern and wide range
        - Mostly high, but occasionally low - only a few low outliers
    - Discard any tests with second two patterns of cross-traffic

NANO (PASSIVE)
- Network Access Neutrality Observatory (NANO)
- "Tries to establish a causal relationship between an ISP's policy and the
  observed degradation of performance for a service using only passively
  collected data."
- Causal effect "X causes Y"
    - X is the "treatment variable"
    - Y is the "outcome variable"
    - X is accessing a particular service through an ISP
    - Y is the observed performance
- Only have the outcome; need to find confounding variables - variable that
  correlates with both X & Y
- Use stratification to adjust for confounding variables
    - Essentially "placing all measurements where everything that could 
      possibly be attributed to performance is equal, except for the ISP"
    - Normally requires enumerating confounding variables - we don't know
      these (although later things assume we know a possible set of these)
    - Create "bins" for each confounding variable - ex. client browser
      (discrete) and time-of-day (continuous, split in 1 hour bins)
- Compute causal effect
    - Find average throughput for each strata - baseline
    - Clients with same ISP whose throughput is below baseline indicates
      discrimination
    - Use decision-tree based classification to determine discrimination
      criteria
- NANO clients
    - Gather data using PCAP
    - Filter out private traffic
    - Do local stratification
    - Send results to NANO server via encrypted connection
- NANO server
    - Receives data from clients
    - Runs stratification & computers causal effect
- Use Click, Emulab, and PlanetLab for emulation - show that yes it works

EXPERIMENTATION
- AT&T DSL - no BitTorrent or flash video traffic shaping
- CS Dept - no BitTorrent of flash video traffic shaping
- Class results 

QUALITY OF SERVICE
- Positive use of traffic differentiation
- What are some factors applications care about?
    - Throughput (IPTV)
    - Dropped packets (IPTV, alarm signaling)
    - Delay (VoIP, vieo chat)
    - Jitter (VoIP, video chat)
- Implementing QoS
    - Reserve resources ahead of time - Resource Reservation Protocol (RSVP);
      not scalable 
    - Differentiated services (DiffServe) - mark packets in IP header; need
      router support (separate queues for different service levels)
    - Use UDP - avoids being affected by TCP's congestion control

CASE STUDY: MARQUETTE UNIVERSITY
- BitTorrent is the main issue
- Two pipes: 100Mb regular, Internet2
    - Regular pipe at MAX except from 4am to 6am
    - Use a Packeteer middle box at border - at some point bandwidth exceeded
      Packeteer's ability to process packets
    - Allow only certain percentage of traffic based on application -
      primarily determined by port
    - To evade, use non-standard ports
    - Added DPI - students started to encrypt traffic; cannot examine
      encrypted traffic
- Three pipes: 100Mb shared, 100Mb for faculty/staff, Internet2
    - Pipe is based on user type - determined by building or wireless NW
    - Limit to maximum daily bandwidth use; placed in slow bin until midnight
      - based on port or MAC address; lots of accounting overhead
- Glasnost tests (on student pipe)
    - Rate limit uploads on port 6881 (2 Kbps)
    - Rate limit downloads on port 6881 and by DPI (5 Kbps)