Internet Performance

1
Internet Performance

• Presented by
• Deepa Srinivasan
• CSE581, Winter 2002, OGI

2
Papers on this topic

• End-End Effects of Internet Path Selection (99)
• Constancy of Internet Path Properties (01)
• Trends in Wide Area IP traffic (00)
• Resilient Overlay Networks (01)

3
For each paper...

• Introduction, Goal, Benefit
• Concepts
• Methodology tools for measurement
• Results
• Evaluation Conclusion

4
End-End Effects of Internet Path Selection
5
Goals of the paper

• Assess how good is internet routing from a
  users perspective?
• Path taken by a packet depends on large number of
  factors
• Constructed better synthetic paths for 30-80
  of the cases
• does not give a mechanism

6
Definitions

• Path Complete set of hops between two hosts
• Route Data structures exchanged between routers
  to describe connectivity
• Path selection combined set of route selection
  decisions
• Path Quality measured in terms of round-trip
  time, loss rate and bandwidth

7
Routing Overview

• Two-level routing hierarchy using Autonomous
  Systems (ASs)
• Intra and Inter AS routing
• Small ASs use hop-count for intra large use
  internal metrics
• BGP uses custom routing policy (or) default of
  minimum number of ASs.

8
Why is routing non-optimal?

• WAN routing protocols consider only connectivity,
  ignore path quality
• Per-network policies are difficult to manage no
  economic motivation
• Contractual agreements exist

9
For a good path...

• Admin of every AS on that path must have
  incentive
• No contractual or operational obligations
• Express it as a policy
• Not hindered by any other AS

10
Methodology

• Alternate paths constructing weighted graphs
• nodes as vertices
• measurement as edge weight
• Datasets
• using traceroute and tcpanaly
• Conservative estimate of potential inefficiency
• Uses long-term averages

11
Results

• Improvements in
• Mean RTT
• Mean Loss Rate
• Mean Bandwidth

12
Interpreting the graphs

• Probability Distribution is a description of the
  probabilities associated with the possible values
  of X.
• Cumulative Distribution Function
• P(X lt x) -? to x ?f(u)du
• i.e. Probability that X has a value less than or
  equal to x is given by the value of CDF at x.

13
Round-trip time
14
Loss Rate
15
Bandwidth
16
Observations

• RTT
• 30-55 has better RTT smaller fraction has more
  than 20 ms.
• Loss Rate
• 75 - 85 have better loss rates
• 5 - 50 has significant improvement
• Bandwidth
• 70 - 80 have improved bandwidth
• 10 - 20 of the paths, factor of 3

17
Evaluation

• Better alternate paths are chosen by avoiding
  poor quality routes, not because of a selective
  small number of hosts.
• Better path obtained by avoiding congestion,
  rather than minimizing propagation delay.
• (Propagation delay inclusive of physical
  transmission, store forward, processing
  overhead)

18
Conclusion

• Internet routing is non-optimal in terms of user
  perception
• Path quality measured by RTT, loss rate,
  bandwidth
• 30 - 80 of the time, there is a better path

19
Discussion

• Questions?

20
On the Constancy of Internet Path Properties
21

• Introduction, Goal, Benefit
• Concepts
• Methodology tools for measurement
• Results
• Evaluation Conclusion

22
Introduction

• Goal is to study constancy of (or) how steady
  Internet path properties are
• Measurements should be modelled and used for
  prediction

23
Concepts

• Mathematical Constancy
• data can be described with a single time
  invariant mathematical model
• e.g. session arrivals are modelled as a Poisson
  process
• (process behavior of a system observed over
  time)
• important to find appropriate model

24
Concepts

• Operational constancy
• quantities of interest remain within boundaries
  considered equivalent
• Predictive constancy
• past measurements allow for reasonable prediction
  of future characteristics
• Constancy is a more useful concept on coarser
  time scales than for finer time scales

25
Methodology

• NIMI - measurement platforms across Internet
• Uses zing to generate Poisson packet streams

26
Statistics Refresher

• IID
• Independent Identically distributed variable
• Poisson process
• n intervals p probability in an interval
• np is constant ? (Poisson rate)
• Change-free Region (CFR)
• Steady regions delineated by change points

27
Loss Constancy

• Loss episode process
• time series indicating when a series of packets
  were lost
• Loss Rate vs. Loss episodes
• Loss Process is not an IID
• Loss Episode Process is an IID - a Poisson
  process

28
Constancy
29
Conclusions

• Many processes are modeled as IID
• once change-points are identified
• Almost all predictors produce similar error
  levels
• How steady is the Internet?
• depends on particular aspect of constancy and
  dataset under consideration
• constancy on at least a scale of minutes

30
Discussion

• Questions?

31
Trends in Wide Area IP Traffic Patterns
32

• Introduction
• Methodology
• Results
• Conclusion

33
Introduction

• Trends are important for
• optimization of future networking equipment
• modeling effects of new protocols
• Internet Exchange
• a junction between multiple points of Internet
  presence
• peers directly connect to each other to exchange
  local Internet traffic
• NASA Ames IX (AIX)

34
Methodology

• Uses NLANR/MOATs NAI project
• First ATM cell is captured
• contains the TCP header
• Data collected over 10 months

35
Results

• Packet Lengths
• majority are 40, 1500 or 552 576 bytes
• no long term trend
• Protocol Mix
• TCP, UDP, GRE and ICMP
• growth of IPSEC, RealAudio is slower
• decline in FTP

36
Results

• Fragmentation is on the rise
• due to UDP
• e-mail traffic increased just before holidays
• Napster showed 50 increase in last 2 months

37
Results

• Online game traffic is on the rise
• doubles on weekends as compared to weekdays

38
Conclusion

• Analysis of various trends
• Need more work in protocol classification
• Questions?

39
Resilient Overlay Networks
40

• Introduction
• Design goals
• Design
• Implementation
• Evaluations
• Conclusion

41
Goals

• Enable a group of nodes to communicate when
  underlying paths have problems
• Integrate routing into distributed applications
• Framework for implementation of expressive
  routing policies

42
Design Goals

• Fast Failure Detection Recovery
• Link, Path Failures
• Perceived as outages/performance failures by apps
• Tighter integration with applications
• routing based on application-specific metric
• Expressive Policy Routing

43
Routing Path Selection

• Three metrics for each virtual link
• latency, packet loss rate, throughput
• Uses a link-state routing protocol to distribute
  topology information

44
RON Router

• Detects outages using active probing
• Implements 3 different metrics
• latency minimizer, loss-minimizer, TCP
  throughput-minimizer
• Policy Routing
• Classification
• Routing information from one RON node to all
  others
• Multi-Level Routing Tables

45
Resilient IP Forwarder

• No modifications to transport layer
• Uses FreeBSD divert sockets to send IP traffic
  over the RON
• Adds about 220?s to packet delivery

46
Evaluation

• Improvements for commercial sites came from
  commercial links
• Outage(t, p) if loss rate over interval t is
  greater than p on path
• 60 - 100 of path outages overcome

47
Improvements in RON

• Loss Rate
• improved by 0.05 a little more than 5
• worse in some cases
• Latency
• improved by 10s to 100s ms
• Throughput
• only 1 received lt 50
• 5 double throughput

48
Discussion

• Possible misuse of BGP transit policies
• Requires cryptographic authentication and access
  control
• Design scales only to 50 nodes
• sufficient for many distributed apps
• Problems with NAT devices

49

• Questions?

50
Conclusion

• Current Internet routing is non-optimal
• Constancy of metrics
• Trends in Wide Area IP Traffic
• RONs - improving end-end performance