Measurement of OneWay Transit Time in IP Routers

1
Measurement of One-Way Transit Time in IP Routers

• HET-NETs05 Working Conference
• 18 20 July 2005
• Ilkley, West Yorkshire, United Kingdom
• Adrian Popescu and Doru Constantinescu
• Dept. of Telecommunication Systems
• Blekinge Institute of Technology
• Karlskrona, Sweden

2
Outline

• Introduction
• Router Architecture
• One-Way Transit Time
• Queueing Delay in Chained IP Routers
• Measurement Setup
• Estimation of OWTT and Other Router Delays
• Sources of Errors
• Modeling Methodology

3
Outline (cont.)

• Experiments
• Processing Delay of a Router
• Router Delay for a Single Data Flow
• Router Delay for More Data Flows
• End-to-End Delay for a Chain of Routers
• Conclusions
• Future Work

4
Introduction

• Measurements of One-Way Transit Time (OWTT) and
  other router delays
• Goals
• Design of a measurement system to follow
  specifications of IETF RFC 2679
• Delay measurements
• Understanding the delay process in IP routers

5
Router Architecture

• Basic activities
• Routing
• Datagram forwarding

6
One-Way Transit Time

• OWTT has several components
• where the delay per node i
• OWTT can be partitioned into a deterministic
  component and a stochastic component

7
Queueing Delay in Chained IP Routers

• Fundamental problem
• Traffic merging
• Main consequences
• Character of arrival process at a downstream
  queue changes
• Appearance of correlations
• Important classes of correlations
• Autocorrelations in packet interarrival times
• Autocorrelations in packet service times
• Crosscorrelations between packet interarrival
  times and packet service times
• Crosscorrelations between packet service times in
  tandem queues
• Another important consequence
• Almost impossible to do precise queueing analysis
• Actual solution used
• Kleinrock independence assumption

8
Measurement Setup

• Dedicated Measurement Points (MPs) equipped with
  (synchronized) DAG 3.5E
• Control in generating and capturing network
  traffic
• UDP traffic generated with TCP-like
  characteristics
• High accuracy of timestamps
• Off-line data analysis

9
Measurement Setup (cont.)

• Control of link utilization Lu and Hurst
  parameter H
• Pareto distributed traffic for the packet length
  generated at the application level with the shape
  parameter a
• Inter packet gap-time exponentially distributed
  with parameter ?
• Number of traffic sources n
• Traffic generation
• Traffic generators
• Generated traffic type
• World Wide Web-like traffic at the application
  level
• Fractional Brownian Motion (fBm) at the network
  level
• Packet identification
• Hashing and masking
• SHA-1 algorithm
• Packet matching
• Use of template containers defined by the
  Standard Template Library

10
Estimation of OWTT and Other Router Delays
11
Estimation of OWTT and Other Router Delays (cont.)

• Matrices used in estimation of different delays
• Timestamps for packet n captured by DAGi Ti(n)
• Interarrival times for packet n captured by DAGi
  IntArri(n)Ti(n)-Ti(n-1)
• Service times for packet n Serv(n)
• One-Way Transit Times for packet n measured
  between DAGj and DAGi OWTTij,i(n)Tj(n)-Ti(n)
• Router transit times for packet n
  RTT(n)OWTTj,i(n)-Serv(n)
• Minimum delay for a specific packet size L
  DminlnL
• Queueing delay for packet n Queue(n)RTT(n)-Dmi
  nlnL

12
Sources of Errors

• Duplicate packets
• Very low probability of occurrence, due to
  strictly controlled environment as well as own
  generated traffic
• Unmatched packets
• Mostly because of other interfering traffic,
  e.g., ARP and inter-router traffic as well as
  because of congestion avoidance in the router
  during heavy-load traffic conditions
• Low probability of occurrence, 0.01 to 5 for
  more than one million packets processed

13
Modeling Methodology

• Selection of candidate distribution(s)
• Use of visual techniques (CCDF plots, EDF plots,
  PDF plots, Hill plots, a-estimation
  plots)
• Determining whether a single or mixture of
  distributions is required
• Parameter estimation
• Maximum Likehood Estimation (MLE) method
• Use of successive right censoring in the case of
  mixture of distributions
• Fitness assessment
• Goodness-of-fit significance tests (null
  hypothesis) Kolmogorov-Smirnov, l2 and
  Anderson-Darling can be used
• Drawback they always tend to reject the null
  hypothesis in the case of large sample
• Own developed method (David Erman) similar to
  the EDF test, but it does not suffer as much with
  increasing size of sample space

14
Experiments

• Classes of OWTT experiments
• One router with single data flow
• One router with more data flows
• Chain of routers with more data flows
• 9 experiments done for each class of experiments
  with different H, Lu and combinations of traffic
  mixture

15
Experiments (cont.)
16
Processing Delay of a Router

• Example of CISCO 3620 router processing delay
  for ICMP and UDP payloads

17
Processing Delay of a Router (cont.)

• Minimum router transit time for a specific
  packet size DminlnL obtained, with 95
  confidence bounds, in experiments 1-3 and 1-7

18
Router Delay for a Single Data Flow

• Measurement configuration

19
Router Delay for a Single Data Flow (cont.)
20
Router Delay for a Single Data Flow (cont.)

• Main observations
• Limited disparity in summary statistics
• Variance and mean slightly increasing with Lu and
  H
• One delay has a maximum that is unusual large
  (55.5 ms)

21
Router Delay for a Single Data Flow (cont.)

• Delay distributions obtained in
  experiment 1-9

22
Router Delay for a Single Data Flow (cont.)

• Modeling results obtained for delays in
  experiment 1-9

23
Router Delay for a Single Data Flow (cont.)
Modeling of delays obtained in experiment 1-9
24
Router Delay for a Single Data Flow (cont.)
25
Router Delay for a Single Data Flow (cont.)

• Results obtained in experiment 1 on power
  spectrum

26
Router Delay for More Data Flows

• Measurement configuration

27
Router Delay for More Data Flows (cont.)
28
Router Delay for More Data Flows (cont.)

• Main observations
• Larger disparity for some OWTT statistics
• Samples with large delays are more common
• Heavy tail observed in histograms (dependent on
  Lu and H)

29
Router Delay for More Data Flows (cont.)

• Modeling results obtained in experiment
  2-5

30
Router Delay for More Data Flows (cont.)
Modeling of delays obtained in experiment 2-5
31
Router Delay for More Data Flows (cont.)
32
End-to-End Delay for a Chain of Routers
Measurement configuration
33
End-to-End Delay for a Chain of Routers (cont.)

• Inte mycket text här

34
End-to-End Delay for a Chain of Routers (cont.)

• Main observations
• Large disparity for all statistics (except
  minimum)
• Large number of samples with large delays
• Heavy tail observed in histograms (dependent on
  Lu and H)

35
End-to-End Delay for a Chain of Routers (cont.)

• Modeling results obtained in experiment 3-7 at
  routers R1, R2 and R3

36
End-to-End Delay for a Chain of Routers (cont.)

• OWTT distributions obtained in experiment 3-7 at
  routers R1, R2 and R3

37
End-to-End Delay for a Chain of Routers (cont.)

• Summary of measured OWTT performance

38
Conclusions

• Dedicated measurement system for delay
  measurements in IP routers (IETF RFC2679)
• Measurement study of delay through IP routers
• We confirm earlier results about the dependency
  of delay on traffic characteristics, link
  conditions, hardware implementations and IOS
  releases
• New results indicate that the delay in IP routers
  can be well modeled with the help of three
  distributions

39
Future Work

• Analytical models for OWTT, to include possible
  correlations between packet service times at
  adjacent nodes
• E2E delay formula in a chain of IP routers
• Targets