User-Perceived Performance Measurement on the Internet

1
User-Perceived Performance Measurement on the
Internet

• Bill Tice
• Thomas Hildebrandt
• CS 6255
• November 6, 2003

2
Introduction

• The Internet is different from LANs
• Diversely administered
• Users are relatively distant from the network
  administrators
• IP is the only service provided by the network
• SNMP or something like it? No

3
Overview

• Why user-perceived performance measurement is
  important and also difficult
• Existing solutions to the global network
  measurement problem
• NIMI
• E2E piPEs
• Client-side proxies and other solutions

4
Performance Measurement on the Internet

• IP was designed around the end-to-end argument
  certain functions are not required at low levels
  of a system 1
• As a result, we have a dumb network with no
  built-in performance measurement architecture
• The tools we can count on are limited

5
Performance Measurement on the Internet Example
Tools

• Packet Internet Groper (ping) Useful to test
  connectivity and RTT between hosts
• Traceroute Provides an approximation of the
  network topology in the forward direction
• Iperf Measures TCP and UDP performance,
  including bandwidth, delay jitter and packet loss

6
Performance Measurement on the Internet

• Ping and traceroute rely on ICMP, which is
  increasingly being blocked by network
  administrators
• Active measurement based on the use of common
  network services is more robust and perhaps more
  realistic
• DNS queries
• HTTP requests

7
Performance Measurement Architectures

• There have been projects to create measurement
  architectures either by deployment at nodes
  within domains or at the users
• NIMI
• E2E pIPEs
• AMP, Medusa Proxy, Liston Proxy, Network Weather
  Service

8
NIMI(National Internet Measurement
Infrastructure)

• Software system for building network measurement
  infrastructures
• Diversely administered
• Facilitate many kinds of measurements
• Extensible and Modular

9
NIMI Architecture

• Measurement servers (probes)
• 2 daemons nimid and scheduled
• Measurement Modules

10
NIMI Architecture

• Configuration Point of Contact (CPOC)
• Configuration and Control Servers
• Several per domain
• Delegation of configuration access

11
NIMI Architecture

• Measurement Client (MC)
• Only user point of contact with NIMI system
• Measurement requests

12
NIMI Architecture

• Data Analysis Client (DAC)
• Data Collection
• Post-Processing
• Run at central location or at MC

13
NIMI Measurement Modules

• NIMI has no knowledge of measurements
• Plug-ins
• Wrapped for a standard API
• Current Modules
• traceroute, treno, zing, mflect, traffic, ftp

14
NIMI Measurements

• Request Received from MC
• Access Control List
• scheduled creates pending measurement
• Results sent to DAC by nimid
• All communications encrypted

15
NIMI Deployment

• 35 hosts (as of 2001)
• Research laboratories and Universities
• Georgia Tech
• Dead?

16
NIMI Difficulties

• Laboratory Conditions
• High bandwidth
• Dedicated Resources
• Not representative of Internet as a whole

17
NIMI Difficulties

• Meaningful data for User Perceived measurements
• Not widely distributed

18
NIMI Difficulties

• Hosts administration difficulties
• Tools requiring privileged access
• Updates

19
NIMI Difficulties

• Difficult to distribute
• Whats in it for me?
• Not attractive to average user

20
E2E piPEs

• End-to-end Performance Improvement Performance
  Environment System
• A framework to indicate performance capabilities
  and locate performance problems along the path
  between two computers connected by the Abilene
  network

21
E2E piPEs Architecture
22
E2E piPEs Architecture
23
OWAMP

• One-Way Active Measurement Protocol An
  Internet2 project
• A UDP-based protocol to precisely measure network
  characteristics
• Loss
• Delay
• Jitter
• http//owamp.internet2.edu/ - Under construction

24
E2E piPEs Status

• The initial deployment, which includes the
  Abilene backbone network and two campuses only,
  is scheduled for Fall 2003.
• piPEfitters are still developing the system
  one suggestion is to place PMPs at the end hosts.

25
Other Tools Proxy-Based

• Liston Proxy
• Between browser and the Web
• Handles DNS resolution and content requests
• Logs information of interest
• DNS responsiveness
• Response time

26
Proxy-Based Tools

• Medusa Proxy
• Cool name
• Monitors performance
• DNS
• Akamai edge servers vs. origin servers

27
Proxy-Based Disadvantages

• Privacy issues
• Overhead
• Limited Scope

28
Distributed Tools

• AMP Network
• Distributed physical nodes
• High Performance Computing (HPC) Network
  measurement, not user-perceived

29
Distributed Tools Disadvantages

• Wide Distribution
• Need large amounts of diverse data
• How to do it
• Updates
• Security

30
Applications

• Network Weather Service
• Predicting network performance for applications
• Can be run by user to predict their performance

31
Applications

• Lots
• Zing
• Ping
• IPerf
• Traceroute

32
NM Applications Disadvantages

• Limited Scope
• Generally Stateless

33
User-Perceived Performance Measurement Why?

• Because the Internet was designed following the
  end-to-end principle, end-to-end performance is
  ultimately the most meaningful to measure
• It is difficult to deploy a measurement
  architecture in the Internet backbone
• Network users see end-to-end performance directly
  and could be effectively used as monitor points

34
User-Perceived Measurement How?

• The infrastructure
• NIMI-like plug-in measurement modules
• Standardized communication between components

35
User-Perceived Measurement How?

• Distribution
• Have to make it something users want to run

36
User-Perceived Measurement How?

• Processing
• Centralized data collection and post-processing

37
References

• 1 J.H. Saltzer, D.P. Reed, and D.D. Clark.
  End to End Arguments in System Design.
  http//web.mit.edu/Saltzer/www/publications/endtoe
  nd/endtoend.txt
• 2 Vern Paxson, Andrew Adams, and Matt Mathis.
  Experiences with NIMI. In Proceedings of
  Passive and Active Measurement, 2000.
  http//citeseer.nj.nec.com/paxson00experiences.htm
  l
• 3 Internet2. E2E piPEs. http//e2epi.internet2.
  edu/E2EpiPEs/e2epipe_index.html
• 4 Richard Liston and Ellen Zegura. Using a
  Proxy to Measure Client-Side Web Performance.
  Proceedings of the 6th International Web Caching
  and Content Distribution Workshop, Boston, MA,
  June 1999. http//www.cc.gatech.edu/liston/pubs/
  proxy_wcw01.ps.gz
• Network Weather Service http//www.npaci.edu/envis
  ion/v15.2/nws.html