IP Performance Measurements using

SurveyorMatt Zekauskas

matt@advanced.org

Guy Almes, Sunil KalidindiAugust, 1998

ISMA 98

Outline

• Background• Surveyor infrastructure• Reporting and analysis• Status

I: Background

• Internet topology is increasingly complex

• Commonly used measurement tools (like ping and traceroute) are inadequate

• Result: users don’t understand the Internet’s performance and reliability

IP Performance Metrics

• IETF IPPM effort– Framework RFC– One-way delay and packet loss drafts– Others: connectivity, bulk transfer, DV

• Surveyor: implementation of one-way delay and packet loss metrics

Motivation for measuring delay • Minimum of delay:

transmission/propagation delay • Variation of delay: queuing delay• Large delay makes sustaining

high-bandwidth flows harder• Erratic variation in delay makes

real-time apps harder

Uses

• Problem determination• Engineering (trends, loads)• Feedback to advanced applications

(e.g., Tele-Immersion, CMU’s Odyssey)

• Monitor QoS

One-way versus round trip

• Paths are asymmetric • Even when paths are symmetric,

forward and reverse paths may have radically different performance - asymmetric queuing

II. The Surveyor Infrastructure• Measurement machines at campuses

and at other interesting places along paths (e.g., gigaPoPs, interconnects)

• GPS to synchronize clocks• Centralized database to store

measurement data• Web based reporting and analysis

tools

II. Surveyor Infrastructure

Measurement machines

Measurement Machines

• Dell 400 MHz Pentium Pro • 128 MBytes RAM; 2 GBytes disk• BSDI Unix• TrueTime GPS card and antenna• Network Interface (10/100bT,

FDDI)• Special driver for the GPS card

Measurement Technology• Active tests of one-way delay and loss

– Measurement daemon– Test packets time-stamped with GPS time– Back-to-back calibration: 95% of

measurements ± 50 s– Measurements centrally managed

• Truer-time daemon to watch clocks

Ongoing Tests

Ongoing Tests - Delay

• Type-P– 12 byte UDP packets, 40 bytes total– Port “random” per session

• Scheduled using a Poisson Process– average rate: 2 per second

• “Mostly” full mesh

Ongoing Tests - Routing

• Traceroute to same sites as One-Way delay

• Scheduled with Poisson process– average rate: one every 10 minutes

Collecting Results

Central Database Machine

• SGI Origin 200• 2 processors, 256MB• 327GB Fibrechannel-attached RAID

for data storage(DataDirect Networks EV-1000)

Central Database Machine

• Collects performance data from the measurement machines [ssh, pull]

• Stores the data in a home-grown database

• Serves data and summaries to reporting and analysis tools [http]

Current Surveyor Deployment• 28 machines, 623 paths

– CSG Schools– Tele-Immersion Labs– National Labs– NASA Ames– CA*net2 Ottawa site– Auckland, NZ– …others

Surveyor Map (N. America)

III. Reporting and analysis tools• Web based Tools• Daily summary reports• Integration with traceroute

measurements

Daily summary reports

• Take a 24-hour sample for a given path

• Divide it into one-minute sub-samples

• For each one-minute sub-sample:– Minimum delay (blue)– 50th percentile (green)– 90th percentile (red)

Example daily reports

• Advanced Network & Services and University of Chicago– path is symmetric– asymmetric queuing

Examples (continued)

• Advanced Network & Services and University of Pennsylvania– path asymmetric

Examples (continued)

• CMU to Brown University

Examples - Route Change

• Advanced Network & Services toPenn State University

• Route change switched providers, and removed one provider from the path

Examples - Auckland

• University of Auckland, NZ toUniversity of Washington, Seattle

• Asymmetric queuing, congested trans-pacific path

IV Status

• Deployment rate: 1/week• Planned: Abilene backbone

– probe at each backbone router– experiment with piecewise delay

Full Mesh of End-to-end Paths

O(N2) paths

Paths with Exchange Points

O(X2+N)

Abilene

Universities

gigaPoPs

Router Nodes

Near-term improvements

• Improve measurement software– time stamping in-kernel: to scale

without losing accuracy

• New and improved analyses– real-time display tools– flag interesting paths– trends …– improved data export to other sites

Summary

• One-way Delay and Loss are – practical – useful

• Surveyor infrastructure growing• Now focus on analysis and

applications

More info

• Surveyor project info– http://www.advanced.org/surveyor/– Email: mm-info@advanced.org

• Access to plots– Email me - matt@advanced.org

• IETF IPPM WG– http://www.advanced.org/IPPM/