h t t p : / / w w w . e e c s . c a s e . e d u /

DipZoomA Marketplace for Internet Measurements

Faculty: Michael Rabinovich

Students: Sipat Tiukose, Zhihua Wen

Ack: Limin Wang EECS Department

Case Western Reserve University

DipZoomDipZoom 2

Internet Measurements

A-priori measurement platforms (e.g. IDMaps) Great for large-scale characterizations Not focused enough for some applications

On-demand measurements (Keynote, Scriptroute/PlanetLab, traceroute servers) Hard to deploy sufficient platform to serve unpredictable needs Vulnerable to being “gamed” Limited choice of measurements

DipZoomDipZoom 3

Example: Latency Measurements

Measuring latency between a laptop connected through VZACCESS BroadbandAccess and a Linux PC on Case network

RTT measured by average of 2000 pings: 280 msec

King Measurements

DipZoomDipZoom 4

DIPZoom Goals

• Fulfill the need for zoomed-in, on-demand measurements

• Provide infrastructure that scales with The Internet– Scaling with Internet size – Scaling with Internet diversity (devices, connectivity)– Scaling with measured metrics and regime

• Simplify staging of measurement experiments– Complex multi-step measurements– Distributed measurements across multiple measuring hosts

(Hence the name: Deep Internet Performance Zoom)

DipZoomDipZoom 5

Coax Internet users to become measurement providers

Deploy a matchmaking service instead of measurement infrastructure

Use market driven approach with real money as the means to control the system

DIPZoom Approach

DipZoomDipZoom 6

DipZoom: Deep Internet Performance Zoom-In

Anyone can offer measurements Anyone can request measurements Anyone can offer measuring software Participants are free to set their prices, compete for requests, bid

and solicit bids, etc. Facilitates an open echosystem, “ebay for Internet

measurements” Client library allows programmatic access to the entire platform

from one computer

DipZoomDipZoom 7

Some Questions

Will anyone want to become a provider? Seti@home, upromise.com, gomez.com

suggest “yes”.

Will anyone pay for the measurements? Keynote and Gomez suggest “yes” Keynote is in business with 53M dollars revenue

last year.

DipZoomDipZoom 8

Related Work

Gomez.com Closed commercial system

DIMES, Traceroute@home No incentives Users participate in a particular measurement experiment Users can’t ask for a measurement

On-demand measurement systems (Keynote, Scriptroute, NIMI, GNP, King) Hard to scale with Internet Hard to stage experiments

DipZoomDipZoom 9

System Overview

DipZoomDipZoom 10

Issues

Security Induced DoS attacks against measurement target Open ports on measuring hosts High-level DoS attack against measuring host Unauthorized resources access

Payment trust Trusted core helps Replay-based cheating

DipZoomDipZoom 11

Issues

Measurement trust Fake Measuring Point (MP) registrations Measuring Point (MP) impersonation Fake measurements

Traversing firewalls and NATs

DipZoomDipZoom 12

Core Needs

Integrity of measuring software Globally unique ID of measuring host (MHID) Duplicate detection + request/response

matching Measurement rate limiting

• Building blocks of a solution:– Unique embedded secrets– GUID or MAC address or Host ID– Nonces (“numbers used once”)– Rate limiting (per measurement host and per target)– Ranking and calibration

DipZoomDipZoom 13

Ranking and Calibration

Security measures raise the bar but do not guarantee protection from malicious MH.

If can’t protect - detect and blacklist! Deploy calibrating measurement targets Purchase measurements from suspect MHs Compare responses with passive measurements by calibrating

hosts Not all measurements can be calibrated (wget vs. ping vs.

nslookup)

• Can calibrating hosts be gamed?–Keynote advertises its measuring hosts location–Calibrating hosts are secret–The risk of blacklisting deters data mining

DipZoomDipZoom 14

Usage Example 1:Accuracy of King Measurements

King (between well-connected)

King (between well-connectedand residential)

King validation using DIPZoom

• King accuracy is sensitive to the type of access network

• Inaccuracy is biased to under-estimation

DipZoomDipZoom 15

Usage Example 2: Accuracy of GNP measurements

• GNP is more accurate and less sensitive to the type of access connection

• The number of measurements needed to achieve accuracy depends on access connection

10 Measurements per pair 2.5 measurements per pair

DipZoomDipZoom 16

Usage Example 3: Lowering the Barrier of Entry for Experimenters

Assessing the quality of server selection by Akamai Discovering a large number of Akamai servers Comparing the performance of Akamai-selected server with

other Akamai servers

Currently: Intricate experiments Subject of research papers (Su et al. SIGCOMM’06;

Krishnamurthy et al. IMW’01)

With DIPZoom: 1-month course project for undergrads Done from students’ laptops with no special arrangements from

measurement providers

DipZoomDipZoom 17

Status

First release available for download http://dipzoom.case.edu/ Includes a GUI client and client class library for scripted

experiments Includes NAT/firewall traversal Works on Windows, Linux, Mac OS X Now offers ping, wget, traceroute, and nslookup.

No payments yet No ranking and calibration Around 150 MHs deployed on PlanetLab and

several home computers

DipZoomDipZoom 18

Next Up

Pluggable measurement tools DipZoom veneer over existing infrastructures (e.g.

Scriptroute)

Down the line: Ranking and calibration Payments Distributed DipZoom core

DipZoomDipZoom 19

Summary

Growing Internet diversity (devices, links, applications) entails growing needs for focused measurements

Proprietary platforms are insufficient Measurements are increasingly complex DipZoom: a facilitator instead of infrastructure

Open system (pricing, measurements, participants) Unified programmatic access to entire platform Market approach to system control

Many challenges ahead

Web site for more information:http://dipzoom.case.edu/

h t t p : / / w w w . e e c s . c a s e . e d u /

Questions and Comments