RMcgeer.ppt

04/13/23 Copyright © 2007 HP corporate presentation. All rights reserved. 1

© 2004 Hewlett-Packard Development Company, L.P.The information contained herein is subject to change without notice

PlanetLab, DARPA Control Plane, And Collaborative Research

Rick McGeerIdes of March, 2006


What is PlanetLab?• A planetary-scale “overlay” network

− Just a bunch of Linux machines that have agreed to communicate− Global account: account on the whole network− Go beyond protocols to services and applications − “The next internet will be created as an overlay in the current one”

(NRC)

• What’s it good for?− A virtual machine creation service− A platform for running planetary-scale distributed applications− A platform for critical, pervasive, robust services− Characteristic of the next generation of the Internet

• Joint Academic/Government/Industry Consortium has formed

− Formally announced in late June ‘03− Hosted by Princeton, U Washington, and UC Berkeley− Google has joined HP and Intel as founding industrial members− NSF funded

• Extension to Japanese National Research Network Fall 2006


PlanetLab’s GrowthPlanetlab Growth

0

100

200

300

400

500

600

700

800

Date

Nu

mb

er

Start of Project

SOSP ’03: 2/3submissions areon PlanetLab

HP Announces support

Consortium Formed

Brazil RPN Joins

CERN Joins

JGN-II Joins


PlanetLab Today

718 Nodes at 345 Sites


Today’s Internet• A 1973 Buick Retroffited with airbags and emission

controls (Technology Review)− Local services distributed across the wide area− Smart client, smart server, dumb pipe in between

Vulnerable•To local outages, flash crowds•Fails when it’s needed most

•cnn.com overwhelmed on 9/11 •When you need it, so does everybody else!

•No intelligence or adaptivity

•Fails on weakest link•Unable to adapt to single point of failure


Tomorrow’s Architecture:PlanetLab• Intelligence throughout

the cloud• Planetary-scale

distributed services

• Service that runs everywhere, all the time

• Means location of service virtualized − In fact, meaningless

• Gives attributes:− Critical: There when it’s

needed most− Pervasive: Always

available, anytime, anywhere

− Robust: Can’t be stopped, killed, crashed

− Secure: Can’t be spoofed or compromised


Case Study: Reinventing the Internet

1980 Today

Number of Users ~10,000 ~2,000,000,000Line Speeds ~10 kb/s 300kb/s-10 Gb/sNumber of Nodes

~1,000 ~1,000,000,000

Applications •Ftp•Telnet•Email

ftp, telnet, email, im, http, mmpg, voip, iptv…

Question: How would we reinvent the Internet today?


Key Point• 1980 Internet model: no intelligence in the network• Network behavior guessed at endpoints

− Led to complex endpoint behavior− Poor behavior under minimal degradation

• Network elements had limited communication− Could only communicate gross network behavior− Couldn’t adjust to minor changes

• Led to architecture where global conditions guessed from local observations

• Suppose endpoints could talk to the network?• Suppose network elements could sense network

conditions• Simpler endpoints• More adaptive network


Research Model• Two levels

− Improve core internet services− Leave artifacts for application services− Application services on top

• Open Research− Developed by coalitions− Build coalitions to solve real-world problems


Better Internet Behavior• Problem: TCP degrades badly in presence of loss

− 5-20% packet loss common in military environments− Weather loss, satellite tracking,…− Means most Internet apps fail

• Assumptions− TCP fixed (don’t want to change every client)− Can add some hardware to each side− Losses are transient− Alternate paths exist

• Solution− Avoid packet loss by rapidly switching away from lossy

links− Software Routing based on Distributed Hash Tables

(DHTs) and overlays


CHART• Comprehensive Hyperplane for Adaptive

Response and Throughput• Set of five loosely-coupled services

− PlanetLab Classic: Base VM creation service− DHT-based software router: Better IP layer on the overlay− Information Plane: comprehensive network sensing

service− Explicit-rate routing

• Hardware Routing Service: next generation routers for high-bandwidth links

• Overlay software routing service: software implementation− Security Service: authentication for information plane

• Big idea− Fix TCP/IP and leave artifacts for application services


CHART Solution Concept

•Conventional TCP: Slowly ramp up to line rate

•CHART TCP:•Get line rate from routers•Transmit immediately at line rate

•Conventional TCP: Backoff in presence of loss since congestion-free transmission rate unknown

•CHART TCP: Keep going in presence of loss since congestion-free transmission rate known

Click for next slide

0 60

4020

0 60

4020


Chart Routing Concept

Conventional: losses in network undetected

Continued transmission over poor links

When combined with TCP backoff, leads to very poor performance

CHART: Links constantly monitored

Rapid switch away from failing links

When combined with TCP explicit rate, line rate performance


Adaptive Routing Scenario

Sw

itch

10% link loss detected


Adaptive Routing Implementation

Sw

itch

Sensor Servers

Router continually queries sensor servers for loss/latency/bandwidth data

Router switched routes when sensor servers report better path


Phase I accomplishments

0

2

4

6

8

10

12

14

16

18

20

Throughput Ratio

Phase 1 Goal

Program Goal

Phase 1 Achieved


Do More Than Routing• Control Plane makes networks intelligent,

adaptive• “Send me the best picture you can of

downtown Baghdad in the next five minutes”− Sensing of link quality, QoS can optimize request

• End-system multicast (CMU)− On-the-fly multicast tree based on load, b/width,

latency

• Intelligent, bandwidth-aware media/file distribution


Today’s Web Service


Tomorrow’s Web Service: CoDeeN• Network of proxies spread across the whole net

− Looks in cache for page− Finds it? Done− Doesn’t find it

• hashes URL• Goes to hash proxy for page• Replicates requested pages

• Load on each host manageable• Load on each proxy manageable• No flash crowds unless whole Internet

overloaded− Has never happened

• Running today on PlanetLab


Tomorrow’s Web Service


Today’s Filesystem• One location only

− If it is down, so are you− Major security concerns on the system


Today’s File System• Files attached to one computer

− Accessable (easily) there only− Vulnerable to local disruption− Hard to back up (media expensive, hard to

administer)− Vulnerable to multiple failures

• crash• media failure• power failure• localized network disruption


Tomorrow’s File System: OceanStore

• Spread Across the network− n pieces− m sufficient to reconstruct− pieces spread across

network

• File Lives everywhere!− Can’t be destroyed unless

n-m+1 destroyed− Accessable everywhere− Backed up by copy in

network− Lasts for 1000 years

• Running today on PlanetLab


PlanetLab Grand Challenges• Intel/HP/AT&T Initiative• Largely driven by Mic Bowman (Intel)

− Help from Jack Brassil, Rick McGeer (HP)− Rick Schlichting, Lee Breslau (AT&T)

• Mobilize academic/research community to solve large problems

• Model: Work with end customer to define RFP• Offer opportunity to research groups• One initiative: PBS Content Distribution• Coming: Large-scale multicast


Media Distribution• Currently done through

satellite or conventional Internet distribution− Satellites cheaper than

conventional Internet for > 40 sites

− Satellites have a variety or problems• Theft• Vulnerability to weather

disruption (rain fade)


Conventional Internet Distribution• Send media to each

endpoint (e.g., TV station)

• Bandwidth required− size of media x

number of endpoints

• More expensive than satellite for many endpoints


New Idea: Use PlanetLab distribution techniques (CoBlitz)• Send media

collaboratively from endpoint to endpoint

• Send chunks of file to each endpoint

• Endpoints send to each other

• Much more efficient than conventional Internet

• More secure, cheaper and reliable than satelite


Conclusion• CHART

− Model of new research model− Coalition of Industrial Lab/Startup/University

Research to solve large problem− Solve specified problem and leave artifacts for

future research

• Grand Challenges− Use distributed infrastructure to solve real-world

problems− Mine industrial/academic consumer community

for research opportunities

• We want to do more of this!

RMcgeer.ppt

Documents

network network behavior

adaptive network

line rate chart tcp

gross network behavior

tcp explicit rate

line rate conventional

tcp backoff

core internet services