History : Before W3C-LA

• Web Servers in
  • Boston
  • Sophia
  • Grenoble
  • Keio
• One Common Name : www.w3.org resolved by a modified DNS
  • Boston serving the World + W3C-MIT Team
  • Sophia serving W3C-Sophia Team + french users
  • Grenoble serving W3C-Grenoble Team + french users
  • Keio serving W3C-Keio Team
• Three Contents :
  • W3C Team Site
  • W3C Member Site
  • W3C Public Site
• Based on a Common Shared Filesystem (AFS) and a leased line USA-France
• Bad Reliability : numerous risks of failures

W3C-LA objectives

Providing a better access to the W3C Production : Place the W3C web site near customers

• One Mirror Site per Office (at least)
• Mirrors in badly connected countries

In terms of QoS, for members and public  :

• Improvement of performance : bring W3C production in badly connected countries
• Improvement of reliability :  decrease the risks of W3C resources unavailability

Questions to Solve

• How to update mirrors
• How to deal with partial mirrors
• How to deal with failures
  • Local Failures : a mirror goes out of sync
  • Global Failures : the general broadcasting system is down/unreachable
• How to share the Internet among an evolving set of mirrors

How to Transmit Pages/Resources ?

• Which Resources to transmit ?
  • in terms of modification time
  • in terms of ACL
• When to transmit Resources ?
  • System or user (W3C Customers) driven updates ?
  • Periodicity of updates :
    • True Real Time ?
    • As soon as possible / as soon as detected ?
    • Differed ?
• How to transmit data between master and mirrors ?
  • Which Protocol ?
  • Which Underlying Network ?

Classical Mirroring Solutions

HTTP & Proxy-Cache Solutions

Advantages

• Include Authentication (delegated to master servers) :
• Include Management of Modification Time
• Caches store only served documents

Drawbacks

• No caching of protected pages : no improvement for members
• User (W3C Customers) failure driven
• No 'Perhaps Out of Date' error Code : no way to serve pages if masters down

Conclusion : will not fit to badly connected countries and will not improve reliability

HTTP and Prefetching

• HTTP not designed as a transport protocol
• Poor compared to dedicated algorithms/protocols :
  • Diff Algorithm (dedicated 1-to-1)
  • Multicast (dedicated 1-to-many)

Experimentation

• 2 transport protocols tested :
  • RSYNC
  • MFTP : commercial reliable multicast product
• 2 physical networks :
  • Internet
  • Satellite
• 4 test sites :
  • RAL - England
  • CWI - Netherland
  • SICS - Sweden
  • FORTH - Greece

Results of the Experimentation

• MFTP over Internet
  • Not reliable with badly connected countries (Greece) due to :
    • The way multicast works over UDP
    • The limits of MFTP not implementing technologies on bandwidth adaptability
• RSYNC over Internet
  • Very good reliability due to TCP
  • Cheap solution
  • Few maintenance
  • Bad performance with badly connected countries
  • Bad scalability
    • if the traffic of www.w3.org increases
    • if the number of mirrors increases
    • if the connectivity between W3C sites and mirrors becomes worse :
      • Greece is part of the European backbone
      • Taiwan is not working
      • Morocco and Tunisia have a very bad Internet connectivity
• MFTP over Satellite
  • Very good performance
  • Very good scalability
  • Limited cover

Solution : Rsync + Mftp-Satellite

• MFTP-Satellite on all mirrors covered by satellite
• RSYNC between uplinks

How to deal with partial mirrors

• Problem : Mirrors will not contain the whole W3C site (no Team-only resources)
• Aim : Keep a common name space and keep transparency for all users (W3C staff included)
• Solution : development of a dedicated Apache module for transparent proxying

How to deal with failures

• Local Failures :
  • Neighbouring
  • Proxy (with IP authentication developed)
• Global Failures :
  • Implementation of "Perhaps out of date" error through CGI

How to assign Clients to mirrors

A two steps method

• An apriori assignment according to distance between clients and mirrors (AS path analysis)
• A check after assignment given the time to serve requests (Statistics on HTTP logs)

Deployment

• One satellite uplink in Sophia
• 5 mirrors in Europe :
  • RAL serving Janet Network
  • SICS serving the most of Swedish providers
  • FORTH serving most of Greek providers
  • CWI (*) no users
  • GMD  (**) : GMD users - rsync
• 2 mirrors in Asia :
  • Hong-Kong : serving HKUST
  • Taiwan (*) : no users
• 3 other mirrors scheduled :
  • CNR - Italy
  • Morocco
  • Tunisia

Achievements

• Improvement of reliability for members and public :
  • One week of failure in April for not European users
  • One day of failure in May
• Improvement of response time / performances
• Cheap, quick and reliable  way to start mirrors everywhere in "big Europe"
  

Web Mirroring : an internal W3C project ?

International acknowledgement of the technical work done :

• Accepted Informational RFC
• 2 Apache modules integrated in the Apache Core Distribution

Industrial Cooperation :

• Eutelsat
• StarBurst Communication
• JCSat

Academic Cooperation :

• Inria - Rodeo Group
• Bremen University
• University of Pisa
• Keio University

Interested Users :

• French Department of Education

Possible Outlets :

• All companies working with the same 1-to-many 'extranet' scheme
• Replication of any kind of files hierarchies

Involved People

• Luc Ottavj (Inria)
• Jose Kahan (W3C - Sysweb Team)
• Emmanuel Duros (Inria - Rodeo) & Patrick Cipiere (Inria - Rodeo)
• Jean-Laurent Wotton (W3C Summer Student)
• W3C-LA Technical Contacts :
  • Neil Calton (RAL)
  • Piet Beertema (CWI)
  • Lars Nilsander (SICS)
  • Demos Panagopoulos & Panos Psikos (FORTH)
  • Juergen Christoffel (GMD)
  • Samuel Kwan & Ban Szeto (HKUST)
  • Hao-Ren Ke (Taiwan University)
• The W3C Sysweb Team
  • Alan Kotok (Team Leader)
  • Gerald Oskoboiny
  • Renaud Bruyeron
  • Eric Prud'Hommeaux
  • Daniel Veillard
  • Takeshi Yamane (till april 31 1999)
  • Stephan Montigaud (till Dec 1 1997)
  • Pierre Fillault (till Jul. 31 1998)
  • Touyama Norio
  • Hirakawa Yasuyuki