A long-lived home-grown Linux-based low-cost leading mass-market service Pierre AUBERT - PowerPoint PPT Presentation

1 / 22
About This Presentation
Title:

A long-lived home-grown Linux-based low-cost leading mass-market service Pierre AUBERT

Description:

A long-lived home-grown Linux-based low-cost leading mass-market service Pierre AUBERT & Eric OLIVERI FranceTelecom/Home Communication Services/Broadband & Internet – PowerPoint PPT presentation

Number of Views:32
Avg rating:3.0/5.0
Slides: 23
Provided by: EricOl1
Learn more at: https://docbox.etsi.org
Category:

less

Transcript and Presenter's Notes

Title: A long-lived home-grown Linux-based low-cost leading mass-market service Pierre AUBERT


1
A long-lived home-grown Linux-based low-cost
leading mass-market servicePierre AUBERT Eric
OLIVERI FranceTelecom/Home Communication
Services/Broadband Internet
Wanadoo Portals
2
We talk about mass market net computer
engineering
  • the 1 in France
  • a simple web page
  • the Big Picture
  • pure Open source Linux PC hardware
  • some thoughts about the past and the future
  • works well, costs little ? delivers more from
    less

3
The 1 portals in France
  • wanadoo.fr voila.fr MMGs IM/IRC
  • 65 reach 13 M unique visitors
  • 80 M pages to 4.2 M customers daily
  • Every days peaks
  • 150 000 simultaneous users
  • 1 Gb/s Internet content
  • 5 000 new users/mn
  • 100 000 web pages/mn
  • transport conn/s, forks/s, SELECTs/s

4
We talk about mass market net computer
engineering
  • the 1 in France
  • a simple web page
  • the Big Picture
  • pure Open source Linux PC hardware
  • some thoughts about the past and the future
  • works well, costs little ? delivers more from
    less

5
A simple web page
6
, but
7
We talk about mass market net computer
engineering
  • the 1 in France
  • a simple web page
  • the Big Picture
  • pure Open source Linux PC hardware
  • some thoughts about the past and the future
  • works well, costs little ? delivers more from
    less

well, mostly
8
Not only a web server
  • Front office
  • Network Load Balancers Firewalls DNS
  • Web servers Apache/Mathopd/TUX, mostly PHP
  • Databases mySQL, postgreSQL
  • Back office
  • Publishing Content management
  • Audience collection
  • Databases
  • Supervision, backups, etc.
  • sometimes open source/Linux/PC hardware
  • most often non-open source/non-Linux/non-PC
    hardware

9
content management
audiencecollection
10
3 ways to deliver Web service
  • Apache for dynamic content
  • Mostly PHP, some C, some SSI
  • 20 M pages for 2.5 M users daily
  • 12 servers Bi-CPU Intel, type 1 60 Mb/s
  • Mathopd for static content
  • Maximize browser cache effect
  • 2000 HTTP/second
  • 6 servers 1-CPU Intel, type 0 25 Mb/s
  • TUX for redirection service
  • Built-in Linux kernel
  • 2500 HTTP/second 2 Mb/s
  • 3 servers (1 is enough) 1-CPU Intel, type -1
    )

11
Some other services
  • Cookies (identity, tracking)
  • 1250 HTTP/sec
  • 4 servers 1-CPU Intel type 0 10 Mb/s
  • Databases authentication
  • 1000 reads/mn 150 simultaneous connections
  • 1 server 1-CPU 1G RAM, type 1
  • Databases search engine crawler
  • 500 M docs 4 Tb 100k writes/mn
  • 12 servers 2-CPU 5x72Gb each, type 1
  • WAN-LAN
  • DNS 4 servers 1-CPU Intel type 0 Zebra
    (BGP/OSPF)
  • Firewall load-balance appliance 1-CPU Intel,
    type -1
  • Tricks lingerd, nscd, multicast
  • tuning 20 000 sockets per IRC server

12
Supervision etc.
  • 60 parameters per server on each of 1000
    servers
  • Net-SNMP, big brother/big sister, Perl
  • ad hoc monitoring ( 30) 1 h to develop and
    deploy
  • 20 servers 1-CPU Intel, type 1
  • At-the-fly graph generation 50 000 parameters
  • 300 samples/s 3 servers Bi-CPU Intel, type 1
  • RRD-Tool SNMP managers (Perl) Apache
  • Env Network from HWCPU to application
    internals
  • Backups, storage
  • Amandas self-planning, no routine restore ?
    1/10th FTE
  • 2 servers P2-400MHz 2,4To (2/GO/y) cheaper
    than tape (6/GO/y)
  • backs up daily 300 servers

13
We talk about mass market net computer
engineering
  • the 1 in France
  • a simple web page
  • the Big Picture
  • pure Open source Linux PC hardware
  • some thoughts about the past and the future
  • works well, costs little ? delivers more from
    less
  • well, mostly

14
Lessons from the past
  • Choose what fits best
  • 1996 from FreeBSD to Solaris
  • 2003 from Solaris to Linux
  • voila.fr launched 1998 Apache/PHP/Linux
  • Service first
  • system administrators are front-line
  • diagnose and fix fast (or alternate
    service/systems)
  • end-user satisfaction systems health
  • accounting transparency (CAPEX OPEX per
    server/day)

15
Lessons from the past
  • PC hardware is a key factor
  • Failures are facts not problems
  •  Small is beautiful  cost of redundancy is
    1/N
  • Competitive mass market sourcing
  • PC operating system is hw-independent
  • Reuse assemble components
  • Open source OS is a key factor
  • Best use of hardware resources
  • Flexible kernels Excellent network support
  • Application level
  • Open source platforms Internet applications
  • Interoperability is in DNA

16
Lessons from the past
  • Design skills
  • Not a sequential  product  process
  • Expert staff allow short long term savings
  • Efficiency
  • Low-range (90) 3 to 5 k/year
  • High-end PC servers cost 5 times more
  • Mid-range Unix servers cost 20 times more
  • Numerous distributed small systems
  • scale at front office
  • keep back office infra stable
  • distribute bandwidth, IO buffers, axis

17
We talk about mass market net computer
engineering
  • the 1 in France
  • a simple web page
  • the Big Picture
  • pure Open source Linux PC hardware
  • some thoughts about the past and the future
  • works well, costs little ? delivers more from
    less
  • well, mostly

18
Costs little
  • Some figures
  • 1 system administrator for 40 servers
  • 1 low-range server 3 to 5 k/year OPEX
  • Per unique visitor 1 /year
  • Per 1k web pages 0.5
  • margin cost 2 c OPEX, 3 c CAPEX
  • Cost breakdown OPEX
  • 15 shell core
  • 15 hardware
  • 15 bandwidth
  • 5 maintenance/support
  • 50 man power

19
Costs little  works  well
  • Home page Solaris
  • 12 Sun 4-CPU 6 dynamic 6 static content
  • 400 k/year OPEX (amort. included)
  • Additional CAPEX 10 per 1k daily pages
  • Home page Linux
  • 18 PC 12-CPU 12 dynamic 6 static content
  • 100 k/year OPEX (amort. included)
  • Additional CAPEX 3 per 1k daily pages? works
    well

20
Works well
  • Voila.fr ranked 1, Wanadoo.fr ranked 3
  • () Googles home page always 1 but thats not
    a real portal home page
  • Both fastest real portal home pages
  • Unavailability mostly due to a sub-service
  • freeware expensive fees maintce poor
    support
  • Open source proprietary code cant fix
  • This sub-service delivers less from more
  • )

21
Next Steps
  • Convince top management that Open-Source
    solutions can replace proprietary one!
  • Most agree on web front office (1-tier)
  • Most disagree on 2nd tier Jonas,Jboss /
    Websphere, 9iAS
  • All disagree on DB MySQL, PostgreSQL / Oracle,
    DB2, SQLs
  • Carefully select an appropriated database for
    your needs
  • Deploy Oracle on PC Linux attached disks
    (Instead of Sun Solaris SAN or NAS )

22
Thank you
Some Open Source Oriented Question ?
Pierre.Aubert_at_francetelecom.com Eric.Oliveri_at_franc
etelecom.com
Write a Comment
User Comments (0)
About PowerShow.com