How Facebook Will Power Graph Search

Nice slashdot article: how facebook build compute resources as services to fine tune utilisation.

How Facebook Will Power Graph Search

[space height=”20″]

Facebook

  • 1 000 000 000 users
  • 230 000 000 photos a day
  • 220 billion photos stored
  • servers hot 14 to 16 hours a day

 

Infra

  • front-end web clusters: 12 000 servers
  • 20 to 40 servers a rack
  • web pages: 250 racks
  • cache: 30 racks
  • ads: 30 racks
  • maximize the utilization of the Web CPU

 

Racks of multi-feed servers = power the wall

  • multi-feed rack: got complete user history for user the last 2 days
  • two services:
    • leaf: contain all data -> in RAM
    • aggregator: compute -> CPU

 

Web serveur wall generation
  • get all friends
  • query all in parallel to leaf servers
  • agregrator rank and merge feed to keep top 40 stories

 

Five type of servers

  • web
  • database
  • hadoop
  • haystack photo
  • feed
[space height=”20″]
Limited set of servers -> easier setup -> maximise pricing
[space height=”20″]
Disaggregated rack
  • hardware as services == ram as service == cpu as service etc…
  • scale each resource independently

[space height=”20″]

Hardware services / pieces
  • Compute
  • RAM
  • Storage
  • Flash
[space height=”20″]
Graph Search
  • 320 cpu cores
  • 3 TB of RAM
  • 30 TB or Flash
  • ram flash ratio: 1:10 -> critical -> flexibility enable to fine tune this ratio !!
[space height=”20″]

Racks of multi-feed servers = power the wall

0 replies

Leave a Reply

Want to join the discussion?
Feel free to contribute!

Leave a Reply

Your email address will not be published. Required fields are marked *