facebook architecture for 600M users

Architecture for
(over) 600M Users

’2011 1st Issue of TechLab Report
Neowiz Games. TechLab
http://rainblue.kr
kikiki.blue@gmail.com
2

4
출처 : http://blog.fliptop.com/facebook-nation/

Facebook Users in Korea
Total Facebook Users: 4 010 980 Penetration of population: 8.25%

Position in the list: 27. Penetration of online pop.: 10.17%

Average CPC(Cost per Click): $0.21 Average CPM(Cost Per Mile): $0.07

출처 : http://www.socialbakers.com
5

Social Network Users in Korea
Cyworld - 2010년 11월을 기점으로 traffic / unique visitor도 감소중

Facebook - 2011년 1월, 740만의 unique visitors, 지속적인
상승곡선

6
출처 : http://iblur.tistory.com/328

7
http://www.socialbakers.com/facebook-applications/category/8-games

"High Performance at Massive Scale – Lessons learned at Facebook‖

Jeff Rothschild - Vice President of Technology Facebook
Calit2 Auditorium - University of California, San Diego

http://cns.ucsd.edu/lecturearchive09.shtml#Roth 8

Web 2.0 Summit 09:, ―High Order Bit: The Infrastructure of Facebook‖

Mike Schroepfer - Vice President of Technology Facebook

http://www.youtube.com/watch?v=iddTbLo5s1M 9
http://channy.creation.net/blog/759

출처 : http://blog.daum.net/amorfati2005/53 11

• 500 million active users
• 100 billion hits per day
• 50 billion photos
• 2 trillion objects cached, with hundreds of millions of requests per
second
• 130TB of logs every day

2010 / 07 / 22

12

• 전통적인 웹사이트에서는
장비를 추가하고 DB를 나누어서 해결
• 하지만 Facebook은 서로 연결된
데이터 구조이기에
전통적 방식으로는 불가능하다

 Users spend 8 billion minutes online everyday using Facebook
 There are some 2 billion pieces of content shared every week
on the service
 Users upload 2 billion photos each month
 There are over 20 billion photos now on Facebook
 During peak times,
Facebook serves 1.2 million photos a second
 Yesterday alone, Facebook served 5 billion API calls
 There are 1.2 million users for every engineer at Facebook

2009년 기준 !!

13

File Storage at facebook
haystack: efficient storage of billions of photos

Facebook에는 약 20빌리얶(200억)장의 사짂이 보존되고 있으며 각각 4 종류의
해상도로 보존되고 있기 때문에 전체 800억장의 사짂이 보존되고 있다고 말할 수
있다. 이것은 사짂을 전부 이으면 지구를 10회 둘러 쌀 만한 면적이 있다.

15

Traditional Filesystem (Posix Standard) Metadata
=> It’s Too Many & Big. We couldn’t cache it at all !!

• File length
• Device ID
• Storage block pointers
• File owner
• Group owner
• Access rights on each assignment: read, write execute
• Change time
• Modification time
• Last access time
• Reference counts

16


17


18


Haystack : a generic HTTP-based object store containing needles that map to stored opaque objects
• HTTP server : simple evhttp server provided with the open source libevent
library

• Photo Store Server : responsible for accepting HTTP requests and translating
them to the corresponding Haystack store operations (Load all meta data on
Memory)

• Haystack Object Store : Index(meta data) + Data

• Filesystem(XFS) : Haystack object stores are implemented on top of files stored
in a single filesystem created on top of the 10TB volume.

• Storage(Blade server) : 2 x quad-core CPUs, 16GB – 32GB memory, hardware
raid controller with 256MB – 512MB of NVRAM cache, 12+ 1TB SATA drives

19


20


But It’s not opensource yet !!
Pomegranate - Storing Billions And Billions Of Tiny Little Files
Hadoop Archive: File Compaction for HDFS

21

Web Tier at facebook

23


24


25

• HipHop is a source code transformer which transforms PHP source code into
highly optimized C++ and then compiles it using g++.

• It will be useful to companies running very large PHP infrastructures who do not
wish to rewrite complex logic within C or C++

• 50% reduction in CPU usage than Apache + PHP

• Facebook’s API tier can serve twice the traffic using 30% less CPU

• It has embedded simple webserver on top of libevent.
90% of Apaches disappear in history.

Reference
 Drupal
 MediaWiki
 phpBB
 WordPress (WordPress has become 2.7x faster)
26


 Almost-Serialization-Free APC
 Faster Serialization and JSON Encoding
 Less Reference Counting
 More Compact Binary Code
 Improved Memory Allocation(jemalloc)

Throughput : the average rate of Web requests successfully served over a given period of time

27

Web Tier in facebook

Tornado is a relatively simple, non-
blocking web server framework written
in Python.

It is designed to handle thousands of
simultaneous connections, making it ideal
for real-time Web services.

28


Node's goal is to provide an easy way to build scalable network programs.

Node is similar in design to and influenced by systems like Ruby's Event
Machine or Python's Twisted. Node takes the event model a bit further—it
presents the event loop as a language construct instead of as a library

29

Chat Service in facebook
Real-time presence notification is biggest challenge. Not sending messages

31

Chat Service at facebook
Real-time presence notification is biggest challenge. Not sending messages

Each Facebook Chat user now needs to be notified whenever one of his/her friends
(a) takes an action such as sending a chat message or loads a Facebook page (if
tracking idleness via a last-active timestamp) or
(b) transitions between idleness states (if representing idleness as a state machine
with states like "idle-for-1-minute", "idle-for-2-minutes", "idle-for-5-minutes", "idle-
for-10-minutes", etc.).
Note that approach (a) changes the sending a chat message / loading a Facebook
page from a one-to-one communication into a multicast to all online friends, while
approach (b) ensures that users who are neither chatting nor browsing Facebook
are nonetheless generating server load.

http://www.facebook.com/note.php?note_id=14218138919
http://www.facebook.com/note.php?note_id=51412338919
32


Facebook Chatting Sub-Systems

For Facebook Chat, we rolled our own subsystem for logging chat messages (in
C++) as well as an epoll-driven web server (in Erlang) that holds online
users' conversations in-memory and serves the long-polled HTTP requests. Both
subsystems are clustered and partitioned for reliability and efficient failover. Why
Erlang? In short, because the problem domain fits Erlang like a glove. Erlang is a
functional concurrency-oriented language with extremely low-weight
user-space "processes", share-nothing message-passing semantics,
built-in distribution, and a "crash and recover" philosophy proven by two
decades of deployment on large soft-realtime production systems.

http://www.facebook.com/note.php?note_id=14218138919&id=9445547199&index=2


Erlang? Why?

• 멀티코어에 적합한 프로그래밍 모델
• CPU가 늘어날수록 속도도 빨라짂다.

• 함수형 얶어

• Light weight process(경량 프로세스)

• 변하지 않는 변수

• 속도와 과부하 문제에 대한 탁월한 해결력

• 무정지(Fault-tolerant) 애플리케이션

출처 : 2008-2009 Neowiz Devlelopers’ Day


1,000
erlang
java
메시지 1건 전송 소요시간(micro second)

C#

100

10

> 200,000
processes

1
10 100 1,000 10,000 100,000
프로세스 수

Source:

출처 : 2008-2009 Neowiz Devlelopers’ Day
Joe Armstrong
SICS

Another Challenge, Communicate with other parts of system

Thrift is a software framework for scalable cross-language services development. It
combines a software stack with a code generation engine to build services that work
efficiently and seamlessly between C++, Java, Python, PHP, Ruby, Erlang, Perl, Haskell, C#,
Cocoa, JavaScript, Node.js, Smalltalk, and OCaml.

36

Chat Service Trio’s Challenge
9 months after launching Service
1. Erlang Channel Server
1. String problem of Erlang makes Memory footprint big.
2. Garbage collecting when wait for new message.
3. We could monitor channel server at runtime.
1. debug & load new code on the fly => Erlang’s feature

2. C++ Logger
1. Memory Fragmentation

3. Presence Server
1. Update user info from channel server data
1. It’s far simpler than the Thrift example service bundled.
2. Collecting user presence to Presence server
3. Use zlib for compressing data to solve network bandwidth problem

4. Facebook Chat Now Available Everywhere
1. They start supporting Jabber(XMPP)
37

Social Messaging Service in facebook

over 350 million users sending over 15 billion messages per month
http://blog.facebook.com/blog.php?post=452288242130
http://blog.hankyung.com/kim215/blog/3653251
38

The Underlying Technology of Messages
Eventual Consistency Performance

Haystack

http://www.facebook.com/notes/facebook-engineering/the-underlying-technology-of-messages/454991608919
http://zookeeper.apache.org/ 39
http://hbase.apache.org/

MySQL at facebook
We use MySQL as a Simple Data Storage(Key-value).
It’s a good RDB, but We do not use that as RDB.
• Logical Migration은 어렵다.
• 많은수의 Logical DB를 생성하고, 다양한 물리노드로 load balancing을 한다
• Scale at Web-Tier
• Business 로직 처리는 Front-end 에서
• No-Join, Compute on Web-tier
• Non-Static data 는 DB에 두지마라
• 많이 참조되는 Static Data는 DB에 두지마라 (Cache하라)
• Service나 Memcache를 이용해서 Global Query를 하라(?)

41

Memcached & MySQL at facebook

42

Memcached & MySQL at facebook

Global Load Balancing(IP기반)

미국 서부와 아시아권 유럽

43
http://bcho.tistory.com/416

Memcached at facebook

1. Per-connection memory buffer
1. Facebook implemented per-thread shared connection buffer pool for TCP/UDP
2. Choose UDP for reduce network traffic / Multi-get
1. UDP Socket Lock ( Linux kernel use single socket lock for multiple thread )
1. separate UDP sockets for transmitting replies
3. On Linux, network interrupt is delivered to one cores => all cores receive soft
interrupt. And some NIC card has high rate of interrupts
1. Solve that combination of interrupt driven and polling driven network I/O
2. Now every core do their work!!
4. Memcached’s stat use global lock
1. moving stats collection per-thread and aggregating results on-demand
5. contention on the lock that protects each network device’s transmit queue
1. changed the dequeue algorithm
2. scale memcached to 8 threads on an 8-core system

https://github.com/facebook/memcached 44

Memcached at facebook

1. 800대의 Memcached 운용중 (2008년 12월)

2. 28 terabytes of memory (2008년 12월)
1. 초당 12억건의 Request 처리중

3. Original memcached can serve 50,000 UDP requests/s
1. 200,000 UDP requests/s with 173 microseconds.
2. 300,000 UDP requests/s with high latency.

https://github.com/facebook/memcached 45

Data Warehousing at facebook

Data: How much?

• 200GB per day in March 2008
• 2+TB(compressed) raw data per day in April 2009
• 4+TB(compressed) raw data per day today ( 2009, 11 )
47


Scribe :
A Thrift service for distributed logfile collection. Scribe was designed
to run as a daemon process on every node in your data center and to
forward log files from any process running on that machine back to a
central pool of aggregators. Because of its ubiquity, a major design point
was to make Scribe consume as little CPU as possible.

Scribe is a server for aggregating streaming log data. It is designed
to scale to a very large number of nodes and be robust to network and
node failures. There is a scribe server running on every node in the system,
configured to aggregate messages and send them to a central scribe
server (or servers) in larger groups. If the central scribe server isn’t
available the local scribe server writes the messages to a file on
local disk and sends them when the central server recovers. The
central scribe server(s) can write the messages to the files that are their
final destination, typically on an nfs filer or a distributed filesystem, or send
them to another layer of scribe servers.
48


49


50


51


52


53

Culture of facebook

55

Culture of facebook

Hackathon

56
http://mimul.com/pebble/default/2010/05/23/1274600280000.html

Culture of facebook
1. 개발 과정에서 Product Manager의 권한은 미약하고, 오히려 개발자들이 자기 프로젝트에 관심을
가지도록 하기 위해 개발자들을 설득, 로비하는 일이 잦다.

2. 모든 엔지니어는 입사 후 4주~6주 과정의 부트 캠프를 거친다. 부트 캠프에서 페이스북 시스템의
버그를 직접 수정하면서 배우고, 시니어 엔지니어들의 강의를 듣는다. 약 10% 정도의 엔지니어들이 이
과정을 완료하지 못하며 이 경우 권고 사직 대상이 된다.

3. 엔지니어는 백엔드부터 프론트엔드까지 보통 혼자서 다 구현을 한다. 클라이얶트, UI 일은 기술적
난이도가 낮고 반복이 많아 엔지니어들이 기피한다. 뉴스피드나 광고 알고리즘, memcache 최적화 등
백엔드 구현을 선호한다.

3. 모든 변경된 코드는 의무적으로 내부 리뷰를 거친다. 주커버그는 뉴스피드 관렦 코드는 직접 리뷰를
한다. 최소 1명 이상의 리뷰를 거치지 않은 코드를 릴리스해서는 안된다.

4. 페이스북에는 전문 QA가 없다. 모든 개발자가 테스트, 버그 수정, 유지보수까지 자기 코드에 대해
책임짂다. 내부 개발자가 사용해보는 테스트는 있다. 서비스 런칭을 위해 거쳐야 하는 자동화된 테스트도
있다. 하지만 기본적으로는 개발자가 UI 구현부터 테스트까지 모두 책임을 지는 구조이다.
"대부붂의 엔지니어들은 버그 없는 코드를 작성할 줄 안다"는 믿음도 일부 있는 붂위기.

5. 매주 화요일 릴리스하며 그 주중에 릴리스 대상 코드를 커밋한 개발자는 릴리스 시작 시 on-site(특정
IRC 채널)에서 대기해야 한다. 그렇지 않으면 공개 창피당한다. 9레벨 릴리스 중 문제가 생기면 개발자가
수정 후 1레벨부터 다시 시작한다.

6. SVN을 통해 혼나거나 프로젝트를 자주 지연시킨 엔지니어는 해고된다.(매우 뛰어난 퍼포먼스만 허용.
보통 채용 6개월 이내에 해고) 버그나 사이트 다운, 실수 등의 문제로 혼난다고 해서 해고되지는 않는다.
다만 이런 공개 창피를 통해 이슈를 해결하기 위해 매우 열중하며 이 과정에서 모두가 함께 배운다.

57
http://logonjava.blogspot.com/2011/01/facebook-sw.html

Culture of facebook

Commercial Software In-House Product

Opensource Product

58

facebook architecture for 600M users

Recommended

More Related Content

What's hot (20)

Viewers also liked (8)

Similar to facebook architecture for 600M users (20)

Recently uploaded (20)

facebook architecture for 600M users