Implementing a Loadbalanced Web Server System

1
Implementing a Load-balanced Web Server System
2
Architecture of A Cluster-based Web System
Courtesy IBM Research Report, The state of the
art in the locally distributed Web Server systems.
3
Architecture of Our Web Server Cluster
File Service Database Service
http requests
Web Server 1
Load Distributor (grid1.cs.ucr.edu)
File Service Video-On-Demand
http requests
Web Server 2
4
Our Web Server Cluster

• The whole web server only provides one visible
  web address to the outside world.
• Each Web Server is able to provide two kinds of
  web services.
• The load distributor distributes the incoming
  requests among the servers according to either
  content-aware or content-unaware load balancing
  strategies.

5
Tasks to do to set up the system

• Building up the web services on the servers
• File Services
• Video-on-demand Services
• Database Services
• Implementing the load distributor on the
  frond-end node
• Content aware request distribution
• Content unaware request distribution

6
Building Up the Web Services

• File service
• Built on top of Apache server.
• File set is generated by SPECWEB99.
• Video-On-Demand
• Real MPEG2 movies are stored in a specific
  directory on the Apache server.
• Client video streaming software (VideoLan) is
  installed and automatically launched by the
  Apache server.
• Database service
• Built on top of Apache and MySQL.

7
Video On Demand Service

• VideoLAN project (Open Source Media Streaming
  Solution)
• Targeting multimedia streaming of MPEG-1, MPEG-2,
  MPEG-4 and DivX files, DVDs, digital satellite
  channels, digital terrestial television channels
  and live videos on a high-bandwidth IPv4 or IPv6
  network in unicast or multicast.
• Client-server Architecture
• Server streams MPEG-1, MPEG-2 and MPEG-4 / DivX
  files, DVDs and live videos on the network in
  unicast or multicast.
• Client receives, decodes and displays MPEG
  stream.

8
VideoLan System
9
Building Up Video-On-Demand Service in Our Web
Server

• VideoLan client-server software is installed
• Server can stream movies to the client in
  realtime through UDP/RTP or HTTP/TCP
• For video-on-demand service using HTTP/TCP, only
  the client is needed. The client software (vlc)
  is automatically launched once the Apache server
  detects that it is a video file.

10
Load Balancing Schemes

• Content Unaware Scheme
• Choose a server before receiving the URL request
• Round Robin
• Content Aware Schemes
• Choose a server to dispatch a request after
  receiving and looking at the URL request
• Balance load according to different URL request
• For database service Database Server
• For video-on-demand service Multimedia Server
• For file service Round Robin

11
Implementing the Load Distributor

• Install the TCPSP
• The tcp splicing is a technique to splice two
  connections inside the kernel, so that data
  relaying between the two connections can be run
  at near router speeds.
• Write the Distributor program in C language
• Two load balancing strategies are implemented
• The installed kernel module TCPSP is invoked to
  perform TCP splicing
• Run the distributor program in the application
  level

12
Flow Chart of the Load Distributor(content aware)
Distributor
Child Process
13
Flow Chart of the Load Distributor(content
unaware)
Distributor
Child Process
Listen for incoming connections on port 8888
Accept the connection
Choose a server according to the load balancing
scheme
Create a child process to do further processing
14
Comparison with Gage

• Gage A QoS Aware Web Server System
• Performance Guarantees for Cluster-Based
  Internet Services, Chang Li, State University
  of New York at Stony Brook.
• The load distributor is implemented as a kernel
  module. It is faster but can only implement
  content-unaware load balancing.
• Gage doesnt provide a variety of web services.

15
Planned Performance Measurement

• Let all servers provide file service, use
  SPECWEB99 to test the performance of the
  cluster-based file server.
• Compare the time taken to service a Database
  request through the load distributor with that
  without the load distributor.
• Compare the time taken to service a Database
  request through the load distributor with that
  without the load distributor.

16
SPECWEB99
17
Lets go to the lab to see DEMO!