Nikolaos Vasiliou and Hanan Lutfiyya

1
Differentiated Quality of Service for Web Servers
for Electronic Commerce Applications

• Nikolaos Vasiliou and Hanan Lutfiyya
• The University of Western Ontario
• London, Ontario, Canada

2
Introduction

• Consumers do not want to be kept waiting when
  browsing, selecting, or paying for products
  on-line.
• Consumers are annoyed by delays and failures.
• Consumers expect a high-level of quality of
  service and are sometimes willing to pay for it.

3
Architecture
client
Accept Request
Send Response
Normal Request Processing
Web Server
Read Request
Log Request
Deregister Request
Register Request
Request Complete
Scheduler Module
Registration
QoS Addition
Deregistration
Resource Manager
QoSBoard Module
4
Architecture

• A request is assigned to a process.
• Register Request occurs when a request arrives
  and before any processing begins.
• Registration suspends processing of the request
  until the Scheduler decides processing should be
  resumed.
• The request is put on a queue based on the
  categorization taking place (e.g., by type of
  request or by type of requestor). The queue is
  in the QoSBoard.

5
Scheduling Algorithm (Version 1)

• Two queues A and B.
• The Scheduler always runs class A requests and
  only runs class B requests when there are no
  class A requests.
• This can starve out class B requests if there is
  always a request in the A queue. Obviously, not
  a good solution.

6
Scheduling Algorithm (Version 2)

• Two queues A and B
• Two variables (also stored in QoSBoard)
• max_proc_A The maximum number of A processes
  allowed to run.
• max_proc_B The maximum number of B processes
  allowed to run.
• Similar to Version 1, except there is now a limit
  on the number of A processes allowed to run at
  any one time.

7
Scheduling Algorithm (Version 2)

• Our experimental results show that this works
  best when max_proc_B is relatively small.
• However, by keeping max_proc_B small (and
  static) means that we may waste computing
  resources if there arent that many A requests.
• These resources could have gone to B requests.

8
Scheduling Algorithm (Version 3)

• The difference between this algorithm and the
  second is that the scheduler doesnt distinguish
  between A processes and B processes unless an A
  process complains that it is waiting too long
• If the number of complaints is equal to
  max_complaints then max_proc_B is reduced.
• The variable representing the number of
  complaints is set to 0. If it again exceeds
  max_complaints, max_proc_B is reduced.

9
Scheduling Algorithm (Version 3)

• max_proc_A and max_proc_B are initially set to
  the same value.
• As long as A processes arent complaining then B
  processes can be run.
• This is a more effective use of resources.

10
Implementation

• We use the Apache Web Server
• The QoSmodule is implemented as shared memory
  between request processes and the Scheduler
  (which is a separate process).

11
Results and Future Research

• Initial results are promising, but they show that
  there are many factors to take into account.
• We are especially interested in architecting out
  a solution that makes it easy for developers to
  implement different scheduling algorithms
  tailored for their sites.
• We will also look at other ways of
  differentiating between users.
• Example If a user is idle for more than n time
  units then the process associated with the
  request is given a lower priority.