Transcript Load Distribution among Replicated Web Servers

Load Distribution among Replicated Web
Servers:
A QoS-based Approach
Marco Conti, Enrico Gregori, Fabio Panzieri
WISP99
2000.9.14
KAIST EECSD CALab
Hwang In-Chul
Contents





Introduction
Load Distribution Strategies
A QoS-Based Architecture
Work in Progress
Critique
2/15
Introduction(1/2)

A practical approach to the provision of web
services
– Replicate Web servers(WSs) at distinct sites
– Each client select the “most convenient” WS replica

The success of this approach
– To bind dynamically a client to the most convenient
replica
– To maintain data consistency among the WS replicas
3/15
Introduction(2/2)

In this paper
– Load distribution strategy
• Mirror-based strategy
• DNS-based strategy
• QoS-based strategy
– To minimize the URT(User Response Time)
4/15
Load Distribution Strategies

Mirror-based strategy
– The user manually selects a replica

DNS-based strategy
– “Ideal” round-robin assignment of clients to WS
replicas

QoS-based strategy
– DNS : all addresses of replica WSs
– Browser selects a replica with satisfactory URT by
sending probe
5/15
Load Distribution Strategies
- Performance Comparison

Simulation scenario
Area 1
Area 2
Web Server
replica 1
Web Server
replica 2
Area 1
Network
Delay
Area 2
Network
Delay
Area 4
Network
Delay
Area 3
Network
Delay
Web Server
replica 4
Web Server
replica 3
Browsers
Area 4
Area 3
Internet
Inter area network transfer delay
Intra area network transfer delay
Access line to a Web Server
Simulation scenario
6/15
Load Distribution Strategies
- Performance Comparison

Simulation environment
– Network delay model
• Intra-area delays
– The minimum area round trip time
– The queuing delays in the area router
– The packet transmission time
• Inter-area delays
– Random variables
– Other factors
Consecutive query
Independent and exponential distributed
Each query
Access a geometrically distributed number of pages
Web page size
Avg. 3000 bytes
Dummy req.
1000 bytes
Server capacity
200 request per second(FIFO queue)
7/15
Load Distribution Strategies
- Performance Comparison

Impact of intra-area network congestion
Area 1 Routers
Other Areas Routers
0.98 Util.
Max. 0.8 Util.
– Results
• Utilization of each replica
– QoS-based strategy : (0.58, 0.91, 0.92, 0.92)
– Other strategies : uniformly 0.80
8/15
Load Distribution Strategies
- Performance Comparison

A heavily loaded area
Area 1 User-Query Generation 0.98 of Server Capacity
Other Areas
0.8 of ServerCapacity
– Results
9/15
Load Distribution Strategies
- Performance Comparison

Symmetric case
– All Areas
• The most congested router : 0.80 utilization
• The user-query generation rate : 0.80 of server capacity
– Results
10/15
Load Distribution Strategies
- Performance Comparison

A realistic scenario
– Four distinct areas
• USA, Europe, Asia, Australia
– Daily different loads in different periods of time
– Results
11/15
A QoS-Based Architecture


Do not require modification of any software
Architecture
12/15
A QoS-Based Architecture
DNS
DNS
DNS Request
DNS Request
All Replica’s
IP Address
Replicated
Server 1
One Replica’s
IP Address
Replicated
Server 1
Replicated
Server 2
Probe Reply

Replicated
Server N
Browser
Replicated
Server 2
...
Probe Request
...
Browser
Broadcast Poll Request
Poll Reply
Replicated
Server N
Drawback
– URT estimation : Single measure
– Polling overhead
13/15
Work in Progress

Load Distribution(LD) service
–
–
–
–
To overcome the main limitations
Responsible for distributing the browsers’ requests
Maintain for each WS replica the WS response time
Continuous monitoring of the response time
14/15
Critiques

Contribution in this paper
– QoS-based approach: Minimize URT
– Load distribution considering network delay



Simulation with realistic workload
Not Scalable
More research on LD
– How to evaluate the accurate WS response time
15/15
16/15