Content-aware dispatching algorithms
The typical Web cluster architecture consists of replicated back-end
and Web servers, and a network Web switch that routes client requests
among the nodes. We have designed and developed new content-aware
scheduling policies, with an aim at improving load sharing
in Web clusters that provide multiple services such as static, dynamic
and secure information. In particular, we have proposed the
Client Aware Policy (CAP), which classifies the client
requests on the basis of their expected impact on main server
resources, that is, network interface, CPU, disk. At run-time,
CAP schedules client requests reaching the Web cluster with the
goal of sharing all classes of services among the server nodes.
Simulations and experimental results have shown that CAP is able
to outperform locality-aware assignment policies and least-loaded
schedulers. In particular, dispatching policies aiming to improve
locality in server caches give best results for Web publishing
sites providing static information and some simple database
searches. When we consider Web sites providing also dynamic
and secure services, CAP is more effective than state-of-the-art layer-7
Web switch policies.
Multi-tier dispatching algorithms
We have also compared the performance of several combinations
of centralized and distributed dispatching algorithms
working at the first and second layer of a multi-tier Web
cluster. Throughout the experiments, we evaluated the effectiveness
of several performance metrics, such as the number of established
TCP connections, the CPU usage, the disk usage. We have tried to
evaluate different load estimation strategies, such as computing
mean values of load samples, or introducing alarm mechanisms to
exclude overloaded nodes temporarily.
One of the main results of this analysis is the confirmation that
least-loaded assignments do not work well when the load metric is
subject to high variations and tends to become stale. In these cases,
load coefficients should be used to avoid the worst assignment,
not to choose the best one.
QoWS dispatching algorithms
In a world where many users rely on the Web for up-to-date personal
and business information and transactions, it is fundamental to build
Web systems that allow service providers to differentiate user
expectations with multi-class Service Level Agreements (SLAs).
Starting from the basic principles of network QoS, we
analyze how Quality of Web Services (QoWS) principles can be
realized in a Web site hosted on a Web-server cluster.
We propose a methodology to determine a set of
confident Service Level Agreements (SLAs) in a Web cluster for
multiple classes of users and services. Moreover, we implement at
the Web switch level and compare three QoWS-aware policies and
mechanisms that transform a best-effort Web cluster into a
QoWS-enhanced system. Our experimental results show that the policies
lacking some QoS principles can provide acceptable performance for
some load and system conditions, but they do not allow the system
to guarantee the contractual SLA targets for every expected load.
Moreover, we carry out a simulation analysis to explore
alternative architectures and policies for addressing other issues
such as system scalability and more severe SLA constraints.
All current dispatching algorithms are evaluated through a Web cluster
represented by a
Reverse Proxy, implemented through the
Apache Web Server
(version 2.0), using both the
mod_rewrite and the
modules. The target platform is the Linux
operating system. The proposed architecture allows for a very quick
setup and evaluation of dispatching algorithms on a Web cluster.
The policies which prove more stable and efficient are integrated
in our prototypes.
To see a list of people involved in the HiPerWeb project,
To see a list of publications related to dispatching
algorithms, click here.
Last updated: December 1, 2002