The Internet Encyclopedia (Volume 3)

P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML

Web ̇QOS WL040/Bidgoli-Vol III-Ch-58 July 16, 2003 9:36 Char Count= 0

CONCLUSION ANDFUTURETRENDS 721

by fractional degradation and offers fine-grained control

of delay and server utilization.

PERFORMANCE CONSIDERATIONS

IN WEB PROXY SERVERS

Before this chapter concludes, a word on performance

considerations in proxy servers is in order. Proxy servers

are intermediaries between the clients and the accessed

Web servers. They are the main performance acceleration

mechanism on the Web, which makes their study very

important. QoS architectures should consider the effects

proxy servers have on user-perceived Web performance

and make use of them to satisfy client QoS requirements.

Proxies may be used for caching, transcoding, or con-

tent distribution. A proxy intercepts incoming requests

and attempts to serve them locally. If the requested con-

tent is not locally available the proxy may forward the

request to another server (e.g., content distribution prox-

ies), contact the origin server and save the response (Web

proxy caches), or contact the origin server, transcode the

response, and forward it to the client (transcoding proxy).

Although current proxy servers typically treat all clients

alike, there has been much talk on making them QoS-

aware. For example, the server may offer preferential

treatment to some classes of clients or classes of content.

To illustrate this point, consider a content distribu-

tion network composed of multiple proxy servers sit-

uated around the Internet backbone. The distribution

provider may make agreements with content providers to

distribute their content preferentially for a correspond-

ing fee. Alternatively, the distribution provider may make

agreements with certain ISPs to improve the quality of ser-

vice to their clients by virtue of the content distribution

network. An example of such a network is that introduced

by Akamai.

Several research efforts have looked at biased replace-

ment policies in proxy caches (Kelly, Chan, Jamin, &

Mackie-Mason, 1999). Such policies attempt to maximize

a weighted hit ratio, where weights are set in accordance

with content importance. For example, content fetched

from the preferred providers can have a higher weight

and therefore a lower likelihood of being replaced. An-

other research direction is to determine dynamically the

disk space allocation of a cache or a content distribution

proxy such that content of preferred providers receives a

higher hit ratio. In this approach, the “performance dis-

tance” between different content types can be controlled

(Lu, Saxena, & Abdelzaher, 2001). For example, one can

specify that preferred content is to receive twice the hit ra-

tio of regular content. The underlying adaptive disk allo-

cation policy uses feedback control to translate this speci-

fication into a dynamic disk space allocation that satisfies

the specified requirement in the presence of dynamically

changing load patterns.

In general, with the proliferation of caching proxies

and content distribution servers, it is becoming impera-

tive to integrate server solutions with proxy solutions in a

comprehensive end-to-end framework. Server and cache

cooperation can implement innovative mechanisms

for performance improvement. Generalization of such

architectures to the area of QoS is currently an active area

of research.

CONCLUSION AND FUTURE TRENDS

In this chapter, we briefly introduced the most important

issues and mechanisms for providing quality of service

in the modern Web architecture. The topic of providing

quality of service guarantees is becoming increasingly im-

portant with the pricing of Internet services, and with

the tendency to include performance requirements within

the contractual obligations of service providers. In the ag-

gregate, QoS-assurance mechanisms impose new, signif-

icantly different challenges on the research community,

calling for novel theoretical foundations for delivering

guarantees in an unpredictable environment.

At present, three underlying theoretical foundations

are identified, whose composition may help understand

the dynamics of the Web enough to provide the needed

assurances. These foundations are real-time scheduling

theory, queuing theory, and feedback control theory. Real-

time scheduling theory has traditionally been concerned

with developing scheduling policies and resource alloca-

tion disciplines that ensure satisfaction of timing guaran-

tees in closed embedded systems. In isolation, this theory

requires making strict assumptions regarding the input

load and available resource capacity, which may not be

possible in the open Web environment. Feedback control

theory, on the other hand, is very robust with respect to

assumptions on load and disturbances, but does not ex-

plicitly consider timing constraints. It applies a simple

trial and error method to correct performance gradually

until the desired specifications are met. Guarantees can be

made of the eventual convergence of this approach. Very

little a priori knowledge is needed regarding the system

and its load. Utilizing feedback control theory for Web

server QoS control is becoming an increasingly popular

research topic. Recent efforts investigate merging feed-

back control techniques with real-time scheduling the-

ory to produce dynamic resource allocation schemes with

provable convergence guarantees and predictable tempo-

ral behavior.

A disadvantage of using feedback control theory is that

it is primarily a reactive approach. No correction is made

to resource allocation until the performance deviates from

the desired levels. In an ideal world, a QoS provisioning

mechanism should predict likely future performance de-

viations before they occur and take preventive actions

a priori such that no deviations develop. Fortunately,

queuing theory offers a predictive framework that allows

reasoning about future temporal behavior from stochas-

tic properties of input load. Hence, combining the predic-

tive power of queuing theory with the corrective power

of feedback control may produce tighter guarantees on

the performance attributes of Web applications. Such a

combined approach bas recently been suggested and is

becoming an active topic of upcoming research.

Another interesting development in the Web architec-

ture is the rapid emergence of content distribution net-

works. Currently distributed architectures and protocols

for QoS-aware content distribution are still at their in-

fancy. The topic offers interesting challenges in balancing