Performance improvement of web caching in Web 2.0 via knowledge discovery

Highlights

• Improvement of web caching in Web 2.0 by fragmenting the contents of the pages.

• Definition of a framework to adapt fragment designs in content aggregated web pages.

• The adaptation is done by the use of decision trees obtained using knowledge discovery.

• Our solution shows shorter latencies than the traditional web cache schemes.

• The overhead generated by the new approach is low enough to validate the solution.

Abstract

Web 2.0 systems are more unpredictable and customizable than traditional web applications. This causes that performance techniques, such as web caching, limit their improvements. Our study was based on the hypotheses that the use of web caching in Web 2.0 applications, particularly in content aggregation systems, can be improved by adapting the content fragment designs. We proposed to base this adaptation on the analysis of the characterization parameters of the content elements and on the creation of a classification algorithm. This algorithm was deployed with decision trees, created in an off-line knowledge discovery process. We also defined a framework to create and adapt fragments of the web documents to reduce the user-perceived latency in web caches. The experiment results showed that our solution had a remarkable reduction in the user-perceived latency even losses in the cache hit ratios and in the overhead generated on the system, in comparison with other web cache schemes.

Introduction

Content aggregation systems (CAS) are Web 2.0 applications in which users are able to create their own web pages by the aggregation of contents. The web pages retrieve content from distributed sources and assemble it in a single web page. These applications differ from other content management systems (CMS) in that users do not create the content, they only set up the web pages by aggregating content from public services and remote sources (web services, RSS, etc.). The web application retrieves independent content from these sources (content elements, CEs).

This type of web pages has high update rates because the content is retrieved from different sources and the web page needs to be updated every time a single source changes. They have also a very high customizable degree.

There are many examples of web applications that fit in CAS: social networks, web blogs, feed aggregation tools, etc. We focused our study on personal start-pages (Yahoo! Pipes, iGoogle, Netvibes and PageFlakes).

Depending on the system tier where the assemblies of the content elements take place, the cache is able to manage and to store whole and indivisible web pages, or otherwise, content elements. When the assembling process is done in the web proxy cache, the cache is able to store content elements (CEs) independently. This benefits the hit ratio of the web cache (only the invalidated content elements are requested to the web server), but it worsens the overhead times of assembling all the content elements. On the other hand, when the assemblies take place in the web application server, if one content element is invalidated, the whole web page is also invalidated and the cache needs to request to the server all the contents of the web page. This reduces the hit ratio, but only one server request is generated so the overhead times corresponding to server connections and assemblies are shorter.

Our research was focused on creating intermediate schemes in which some of the content elements were assembled in the application server and other ones in the web proxy cache. Thus, the increases of the hit ratios with the losses in the overhead times were balanced, obtaining shorter user-perceived latencies. Our hypothesis was that this adaptation can be done with decision trees created in a previous and off-line knowledge discovery process. We suggested to obtain this knowledge using the characteristics of the contents and the performance results of an emulation of synthetic content models. The contributions of this research work are: (a) Definition of a framework to adapt the content fragment elements of a web page to reduce the user-perceived latency in content aggregation systems. (b) Deployment of the adaptive core of the framework using knowledge discovery techniques by mining performance data obtained in an off-line process using synthetic content models. (c) Experimental validation of the proposed framework and the use of knowledge discovery in a real system with contents of a real website.