Performance Evaluation of Improved Web Search Algorithms

In this paper we propose an evaluation method for parallel algorithms that can be used independently of the used parallel programming library and architecture. We propose to predict the execution costs using a simple but efficient framework that consists in modeling the strategies via a BSP architecture, and estimating the real costs using as input real query traces over real or stochastically generated data. In particular we apply this method on a 2D inverted file index used to resolve web search queries. We present results for OR queries, for which we compare different ranking and caching strategies, and show how our framework works. In addition, we present and evaluate intelligent ranking and caching algorithms for AND queries.