Abstract

Real time search - a setting in which Web search engines are able to include among their query results documents published on the Web in the very recent past - is a clear evidence that many of the off-line computations performed so far on conventional search engines need to be moved to the on-line arena. This is a demanding case for parallel computing since it is necessary to cope efficiently with thousands of concurrent read and write operations per unit time, all requiring latency times within a fraction of a second. To our knowledge, computations related to capturing user preferences through their clicks on the query result webpages and include this feature in the document ranking process are currently performed in an off-line manner. This is effected by pre-processing very large logs containing millions of queries submitted by actual users in a time scale of days, weeks or even months. The outcome is score data for the set of documents indexed by the search engine which were selected by users in the past. This paper studies the efficiency of this process in the on-line setting by evaluating a set of strategies for concurrent read/write operations executed on a multi-threaded multi-core architecture. The benefit of efficient on-line processing of user clicks is making it feasible to include user preference in document ranking also in a real-time fashion. © 2011 Springer-Verlag Berlin Heidelberg.