Abstract

Large Web search engines are complex systems that solve thousands of user queries per second on clusters of dedicated distributed memory processors. Processing each query involves executing a number of operations to get the answer presented to the user. The most expensive operation in running time is the calculation of the top-k documents that best match each query. In this paper we propose the parallelization of a state of the art document ranking algorithm called Block-Max WAND. We propose a 2-steps parallelization of the WAND algorithm in order to reduce inter-processor communication and running time cost. Multi-threading tailored to Block-MaxWAND is also proposed to exploit multi-core parallelism in each processor. The experimental results show that the proposed parallelization reduces execution time significantly as compared against current approaches used in search engines.