A Parallel Implementation of the K Nearest Neighbours Classifier in Three Levels: Threads, MPI Processes and the Grid

Abstract

The work described in this paper tackles the problem of data mining and classification of large amounts of data using the K nearest neighbours classifier (KNN) [1]. The large computing demand of this process is solved with a parallel computing implementation specially designed to work in Grid environments of multiprocessor computer farms. The different parallel computing approaches (intra-node, inter-node and inter-organisations) are not sufficient by themselves to face the computing demand of such a big problem. Instead of using parallel techniques separately, we propose to combine the three of them considering the parallelism grain of the different parts of the problem. The main purpose is to complete a 1 month-CPU job in a few hours. The technologies that are being used are the EGEE Grid Computing Infrastructure running the Large Hadron Collider Computing Grid (LCG 2.6) middleware [3], MPI [4] [5] and POSIX [6] threads. Finally, we compare the results obtained with the most popular and used tools to understand the importance of this strategy.

Topics: Grid, Parallel Computing, Threads and Data Mining.

The authors wish to thank the financial support received from the Spanish Ministry of Science and Technology to develop the GRID-IT project (TIC2003-01318).