The Simulation Evaluation of Heat Balancing Strategies for Btree Index over Parallel Shared Nothing Machines

The Simulation Evaluation of Heat Balancing Strategies for Btree Index over Parallel Shared Nothing Machines Hisham Feelifl and Masaru Kitsuregawa Institute of Industrial Science, The University of Tokyo {hisham, kitsure} @tkl.iis.u-tokyo.ac.jp In shared nothing machines the data are typically declustered across the system processing elements (PEs) to exploit the I/O bandwidth of the PEs. Notwithstanding, the access pattern is inherently dynamic; therefore, the system performance may degrade as some PEs become hot spot. Consequently reorganization for heat balancing is essential and should be online as well. Meanwhile, in order to achieve efficient query and transaction evaluation, data at each PE are indexed based on the Btree structures. Thus, data reorganization should satisfactorily deal with the index modification as moving data from “hot spot” PEs to cold PEs. In this paper, we propose online heat balancing strategies for Btree index over parallel shared nothing machines based on the principle of distributing the given heat as evenly as possible across the system PEs. Furthermore, the proposed strategies have the capability to reduce the instantaneous migration cost during the heat balancing process, which in turn avoid the harmful migrations that may degrade the system performance. We evaluate the performance of the proposed strategies in comparison with the well-known strategies. The conducted simulation shows their efficiency in correcting the system performance degradation.