Online Heat Balancing for Parallel Indexed Database On Shared Nothing System

Shared nothing systems offer a tremendous processing capacity. In such highly parallel systems the data are typically declustered across the system processing elements (PEs) to exploit the I/O bandwidth of the PEs. However, the access pattern is inherently dynamic, which in turn can lead to performance degradation as some PEs become hot spot (bottleneck). Therefore, reorganization for heat (load) balancing is essential and should be online. Our objective is to migrate the data from the hot PEs to the cold PEs with minimal cost of modifying the index structure of the system, so that the system is heat balancing and consequently higher performance could be achieved with minimal reorganization cost. In this paper, we propose an online heat balancing strategy for parallel indexed database on shared nothing system, in which the data migration process itself is based on the heat statistics generated by the access pattern that may be directed to the distributed index structure of the system. The proposed strategy captures the intuitive goal of distributing the given heat across the system PEs as evenly as possible and the result demonstrates that it is efficient in correcting any degradation in the system performance.