A Study on Data Placement of Extensible Parallel Storage System

For a parallel file system which employs modulo function as their data placement mechanism, a rebalancing operation must be taken after new storage nodes are added in it. The rebalancing operation moves data among system nodes for efficiently using of system resource. However, most data moves are between old nodes and useless to achieving storage balance, and therefore waste plenty of CPU time and I/O bandwidth. In this paper, two data placement mechanisms, ordinal-selection and improved ordinal-selection, are presented to settle this problem, both of the mechanisms place data according to the simulation results of multi times rebalancing operations and make the storage system more flexible to changes in number of system nodes. The analysis and simulation results show that: both of the mechanisms can achieve zero-ineffectual move ratio. As a tradeoff, their parallel degrees are lower than that of modulo function, and the improved ordinal-selection is little better than ordinal-selection.