Optimizing large knowledge networks in spatial computers

This paper presents a novel concept of a Spatially Aware Graph Store, which realizes a Graph Store on top of a spatial computer architecture to manage graphs in one, two or three physical dimensions. In this environment, the physical distance between graph nodes strongly affects graph traversal performance. Consequently, a Spatially Aware Graph Store needs to minimize these distances to operate efficiently. We show that this minimization can be achieved in two ways. First, by increasing the dimensionality of the spatial computer and second by applying optimization methods. For the latter, this work introduces a novel Mid Point Optimization method to quickly optimize large real-world knowledge networks by rearranging nodes in a way that distances between linked nodes are reduced. In addition, a Local Optimization method is subsequently applied to refine the result. Finally, the Node Decomposition method is presented that splits nodes with many edges into several smaller nodes to achieve a further reduction of distances between linked nodes.Our results show that the overall distances between nodes can be reduced by three orders of magnitude for 3D in comparison to one-dimensional (1D) Spatially Aware Graph Stores. The suggested Mid Point Optimization method achieves a reduction by another order of magnitude. In a 3D spatial computer, Local Optimization is capable of reducing distances by another 20%. However, in 1D and 2D spatial computers it becomes a prohibitive time consuming method. Finally, the Node Decomposition enables an additional distance reduction by 40% in Scale Free Graph Data sets.