MapReduce algorithms for the K group nearest-neighbor query

Algorithms for processing closest-pairs and nearest-neighbors queries on big spatial data in parallel and distributed frameworks

“…Contrary to previous methods that are all based on centralized systems, in [52] we pro posed the first MapReduce algorithm to effectively process the GKNN query in a parallel and distributed environment. Utilizing ideas and elements from previous work (query defini tion and heuristics [60], processing without indexes and PS heuristics [65,67], repartitioning data [24,27]), in this paper we present an algorithm consisting of seven phases, local or…”

“…The algorithm presentation in this section is a unified approach of the ones presented in [52] and [53] and their differences will be noted when necessary. The algorithm modification in [55] will be presented separately, because it is quite different.…”

“…In Phase 1 we have counted the points within each cell (Figure 5.9), so we can use this result. In [52] we used all the cells that intersect with the Query MBR (Figure 5.10). It is obvious that the GKNN candidates will most likely be found inside or closely around the MBR than far away from it, while the farthest cells will probably be pruned in Phase 3.…”

“…Extensive experimentation on the algorithm [52] has shown that the subsequent Phase 2, was by far the hardest one to complete. This happens because the use of MBR yields a rather large number of intersecting cells that results to too many calculations.…”

“…The algorithm was first presented in [52] and was implemented in the Apache Hadoop framework. Later in [53] an improved version of the algorithm was presented, along with a SpatialHadoop implementation and a study of what happens internally during the operation of the algorithm, using specialized cluster wide metrics.…”