Citation Context

...er they exhibit slower query time and/or preprocessing time and distributed memory parallelization looks more difficult since the query search space is less localized around source and target. Vetter =-=[12]-=- has developed a shared memory parallelization of the preprocessing stage that is also the basis for our distributed memory parallelization. While sequential CH techniques perform node 2 ordering onli...

Citation Context

...lete. To reduce the preprocessing time, we can parallelize it. We can process different parameter intervals after a split in parallel and additionally we can apply the parallelization techniques from =-=[25]-=-. Also, our parameter splitting is efficient but leaves room for further improvements. One could possibly merge intervals later when they seem to be very similar. This could save preprocessing time si...

Citation Context

...an Intel Core i7-920, clocked at 2.67 GHz with four cores and 12 GiB of RAM. It runs Linux kernel version 2.6.34 Our CH variant implements the shared-memory parallel preprocessing algorithm of Vetter =-=[15]-=- with a hop limit of 5 and 1 000 settled nodes for witness searches and 7 hops or 2 000 settled nodes during the actual contraction of nodes. The priority function is 2 ∗ edgeQuotient + 4 ∗ originalEd...

Citation Context

...n to compute the node priorities. Nodes with higher priority are more important and get contracted later. The nodes are contracted in parallel by computing independent node sets with a 2-neighborhood =-=[10]-=-. 6 Query The query computes a shortest path between two original edges, a source s→ s′ and a target t′ → t. It consists of two Dijkstra-like searches that settle original edges (cf. Section 2) one in...