Combined with the application of circuit partitioning, the hyper-graph model of ISPD98 benchmark is built on the weighted hyper-graph and the formal description of weighted hyper-graph partitioning problem is presented. The weighted hyper-graph partitioning algorithms based on the multi-level method is described. We also propose the matching strategy based on the core property of vertex which improves previous matching strategies based on the local information of vertex. Furthermore, we extend the core notion of graph to the hyper-graph, propose the core notion of hyper-graph and give its formal description. Finally, we carry out the comparative experiment between the degree and core of vertex based on 18 hyper-graphs of ISPD98 benchmark. The experiment and analysis show the core of vertex can more nearly reflects the importance of vertex in the coarse hyper-graph of each level than the degree of vertex.

http://www.joics.com/publishedpapers/2013_10_8_2279_2290.pdf

Comparative Experiment of the Core Property of Weighted Hyper-Graph Based on the ISPD98 Benchmark ⋆

In this era of big data, data are diversified, strongly connected, fragmented, dynamic, and combined with dynamic knowledge fragments to optimize the distributed storage of graphs and enable fast and efficient knowledge graph query problems. Presently, the distributed storage scheme of graph data has a large number of hop accesses between partitions, which leads to a long retrieval response time and is not conducive to fragment knowledge expansion. According to the characteristics of real-time inflow knowledge fragments and the storage structure and principles of graph databases, the Metis+ algorithm is proposed. The label graph is used as the initial initialization segmentation graph, and it is roughened to reduce the cutting of the large-weight edge. The weighted LND algorithm is proposed to run the balancing strategy for storage and assign the similar nodes and closely related nodes to the same partition to the greatest extent, which minimizes jump accesses between the partitions during retrieval.

Graph Data Retrieval Algorithm for Knowledge Fragmentation

During the coarsening phase of multilevel method,this paper introduces the concept of core and proposes the Core-Sorted Heavy-Edge Matching(CSHEM) algorithm in accordance with the compressed storage format of graph.The CSHEM algorithm not only improves previous matching schemes which are based on local information of vertex,using the global information of the finest graph core to develop its guidance role,but overcomes the defect that can only choose the Random Matching(RM) algorithm as a guide matching algorithm.Furthermore,it also presents an effective matching-based coarsening scheme that uses the CSHEM algorithm on the finest graph and the Sorted Heavy-Edge Matching(SHEM) algorithm on the coarser graphs.The scheme plays a guidance role of the core so as to make the coarser graph in accordance with the core-consistent principle.The experiment and the analysis based on ISPD98 circuit benchmark show the scheme produces encouraging performance improvement compared with those produced by the combination scheme of RM and SHEM of MeTiS that is a state-of-the-art partitioner in the literature.

Core-sorted heavy-edge matching algorithm based on compressed storage format of graph

Multilevel-based VLSI partitioner is proposed and implemented,named MVP.The MVP's framework,flow,composition modules as well as their functions are described.The multilevel partitioner of MVP achieves the better partitioning of weighted undirected graph by introducing the graph core during the coarsening phase,the spectral graph theory during the initial partitioning phase and swarm intelligence during the refinement phase.The characteristic of MVP is the conversion from netlist to weighted undirected graph and its objective is to improve the efficiency of VLSI partitioner by avoiding partitioning netlist directly.The experiment and analysis show that the multilevel partitioner of MVP can find the better partitioning and MVP can improve the performance of VLSI partitioner.

YU Song-nian

Papers

Core-sorted heavy-edge matching algorithm based on compressed storage format of graph

Research and implementation of VLSI partitioner