US 12,475,078 B2
Post-exascale graph computing method, system, storage medium and electronic device thereof
Yu Zhang, Wuhan (CN); Jin Zhao, Wuhan (CN); Hui Yu, Wuhan (CN); Yun Yang, Wuhan (CN); Xinyu Jiang, Wuhan (CN); Shijun Li, Wuhan (CN); Xiaofei Liao, Wuhan (CN); and Hai Jin, Wuhan (CN)
Assigned to HUAZHONG UNIVERSITY OF SCIENCE AND TECHNOLOGY, Wuhan (CN)
Filed by Huazhong University of Science and Technology, Wuhan (CN)
Filed on Jul. 27, 2022, as Appl. No. 17/815,436.
Claims priority of application No. 202210234737.4 (CN), filed on Mar. 7, 2022.
Prior Publication US 2023/0281157 A1, Sep. 7, 2023
Int. Cl. G06F 15/80 (2006.01); G06F 9/50 (2006.01); H04L 67/10 (2022.01)
CPC G06F 15/80 (2013.01) [G06F 9/5066 (2013.01); H04L 67/10 (2013.01)] 4 Claims
OG exemplary drawing
 
1. A post-exascale graph computing method, comprising:
a step of performing distributed, asynchronous graph processing and a step of establishing hierarchical, very-large-scale, distributed communication;
in which said step of performing distributed, asynchronous graph processing comprises:
with a core subgraph constructed, having at least at least one computing node select graph blocks including active graph vertices in said core subgraph selected first for processing based on a topology-aware graph processing mechanism to propagate state information of graph vertices and accelerate state convergence among said graph vertices, generating a communications message, wherein each graph vertices contain respective state values;
and in which said step of establishing hierarchical, very-large-scale, distributed communication comprises:
arranging a plurality of computing nodes onto floors in a tree-like structure, partitioning graph data in a community-structure-aware manner, and assigning said graph data to individual said computing nodes on a floor,
partitioning the graph data in a community-structure-aware manner to form graph partitions, wherein partitioning comprises: determining partitioned core subgraphs based on the graph partitions having their graph vertices highly dependent on each other; and grouping the partitioned core subgraphs into the same group of computing nodes on a floor according to the dependency of the state values, so that frequent communication among the partitioned core subgraphs happens inside the same group of computing nodes on the same floor,
when plural said computing nodes on the same floor communicate, merging communication information associated with the same graph vertex based on reduction so as to decrease communication traffic, wherein reduction is by an algorithm in which information sent to the same computing node is merged into a same queue for communications at batch; and
sending the communication information to be sent to the same computing node on a higher floor in a merged and unified form;
after the graph data of each said higher floor computing node and of a lower floor computing node that is subordinate to said higher floor computing node have converged, having said high floor computing node communicate, after reduction merger, and compression of communications messages, reducing communication traffic, with a next higher floor computing node that said higher floor computing node is subordinate to and with said higher floor computing nodes that are of said some floor as said higher floor computing node, so as to synchronize states of the graph vertices that are peers; and
performing communication floor by floor until all said graph vertices in a cluster have their states converged.