Dynamic graph processing method based on FPGA

Abstract

The invention relates to a dynamic graph processing method based on an FPGA. The method comprises the steps that under the condition that increment exists between image mirror images of adjacent timestamps of a dynamic graph, a preprocessing module divides the image mirror image of the next timestamp into at least one path unit in the mode that increment calculation of any graph vertex only depends on a preorder graph vertex of the any graph vertex; the FPGA processing module stores the at least two path units in an on-chip memory directly connected with a thread according to a mode that eachthread unit can independently process the path unit; and the thread unit determines an increment value between adjacent timestamps of the vertex of the preorder graph under the condition that the state value of the vertex of the preorder graph is updated, and transmits the increment value to a subsequent graph vertex adjacent to the vertex of the preorder graph according to the transmission direction determined by the path unit so as to update the state value of the subsequent graph vertex.

Description

technical field [0001] The invention relates to the technical field of graph calculation for computer big data processing, in particular to an FPGA-based dynamic graph processing method. Background technique [0002] With the advent of the big data era, graph processing systems have gradually become important. As a data structure, a graph has a strong expressive ability in terms of structure and semantics, and can well describe the relationship between things. By abstracting relational properties in the real world into graph data structures, graph processing systems are able to execute graph algorithms to analyze these graph data. For example, the importance ranking of graph vertices can be carried out through the PageRank algorithm; the accessibility relationship between graph vertices can be obtained through the strongly connected component algorithm. [0003] However, since the relationship between things in the real world changes every moment, which makes the graph dat...

AI Technical Summary

Problems solved by technology

[0006] Since the FPGA-based full dynamic graph method lacks the perception of dynamic graph structural changes, there are the following problems when processing dynamic graphs: (1) The graph iteration convergence speed is slow: the current FPGA-based dynamic graph full methods all use a synchronous model to Graph data performs iterative calculations, and the calculation results in the current iteration can only be used in the next iteration; therefore, the increment in each iteration can only be propagated in a single step along the edge, and the propagation speed is slow; (2) High memory access latency: In dynamic graph changes, increments are propagated along the path direction; however, current FPGA-based graph processing systems store graph data according to graph vertex numbers; the numbers of graph vertices on a path are usually randomly distributed, and they The location in the memory is discontinuous, so the spatial locality and time locality of the graph data are very poor, resulting in low memory access bandwidth utilization and high memory access delay; (3) Redundant calculation: in the process of dynamic graph processing, only The graph data affected by the increment needs to be incrementally calculated. However, in the existing FPGA-based dynamic graph full-scale method, a large amount of graph data that is not affected by the increment still participates in the calculation process, resulting in redundant calculations.

Images