A GPU-based asynchronous graph data processing system

Abstract

The invention discloses a GPU-based asynchronous graph data processing system, including: realizing a mixed coloring algorithm: changing the original single, heuristic coloring algorithm into a mixed coloring algorithm; data preprocessing strategy: changing the original coloring algorithm based on The preprocessing and segmentation method of vertex degree is changed to the method based on vertex coloring distribution; asynchronous processing execution engine: the execution method of the processing engine is changed from the synchronous BSP method to the asynchronous processing method, combined with the hybrid shading algorithm module, to achieve lock-free on the GPU Asynchronous programming; inside the asynchronous processing execution engine, iterative processing module and data flow execution module are implemented, and resources such as bandwidth and computing power of CPU and GPU are managed and used in a unified manner. The present invention implements a GPU-based graph data processing system. The asynchronous execution engine can greatly accelerate the graph algorithm convergence speed, solve problems such as synchronous waiting overhead and asynchronous programming lock overhead on the GPU, and further improve the GPU's ability to process graph data.

Description

technical field [0001] The invention belongs to the field of graph data processing and high-performance computing, and more specifically relates to a GPU-based asynchronous graph data processing system. Background technique [0002] Large-scale graph data processing is involved in more and more applications. Traditional stand-alone graph processing systems divide large-scale graph data and then process data blocks sequentially. Compared with the cumbersome programming of distributed systems, simple stand-alone systems have attracted more and more attention. The key to a stand-alone system is to solve problems such as poor disk access performance, low task concurrency, and low utilization of device memory bandwidth due to graph data partitioning. In the academic community, some active exploration and research have been carried out on such problems. try. [0003] The GraphChi system adopts the newly designed sliding window PSW (Parallel Sliding Window) algorithm, which divid...

AI Technical Summary

Problems solved by technology

[0008] In view of the low efficiency of the data division strategy of the current stand-alone graph processing system and the poor concurrency of graph vertex processing tasks, the present invention proposes a GPU-based asynchronous graph processing system that takes into account the data exchange overhead between the GPU device memory and the host memory problem, adopt a data partition strategy based on the mixed graph coloring algorithm, improve data concurrency, make full use of the powerful parallel processing capability of the GPU, and greatly improve the performance of single-machine graph processing

Images