A HLS-based Graph Processing Approach On FPGA-HBM

Graph computing are widely used in many real-world scenarios,such as path navigation,social network analysis,and ad recommendations.With the rapid growth of graph data scale,the demand for high performance of graph computing is increasing day by day.FPGAs have become a new computing infrastructure,including data centers,thanks to their fine-grained parallelism,energy efficiency,and programmability.High-level synthesis allows users to program FPGAs in high-level languages,significantly lowers the threshold for FPGA hardware development,and the use of high-level technology is an important technical way to balance the efficiency and ease of use of graph computing.However,the existing graph computing method based on high-level synthesis,due to the high consumption of resources and frequency decrease,results in the poor scalability of the graph computing accelerator system,can not take full advantage of the hardware characteristics of the new High Bandwidth Memory,the graph computing performance accelerator effect is not ideal.In order to solve the above problems,a scalable high-level integrated graph computing method ScalaGP is proposed.With a hierarchical data forwarding mechanism,ScalaGP can reduce the usage of lookup table resources from complexity ()( for the number of vertex processing units, for the number of edge processing units)to (7)2)).With simpler module design and less resource consumption,enabling higher number of parallel processing elements and higher frequency.ScalaGP can significantly improve the efficiency of memory access in the process of graph computing through a series of memory optimization strategies.At the same time,ScalaGP abstracts the basic operators needed for graph algorithms and designs a hardware-friendly accelerator template for graph computing,to help users deploy hardware accelerators.The experimental results show that the number of ScalaGP parallel processing elements can be extended to at least 384 and the frequency can reach about 230 MHZ.Compared with the latest high-level integrated graph computing framework Thunder GP,parallel processing capacity increased by at least 12 times,resource consumption is only about 75% of Thunder GP,performance acceleration of 8.64 to 18.94 times.