Implementation And Collaborative Optimization Of A Fine-grained And Unified Addressing Parallel Architecture For Graph Search

The age of Big Data is coming.Now the data size is growing rapidly and the applications require faster and faster processing speed.These bring a new great challenge for the processor architecture and parallel computer system architecture.The graph search problem is a hot issue in big data research and is widely used in various fields of scientific research.Based on the analysis of load characteristics of graph search problem,the structure of multi-node parallel processing system and the design of multi-node parallel BFS algorithm for graph search are further studied based on the single-node processor structure.Then we carry out hardware and software collaborative optimization research based on the prototype system.The size of the data involved in the search problem is often large and requires many nodes to process parallelly.Aiming at the problem of global and fine granularity and irregularity,a fine-grained and unified addressing parallel architecture is designed.In the paper,we designed a multi-node system with distributed shared memory and a fine-grained access interface between the nodes based on the high-speed communication interface of the private network so that each processor can access the system's arbitsrary memory data through the physical address.Based on the parallel architecture design,an 8-node prototype system platform is implemented.Then we designed the parallel BFS algorithm based on 1-D partition on the prototype system platform and analyze the data structure design and the access mode of the algorithm to carry out further collaborative optimization.In order to verify the correctness of the algorithm,a multi-node parallel BFS algorithm simulator is designed to simulate the parallel execution of the program,which can verify the correctness of the algorithm and obtain the algorithm execution parameters.In order to optimize the overall performance of the system,we conducted further performance testing and collaborative optimization on the multi-node prototype system platform.Based on the test result analysis,we conducted hardware and software collaborative optimization in three perspectives,which are system remote access scheme,local access scheme and global communication scheme,and achieved a good performance on the 8-node prototype system platform of 805.9 MTEPS and a good scale,which is 4.89 times of the performance for the single-node system.