Research On Optimization Of Graph Computing System For Heterogeneous Platform

As the carrier of the graph algorithm,the graph computing system can encapsulate many graph algorithms into a simple and efficient system interface for enterprises and related researchers to use.At present,although there have been many graph computing systems with outstanding performance in scientific research and industrial fields,these systems have not yet given full play to the advantages of the platform architecture.This paper mainly studies the most representative Sp MV algorithm in the graph computing system.By researching and analyzing the current problems of frequent memory access,load imbalance,and inefficient vectorization in the graph computing system and taking reasonable measures to optimize the computing performance of Sp MV,provides a full range of system-level optimization solutions for the graph computing system.First of all,this paper designs an optimization scheme of the graph computing system based on the new storage format CSR2.It is a new single format suitable for processor platforms with SIMD(Single Instruction Multiple Data)vectorization capabilities.This format is easy to implement in source code operations and has a low conversion overhead.Compared with the most advanced single storage format CSR5 and Intel high-performance library MKL,CSR2 has significant performance improvements.On the Intel Xeon E5-2670 v3 platform,compared with CSR5,CSR2 has a mean speedup of 1.401x(up to 1.861x);compared with MKL,CSR2 has a mean speedup of 1.261x(up to 5.921x).Secondly,this paper designs an optimization scheme of the graph computing system based on the new load balancing algorithm ALBUS.It solves the two significant challenges of the current graph computing system: first,to achieve balanced multi-core processor load processing and reduce memory access operations;second,to give full play to the parallel capabilities of the SIMD vectorized instruction set.In this paper,20 sets of regular matrices and 20 sets of irregular matrices are selected to form the benchmark suite.On the Intel Xeon E5-2670 v3 CPU platform,compared with CSR5,Merge,and MKL,for 20 sets of regular matrices,ALBUS has a mean speedup of 1.59 x,1.32 x,and 1.48 x,respectively(up to 2.53 x,2.22 x,and2.31x).For 20 groups of irregular matrices,ALBUS has a mean speedup of 1.38 x,1.42 x,and 2.44 x,respectively(up to 2.33 x,2.24 x,and 5.37x).Finally,this paper designs an optimization scheme of the graph computing system based on GPU segmented optimization strategy GSp MV.In this algorithm,we introduced a segmented optimization strategy to achieve balanced load processing.Experiments show that on the Nvidia Tesla T4 GPU platform,compared with Merge,Cusparse(CSR),and Cusparse(HYB),GSp MV has a mean speedup of 1.07 x,1.86 x,and 1.24 x,respectively(up to 1.54 x,138.65 x,and 2.32x).