In heterogeneous cluster environment

the network topology is irregular

and performance gap exists among computer nodes and among interconnected networks.These factors affect the performance of collective communication.To solve this problem

we present MGO (Multi-Granularity Optimization) algorithm.In MGO algorithm

we optimize the path for collective communication by network topology and the performance of the computer nodes and the networks.The simulation results show that by comparing MGO algorithm with related optimized algorithm

the former can result in notable performance improvement.Moreover

the performance-enhancing ratio increases to some extent as the cluster scales up.