Optimization of the recursive K-nomial algorithm in the UCC communication library
Juan Pang, Wei Wan, Guangyao Zhang, Yongbo Wang · 2025
Collective communication has long been a core concept in the field of parallel computing. However, with the rapid development of this field, the scale and complexity of data have been growing exponentially. To address these challenges, the Unified Communication Framework (UCF) has brought together multiple research teams to develop a unified collective communication library—UCC. This library has been continuously improved through collaboration between academia and industry. Within the UCP layer of the UCC library, the Recursive K-nomial algorithm encounters certain issues related to process communication. Specifically, the original algorithm may lead to uneven distribution of processes across nodes, negatively impacting communication efficiency. In response to this, we propose an optimization to the algorithm that achieves load balancing. Testing on the computer system at Zhengzhou University shows that, for communication on a scale of 128 nodes, the performance of the allreduce_knomial algorithm improved by up to 54.31%, and the performance of the allgather_knomial algorithm improved by up to 46.07%.