@inproceedings{sensi_bine_2025,
  title = {Bine {{Trees}}: {{Enhancing Collective Operations}} by {{Optimizing Communication Locality}}},
  booktitle = {Proceedings of the {{International Conference}} for {{High Performance Computing}}, {{Networking}}, {{Storage}} and {{Analysis}}},
  author = {Sensi, Daniele De and Pasqualoni, Saverio and Piarulli, Lorenzo and Bonato, Tommaso and Ba, Seydou and Turisini, Matteo and Domke, Jens and Hoefler, Torsten},
  date = {2025-11},
  series = {{{SC}} '25},
  publisher = {IEEE Press},
  abstract = {As high-performance computing (HPC) systems scale, optimizing communication locality becomes essential for performance. This paper introduces Binomial Negabinary (Bine) trees, a novel approach to enhancing collective operations by reducing inter-domain communication. Bine trees can be applied to both tree-based and butterfly-based collective algorithms, making them versatile and adaptable. They minimize the distance between communicating ranks, reducing traffic on global links and alleviating congestion. Unlike traditional hierarchical algorithms, Bine trees are topology-agnostic and do not assume a uniform partition of ranks, making them ideal for production supercomputers with irregular process allocations. We design new algorithms for eight different collectives, demonstrating significant performance improvements up to 2x, along with a reduction of traffic on global links by up to 25\%. Our results emphasize their effectiveness in improving performance while reducing the load on global communication links.},
  keywords = {deep neural networks,GPUs,out-of-core}
}