Blink: Fast and Generic Collectives for Distributed ML

  • Guanhua Wang ,
  • Shivaram Venkataraman ,
  • Amar Phanishayee ,
  • ,
  • Nikhil Devanur ,
  • Ion Stoica

Conference on Machine Learning and Systems (MLSys 2020) |

Related File

Model parameter synchronization across GPUs introduces high overheads for data-parallel training at scale. Existing parameter synchronization protocols cannot effectively leverage available network resources in the face of ever increasing hardware heterogeneity. To address this, we propose Blink, a collective communication library that dynamically generates optimal communication primitives by packing spanning trees. We propose techniques to minimize the number of trees generated and extend Blink to leverage heterogeneous communication channels for faster data transfers. Evaluations show that compared to the state-of-the-art (NCCL), Blink can achieve up to 8× faster model synchronization, and reduce end-to-end training time for image classification tasks by up to 40%.