Scaling Distributed Deep Learning Workloads beyond the Memory Capacity with KARMA

Authors: Mohamed Wahib (National Institute of Advanced Industrial Science and Technology (AIST), RIKEN Center for Computational Science (R-CCS)); Haoyu Zhang (miHoYo Ltd); Truong Thao Nguyen (National Institute of Advanced Industrial Science and Technology (AIST)); Aleksandr Drozd and Jens Domke (RIKEN Center for Computational Science (R-CCS)); Lingqi Zhang (Tokyo Institute of Technology); Ryousei Takano (National Institute of Advanced Industrial Science and Technology (AIST)); and Satoshi Matsuoka (RIKEN Center for Computational Science (R-CCS), Tokyo Institute of Technology)

Abstract: The dedicated memory of hardware accelerators can be insufficient to store all weights and/or intermediate states of large deep learning models. Although model parallelism is a viable approach to lessen the memory pressure issue, significant modification of the source code and considerations for algorithms are required. An alternative solution is to use out-of-core methods instead of, or in addition to, data parallelism.

We propose a performance model based on the concurrency analysis of out-of-core training behavior, and derive a strategy that combines layer swapping and redundant recomputing. We achieve an average of 1.52x speedup in six different models over the state-of-the-art out-of-core methods. We also introduce the first method to solve the challenging problem of out-of-core multi-node training by carefully pipelining gradient exchanges and performing the parameter updates on the host. Our data parallel out-of-core solution can outperform complex hybrid model parallelism in training large models, e.g., Megatron-LM and Turning-NLG.




Back to Technical Papers Archive Listing