ShmCaffe: A Distributed Deep Learning Platform with Shared Memory Buffer for HPC Architecture
Shinyoung Ahn, Joongheon Kim, Eunji Lim, Wan Choi, Aziz Mohaisen, Sungwon Kang · 2018
One of the reasons behind the tremendous success of deep learning theory and applications in the recent days is advances in distributed and parallel high performance computing (HPC). This paper proposes a new distributed deep learning platform, named ShmCaffe, which utilizes remote shared memory for communication overhead reduction in massive deep neural network training parameter sharing. ShmCaffe is designed based on Soft Memory Box (SMB), a virtual shared memory framework. In the SMB framework, the remote shared memory is used as a shared buffer for asynchronous massive parameter sharing among many distributed deep learning processes. Moreover, a hybrid method that combines asynchronous and synchronous parameter sharing methods is also discussed in this paper for improving scalability. As a result, ShmCaffe is 10.1 times faster than Caffe and 2.8 times faster than Caffe-MPI for deep neural network training when Inception_v1 is trained with 16 GPUs. We verify the convergence of the Inception_v1 model training using ShmCaffe-A and ShmCaffe-H by varying the number of workers. Furthermore, we evaluate scalability of ShmCaffe by analyzing the computation and communication times per one iteration of deep learning training in four convolutional neural network (CNN) models.