PMLS-Caffe: Distributed Deep Learning Framework for Parallel ML System
PMLS-Caffe (formerly Poseidon) is a scalable open-source framework for large-scale distributed deep learning on CPU/GPU clusters. It is initially released in January 2015 along with PMLS v1.0 as an application under the Bösen parameter server.
PMLS-Caffe builds upon the Caffe (http://caffe.berkeleyvision.org/) CNN libraries and the PMLS distributed ML framework (http://sailing-lab.wixsite.com/sailing-pmls) as a starting point, but goes further by implementing three key contributions for efficient CNN training on clusters of GPU-equipped machines: (i) a three-level hybrid architecture that allows PMLS-Caffe to support both CPU-only clusters as well as GPU-equipped clusters, (ii) a distributed wait-free backpropagation (DWBP) algorithm to improve GPU utilization and to balance communication, and (iii) a dedicated structure-aware communication protocol (SACP) to minimize communication overheads.
PMLS-Caffe's design philosophy is rooted on efficiently harnessing multiple, distributed GPUs on commodity hardware and Ethernet, in order to maximize the speedup with a fully data parallel scheme for distributed deep learning. We empirically evaluate PMLS-Caffe regarding of throughput, convergence and accuracy on the image classification tasks with multiple standard datasets, and show that PMLS-Caffe is able to achieve state-of-the-art speedups in accelerating the training of modern CNN structures, at the same time guarantee the correct convergence.
PMLS-Caffe inherits many functionalities and benefits of PMLS, including the Sufficient Factor Broadcasting (SFB), managed communication and bandwidth management in the Bösen parameter server, etc. Moreover, most of the Caffe interfaces are kept unchanged.
Please consult the documentation page for more details on how to setup PMLS-Caffe on your clusters and start training your model. We also disclose the system architecture of PMLS-Caffe and several distributing strategies for fast parallelization of deep learning in the following arXiv paper: