Local SGD with Periodic Averaging: Tighter Analysis and Adaptive Synchronization

Authors: Farzin Haddadpour, Mohammad Mahdi Kamani, Mehrdad Mahdavi, Viveck Cadambe

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we validate the theory with experimental results, running over AWS EC2 clouds and an internal GPU cluster.
Researcher Affiliation Academia Farzin Haddadpour Penn State fxh18@psu.edu Mohammad Mahdi Kamani Penn State mqk5591@psu.edu Mehrdad Mahdavi Penn State mzm616@psu.edu Viveck R. Cadambe Penn State vxc12@psu.edu
Pseudocode Yes Algorithm 1 LUPA-SGD(τ): Local updates with periodic averaging.
Open Source Code Yes The implementation code is available at https://github.com/mmkamani7/LUPA-SGD.
Open Datasets Yes Epsilon dataset2, a popular large scale binary dataset, consists of 400, 000 training samples and 100, 000 test samples with feature dimension of 2000. (Footnote 2: https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html)
Dataset Splits No The paper mentions 400,000 training samples and 100,000 test samples but does not specify a validation split or details on how data was partitioned for validation.
Hardware Specification Yes Most of the experiments will be run on Amazon EC2 cluster with 5 p2.xlarge instances. ... We also use an internal high performance computing (HPC) cluster equipped with NVIDIA Tesla V100 GPUs.
Software Dependencies No The paper mentions software like PyTorch, TensorFlow, mpi4py, and Open MPI, but it does not specify version numbers for any of these dependencies, which is required for reproducibility.
Experiment Setup Yes The learning rate and regularization parameter are 0.01 and 1 10 4, respectively, and the minibatch size is 128 unless otherwise stated.