Adding vs. Averaging in Distributed Primal-Dual Optimization

Authors: Chenxin Ma, Virginia Smith, Martin Jaggi, Michael Jordan, Peter Richtarik, Martin Takac

ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide a thorough experimental comparison with competing algorithms using several real-world distributed datasets. Our practical results confirm the strong scaling of COCOA+ as the number of machines K grows, while competing methods, including the original COCOA, slow down significantly with larger K. We implement all algorithms in Spark, and our code is publicly available at: github.com/gingsmith/cocoa.
Researcher Affiliation Academia Chenxin Ma CHM514@LEHIGH.EDU Industrial and Systems Engineering, Lehigh University, USA Virginia Smith VSMITH@BERKELEY.EDU University of California, Berkeley, USA Martin Jaggi JAGGI@INF.ETHZ.CH ETH Z urich, Switzerland Michael I. Jordan JORDAN@CS.BERKELEY.EDU University of California, Berkeley, USA Peter Richt arik PETER.RICHTARIK@ED.AC.UK School of Mathematics, University of Edinburgh, UK Martin Tak aˇc TAKAC.MT@GMAIL.COM Industrial and Systems Engineering, Lehigh University, USA
Pseudocode Yes Algorithm 1 COCOA+ Framework
Open Source Code Yes We implement all algorithms in Spark, and our code is publicly available at: github.com/gingsmith/cocoa.
Open Datasets Yes The used datasets are summarized in Table 2.
Dataset Splits No The paper does not provide specific dataset split information (percentages, sample counts, or explicit splitting methodology) for training, validation, or test sets.
Hardware Specification Yes We implement all algorithms in Apache Spark (Zaharia et al., 2012) and run them on m3.large Amazon EC2 instances
Software Dependencies No The paper mentions 'Apache Spark' but does not specify its version number or any other software dependencies with their versions.
Experiment Setup Yes We compare the COCOA+ and COCOA frameworks directly using two datasets (Covertype and RCV1) across various values of λ, the regularizer, in Figure 1. For each value of λ we consider both methods with different values of H, the number of local iterations performed before communicating to the master. For all runs of COCOA+ we use the safe upper bound of γK for σ .