NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:319
Title:SySCD: A System-Aware Parallel Coordinate Descent Algorithm

This paper identifies several performance bottlenecks in parallel implementation of randomized coordinate-descent methods for solving optimization of large scale linear models, and proposes a variant that can effectively mitigate these effects and improve performance. The theoretical analysis of the algorithm follow standard arguments, but look to be necessary due to their modifications. The systematic study of parallel performance against different parallel architectures is refreshing and can be practically very useful.