A Reliable Effective Terascale Linear Learning System

Опубликовано 28 июля 2016, 0:24

We present a system and a set of techniques for learning linear predictors with convex losses on terascale datasets, with trillions of features (the number of features here refers to the number of non-zero entries in the data matrix), billions of training examples and millions of parameters in an hour using a cluster of 1000 machines. Individually none of the component techniques is new, but the careful synthesis required to obtain an efficient implementation is a novel contribution. The result is, up to our knowledge, the most scalable and efficient linear learning system reported in the literature. We describe and thoroughly evaluate the components of the system, showing the importance of the various design choices.