Efficient Use of Limited-Memory Accelerators for Linear Learning on Heterogeneous Systems

TL;DR

This paper introduces a flexible algorithmic approach that enables efficient training of large-scale machine learning models on heterogeneous systems with limited memory, achieving significant speedups over existing methods.

Contribution

It presents a novel primal-dual coordinate method leveraging duality gap information for dynamic data management in heterogeneous environments.

Findings

Order-of-magnitude speedup over existing approaches

Effective training of models exceeding GPU memory capacity

Adaptability to various system memory hierarchies

Abstract

We propose a generic algorithmic building block to accelerate training of machine learning models on heterogeneous compute systems. Our scheme allows to efficiently employ compute accelerators such as GPUs and FPGAs for the training of large-scale machine learning models, when the training data exceeds their memory capacity. Also, it provides adaptivity to any system's memory hierarchy in terms of size and processing speed. Our technique is built upon novel theoretical insights regarding primal-dual coordinate methods, and uses duality gap information to dynamically decide which part of the data should be made available for fast processing. To illustrate the power of our approach we demonstrate its performance for training of generalized linear models on a large-scale dataset exceeding the memory size of a modern GPU, showing an order-of-magnitude speedup over existing approaches.