Design, Implementation and Performance Evaluation of a Stochastic Gradient Descent Algorithm on CUDA
Sapienza University of Rome
Sapienza University of Rome, Technical Report n. 11, 2015
@techreport{de2015design,
title={Design, Implementation and Performance Evaluation of a Stochastic Gradient Descent Algorithm on CUDA},
author={De Falco, Emanuele and others},
year={2015},
institution={Department of Computer, Control and Management Engineering, Universita’degli Studi di Roma" La Sapienza"}
}
Stochastic Gradient Descent, a stochastic optimization of Gradient Descent, is an algorithm that is used in different topics, like for example for linear regression or logistic regression. After the Netflix prize, SGD start to be used also in recommender systems to compute matrix factorization. Considering the large amounts of data that this kind of system has to elaborate, adapt the algorithm on a distributed system or parallelize it is a good idea to improve performance. One way to do this is by using GPGPU, that thanks to its characteristics it’s now days a good solution for parallelize an application. With this work, we are interested in analyze how SGD works on a GPGPU environment that is designed with a CUDA architecture, starting from an existing implementation for parallel environments and then adapting it to exploits all characteristics that a GPU of this kind provide.
November 29, 2015 by hgpu
