arXiv subjects

Franz-Josef Pfreundt

Publications and source records attributed to Franz-Josef Pfreundt.

22 records · Page 2Linked to original sources

arXiv2017

Using GPI-2 for Distributed Memory Paralleliziation of the Caffe Toolbox to Speed up Deep Neural Network Training

Deep Neural Network (DNN) are currently of great inter- est in research and application. The training of these net- works is a compute intensive and time consuming task. To reduce training times to a bearable amount at reasonable cost we extend the popular Caffe toolbox for DNN with an efficient distributed memory communication pattern. To achieve good scalability we emphasize the overlap of computation and communication and prefer fine granu- lar synchronization patterns over global barriers. To im- plement these communication patterns we rely on the the Global address space Programming Interface version 2 (GPI-2) communication library. This interface provides a light-weight set of asynchronous one-sided communica- tion primitives supplemented by non-blocking fine gran- ular data synchronization mechanisms. Therefore, Caf- feGPI is the name of our parallel version of Caffe. First benchmarks demonstrate better scaling behavior com- pared with other extensions, e.g., the Intel TM Caffe. Even within a single symmetric multiprocessing machine with four graphics processing units, the CaffeGPI scales bet- ter than the standard Caffe toolbox. These first results demonstrate that the use of standard High Performance Computing (HPC) hardware is a valid cost saving ap- proach to train large DDNs. I/O is an other bottleneck to work with DDNs in a standard parallel HPC setting, which we will consider in more detail in a forthcoming paper.

cs.LG↗

arXiv2016

Distributed Training of Deep Neural Networks: Theoretical and Practical Limits of Parallel Scalability

This paper presents a theoretical analysis and practical evaluation of the main bottlenecks towards a scalable distributed solution for the training of Deep Neuronal Networks (DNNs). The presented results show, that the current state of the art approach, using data-parallelized Stochastic Gradient Descent (SGD), is quickly turning into a vastly communication bound problem. In addition, we present simple but fixed theoretic constraints, preventing effective scaling of DNN training beyond only a few dozen nodes. This leads to poor scalability of DNN training in most practical scenarios.

cs.CV↗

arXiv2015

Asynchronous Parallel Stochastic Gradient Descent - A Numeric Core for Scalable Distributed Machine Learning Algorithms

The implementation of a vast majority of machine learning (ML) algorithms boils down to solving a numerical optimization problem. In this context, Stochastic Gradient Descent (SGD) methods have long proven to provide good results, both in terms of convergence and accuracy. Recently, several parallelization approaches have been proposed in order to scale SGD to solve very large ML problems. At their core, most of these approaches are following a map-reduce scheme. This paper presents a novel parallel updating algorithm for SGD, which utilizes the asynchronous single-sided communication paradigm. Compared to existing methods, Asynchronous Parallel Stochastic Gradient Descent (ASGD) provides faster (or at least equal) convergence, close to linear scaling and stable accuracy.

cs.DC↗

arXiv2015

Balancing the Communication Load of Asynchronously Parallelized Machine Learning Algorithms

Stochastic Gradient Descent (SGD) is the standard numerical method used to solve the core optimization problem for the vast majority of machine learning (ML) algorithms. In the context of large scale learning, as utilized by many Big Data applications, efficient parallelization of SGD is in the focus of active research. Recently, we were able to show that the asynchronous communication paradigm can be applied to achieve a fast and scalable parallelization of SGD. Asynchronous Stochastic Gradient Descent (ASGD) outperforms other, mostly MapReduce based, parallel algorithms solving large scale machine learning problems. In this paper, we investigate the impact of asynchronous communication frequency and message size on the performance of ASGD applied to large scale ML on HTC cluster and cloud environments. We introduce a novel algorithm for the automatic balancing of the asynchronous communication load, which allows to adapt ASGD to changing network bandwidths and latencies.

cs.DC↗