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.
%0 Generic
%1 KueKeu17Using
%A Kuehn, Martin
%A Keuper, Janis
%A Pfreundt, Franz-Josef
%D 2017
%K GPI deep_learning distributed scaling
%T Using GPI-2 for Distributed Memory Paralleliziation of the Caffe Toolbox
to Speed up Deep Neural Network Training
%U http://arxiv.org/abs/1706.00095
%X 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.
@misc{KueKeu17Using,
abstract = {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.},
added-at = {2018-06-26T08:39:55.000+0200},
author = {Kuehn, Martin and Keuper, Janis and Pfreundt, Franz-Josef},
biburl = {https://www.bibsonomy.org/bibtex/25ace26772c036e4171f9660ebeb7daa6/loroch},
description = {1706.00095.pdf},
interhash = {cec32dfe257aad82640105bedcf059b7},
intrahash = {5ace26772c036e4171f9660ebeb7daa6},
keywords = {GPI deep_learning distributed scaling},
note = {cite arxiv:1706.00095},
timestamp = {2018-06-26T08:41:42.000+0200},
title = {Using GPI-2 for Distributed Memory Paralleliziation of the Caffe Toolbox
to Speed up Deep Neural Network Training},
url = {http://arxiv.org/abs/1706.00095},
year = 2017
}