A Processor Selection Method based on Execution Time Estimation for Machine Learning Programs
A Processor Selection Method based on Execution Time Estimation for Machine Learning Programs
复制标题
DOI：
10.1109/ipdpsw52791.2021.00116
复制
发表时间：
2021-06
期刊：
影响因子：
--
通讯作者：
Kou Murakami;Kazuhiko Komatsu;Masayuki Sato;Hiroaki Kobayashi
中科院分区：
文献类型：
--
作者：
Kou Murakami;Kazuhiko Komatsu;Masayuki Sato;Hiroaki Kobayashi
近年来，机器学习变得越来越普遍。由于机器学习算法已经变得复杂，并且要处理的数据量已经变得很大，因此机器学习程序的执行时间一直在增加。被称为加速器的处理器可以在短时间内执行机器学习程序。然而，包括加速器的处理器具有不同的特性。因此，目前还不清楚现有的机器学习程序是否在适当的处理器上执行。本文提出了一种选择适合于每个机器学习程序的处理器的方法。在所提出的方法中，选择是基于每个处理器上的机器学习程序的执行时间的估计。所提出的方法不需要预先执行目标机器学习程序。从实验结果可以看出，该方法的执行速度比NumPy的原始实现快5.3倍。这些结果证明，所提出的方法可以用于自动选择处理器的系统中，以便每个机器学习程序可以在最佳处理器上轻松执行。
In recent years, machine learning has become widespread. Since machine learning algorithms have become complex and the amount of data to be handled have become large, the execution times of machine learning programs have been increasing. Processors called accelerators can contribute to the execution of a machine learning program with a short time. However, the processors including the accelerators have different characteristics. Therefore, it is unclear whether existing machine learning programs are executed on the appropriate processor or not. This paper proposes a method for selecting a processor suitable for each machine learning program. In the proposed method, the selection is based on the estimation of the execution time of machine learning programs on each processor. The proposed method does not need to execute a target machine learning program in advance. From the experimental results, it is clarified that the proposed method can achieve up to 5.3 times faster execution than the original implementation by NumPy. These results prove that the proposed method can be used in a system that automatically selects the processor so that each machine learning program can be easily executed on the best processor.