摘要
阐述各类Internet服务与智能手机应用程序,每分钟都会产生大量数据,这大大刺激了包括深度学习在内的数据收集、存储、分析需求。但是,探讨如何优化深度学习训练平台的I/O表现,目前的研究成果还比较少。从"类脑智能开放平台"真实数据出发,分析了深度学习训练平台I/O特性,并基于以上,兼顾缓存命中率与节点负载均衡,提出了缓存联合部署策略。仿真结果表明,相比于常用的LRU策略,提出的策略具有更高的缓存命中率,且各计算节点之间的负载较为均衡,从而较好地优化了深度学习应用的I/O表现。
Various Internet services and smart phone applications generate a large amount of data every minute,which greatly stimulates the data collection,storage,and analysis needs including deep learning.However,there are relatively few studies on how to optimize the I/O performance of the deep learning training platform.To this end,this paper analyzes the I/O characteristics of the deep learning training platform based on the real data of the"brain-like intelligent open platform",and based on the above analysis,takes into account the cache hit rate and node load balance,and proposes a cache joint deployment strategy.The simulation results show that compared with the commonly used LRU strategy,the strategy proposed in this paper has a higher cache hit rate,and the load between the computing nodes is more balanced,which better optimizes the I/O performance of deep learning applications.
作者
鲍裕麟
郑烇
BAO Yulin;ZHENG Quan(Laboratory of Future Networks,Department of Automation,University of Science and Technology of China,Anhui 230026,China;Institute of Advanced Technology,University of Science and Technology of China,Anhui 230088,China)
出处
《电子技术（上海）》
2021年第7期61-67,共7页
Electronic Technology
关键词
深度学习训练平台
I/O特性
缓存策略
部署策略
deep learning training platform
I/O characteristics
caching strategy
deployment strategy
