CLIO: enabling automatic compilation of deep learning pipelines across IoT and cloud
CLIO: enabling automatic compilation of deep learning pipelines across IoT and cloud
复制标题
DOI：
10.1145/3372224.3419215
复制
发表时间：
2020-09
期刊：
影响因子：
--
通讯作者：
Jin Huang;Colin Samplawski;Deepak Ganesan;Benjamin M. Marlin;H. Kwon
中科院分区：
文献类型：
--
作者：
Jin Huang;Colin Samplawski;Deepak Ganesan;Benjamin M. Marlin;H. Kwon
近年来，旨在将深度学习分析引入物联网设备的低功耗神经加速器取得了巨大进步;同时，低功耗无线电的设计也取得了相当大的进步，可以实现从物联网设备到云端的高效计算卸载。这两种方法都不是万能的--深度学习模型对于低功耗加速器来说往往太大，而对于低功耗无线电来说，带宽需求往往太高。虽然在智能手机类设备的深度学习方面已经做了大量的工作，但这些方法对于资源受限的小型电池供电的物联网设备并不适用。在本文中，我们通过设计一种可持续调整的方法来利用本地和远程资源来优化深度学习模型的性能，从而弥合了这一差距。克利奥提出了一种新颖的方法，以适应无线动态的渐进方式在物联网设备和云之间拆分机器学习模型。我们证明了这种方法可以与模型压缩和自适应模型分区相结合，以创建一个集成的物联网云分区系统。我们实现克利奥的GAP 8低功耗神经加速器，提供了一个详尽的表征的操作制度，每种方法表现最好，并表明，克利奥可以使优雅的性能下降，资源减少。
Recent years have seen dramatic advances in low-power neural accelerators that aim to bring deep learning analytics to IoT devices; simultaneously, there have been considerable advances in the design of low-power radios to enable efficient compute offload from IoT devices to the cloud. Neither is a panacea --- deep learning models are often too large for low-power accelerators and bandwidth needs are often too high for low-power radios. While there has been considerable work on deep learning for smartphone-class devices, these methods do not work well for small battery-powered IoT devices that are considerably more resource-constrained. In this paper, we bridge this gap by designing a continuously tunable method for leveraging both local and remote resources to optimize performance of a deep learning model. Clio presents a novel approach to split machine learning models between an IoT device and cloud in a progressive manner that adapts to wireless dynamics. We show that this method can be combined with model compression and adaptive model partitioning to create an integrated system for IoT-cloud partitioning. We implement Clio on the GAP8 low-power neural accelerator, provide an exhaustive characterization of the operating regimes where each method performs best and show that Clio can enable graceful performance degradation as resources diminish.