GTC Silicon Valley-2019: Intelligent Operation and Maintenance of Public Cloud Based on GPU-Accelerated Machine Learning
GTC Silicon Valley-2019 ID:S9845:Intelligent Operation and Maintenance of Public Cloud Based on GPU-Accelerated Machine Learning
Zhiming Wang(Alibaba (China))
The hardest part of cloud computing engineering is operations because of the complexity of managing thousands of machines, but machine learning can add intelligence to public cloud operation and maintenance. We use RAPIDS to accelerate machine learning and the NVIDIA TensorRT inference server for GPU load balancing and improved GPU utilization. We'll explain how to use traditional machine learning algorithms such as ARIMA, XGBoost, and RandomForest for load prediction, load classification, user portrait, exception prediction, and other scenarios. Learn how to use GPUs for data preprocessing and algorithm acceleration for large-scale data analysis and machine learning of massive public cloud data. In addition, we'll cover how we implemented a large-scale training and prediction service platform based on Dask and NVIDIA's inference server. The platform can support large-scale GPU parallel computing and prediction requests.