Characterizing and understanding deep neural network batching systems on GPUs

Abstract

As neural network inference demands are ever-increasing in intelligent applications, the performance optimization of model serving becomes a challenging problem. Dynamic batching is an important feature of contemporary deep learning serving systems, which combines multiple requests of model inference and executes them together to improve the system’s throughput. However, the behavior characteristics of each part in deep neural network batching systems as well as their performance impact on different model structures are still unknown. In this paper, we characterize the batching system by leveraging three representative deep neural networks on GPUs, performing a systematic analysis of the performance effects from the request batching module, model slicing module, and stage reorchestrating module. Based on experimental results, several insights and recommendations are offered to facilitate the system design and optimization for deep learning serving.