Characterizing Deep Learning Training Workloads on Alibaba-PAI

TL;DR

This paper analyzes deep learning training workloads on Alibaba's PAI platform, identifying communication as the main bottleneck and exploring hardware/software optimizations for performance improvements.

Contribution

It provides a detailed characterization of training workloads, performance bottlenecks, and potential hardware/software optimizations for Alibaba's AI cloud.

Findings

Weight/gradient communication accounts for 62% of execution time.

Computation is not the primary bottleneck.

Upgrading interconnects and architecture can significantly improve performance.

Abstract

Modern deep learning models have been exploited in various domains, including computer vision (CV), natural language processing (NLP), search and recommendation. In practical AI clusters, workloads training these models are run using software frameworks such as TensorFlow, Caffe, PyTorch and CNTK. One critical issue for efficiently operating practical AI clouds, is to characterize the computing and data transfer demands of these workloads, and more importantly, the training performance given the underlying software framework and hardware configurations. In this paper, we characterize deep learning training workloads from Platform of Artificial Intelligence (PAI) in Alibaba. We establish an analytical framework to investigate detailed execution time breakdown of various workloads using different training architectures, to identify performance bottleneck. Results show that weight/gradient communication during training takes almost 62% of the total execution time among all our workloads on average. The computation part, involving both GPU computing and memory access, are not the biggest bottleneck based on collective behavior of the workloads. We further evaluate attainable performance of the workloads on various potential software/hardware mappings, and explore implications on software architecture selection and hardware configurations. We identify that 60% of PS/Worker workloads can be potentially sped up when ported to the AllReduce architecture exploiting the high-speed NVLink for GPU interconnect, and on average 1.7X speedup can be achieved when Ethernet bandwidth is upgraded from 25 Gbps to 100 Gbps.