Deep Learning Training in Facebook Data Centers: Design of Scale-up and Scale-out Systems

By Maxim Naumov, John Kim, Dheevatsa Mudigere, Srinivas Sridharan, Xiaodong Wang, Whitney Zhao, Serhat Yilmaz, Changkyu Kim, Hector Yuen, Mustafa Ozdal and 6 more. 2020

107 citations recorded by the source corpus.

Large-scale training is important to ensure high performance and accuracy of machine-learning models. At Facebook we use many different models, including computer vision, video and language models. However, in this paper we focus on the deep learning recommendation models (DLRMs), which are responsible for more than 50% of the training demand in our data centers. Recommendation models present unique challenges in training because they exercise not only compute but also memory capacity as well as memory and network bandwidth. As model size and complexity increase, efficiently scaling training becomes a challenge. To address it we design Zion - Facebook's next-generation large-memory training platform that consists of both CPUs and accelerators. Also, we discuss the design requirements of future scale-out training systems.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Efficiency & Compression, Optimization & Training

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Efficiency & Compression

In Optimization & Training

Citations in the corpus

Pathways: Asynchronous Distributed Dataflow for ML (2022, methodology): “D mesh tori (Naumov et al., 2020).”
Chasing Carbon: The Elusive Environmental Footprint of Computing (2020, background): “In less than two years, Facebook hardware devoted to AI training and inference has grown by 4 × and 3.5 × , respectively [22], [23].”
RecSSD: near data processing for solid state drive based recommendation inference (2021, background): “Recent work shows that deep-learning based recommendation systems not only drive product success[13, 37, 40] but also dominate the datacenter capacity for AI training and inference [20, 21, 29].”
Understanding Training Efficiency of Deep Learning Recommendation Models at Scale (2020, background): “month period, we have witnessed the compute capacity for recommendation model training quadrupling at Facebook’s datacenter fleet [41].”
This paper is included in the Proceedings of the 19th USENIX Symposium on Networked Systems Design and Implementation. (background): “Recommendation systems are often used in commercial settings and dominate the datacenter capacity for AI training and inference [22].”
RecShard: statistical feature-based memory optimization for industry-scale neural recommendation (2022, background): “For example, at Facebook, DLRMs account for more than 50% of training demand [32] and more than 80% of inference demand [11].”
High-performance, Distributed Training of Large-scale Deep Learning Recommendation Models (2021, methodology): “Traditionallyadisaggregatedparameter-server(PS)baseddistributed CPUtrainingsystemhasbeenusedfortrainingDLRMsinapro-ductionsetting[13,38].”
TT-Rec: Tensor Train Compression for Deep Learning Recommendation Models (2021, background): “These models represent one of the most resource-demanding deep learning workloads, consuming more than 50% of training and 80% of the total AI inference cycles at Facebook’s data centers (Gupta et al., 2019; Naumov et al., 2020).”
Software-hardware co-design for fast and scalable training of deep learning recommendation models (2021, methodology): “This hybrid parallelism requires AlltoAll communications for the embedding lookup results [42, 43], as well as embedding table input redistribution if the inputs are streamed from database in batches, which is often the case.”
Understanding Capacity-Driven Scale-Out Neural Recommendation Inference (2020, background): “Separate works have described the challenges of scaling training for large models, but they either don’t address deep recommendation systems or don’t explore the inference serving environment that is the focus of our paper [12], [14], [15], [36]–[38].”