Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Maxim Naumov, John Kim, Dheevatsa Mudigere, Srinivas Sridharan, Xiaodong Wang, Whitney Zhao, Serhat Yilmaz, Changkyu Kim, Hector Yuen, Mustafa Ozdal, Krishnakumar Nair, Isabel Gao, Bor-Yiing Su, Jiyan Yang, Mikhail Smelyanskiy

TL;DR

The paper addresses the rising compute and memory demands of training DL models at Facebook, with a focus on deep learning recommendation models (DLRMs). It introduces Zion, a scale-up platform that decouples memory, compute, and networking to deliver large memory and accelerator-enabled throughput, and it analyzes design choices for future scale-out systems. It covers the mapping of DLRMs to data- and model-parallel training, the communication primitives (allreduce/alltoall), and how interconnect topology and transport influence performance, including an analytical comparison of ring versus fully connected topologies. The discussion extends to scale-out fabric topologies, RDMA-based transports, and topology-aware collectives, comparing Zion to existing scale-out systems and outlining practical considerations for future accelerator-centric infrastructures.

Abstract

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.

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

TL;DR

The paper addresses the rising compute and memory demands of training DL models at Facebook, with a focus on deep learning recommendation models (DLRMs). It introduces Zion, a scale-up platform that decouples memory, compute, and networking to deliver large memory and accelerator-enabled throughput, and it analyzes design choices for future scale-out systems. It covers the mapping of DLRMs to data- and model-parallel training, the communication primitives (allreduce/alltoall), and how interconnect topology and transport influence performance, including an analytical comparison of ring versus fully connected topologies. The discussion extends to scale-out fabric topologies, RDMA-based transports, and topology-aware collectives, comparing Zion to existing scale-out systems and outlining practical considerations for future accelerator-centric infrastructures.

Abstract

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.

Paper Structure

This paper contains 20 sections, 2 equations, 14 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Recommendation Model
  4. Increase in Complexity
  5. Distributed Training
  6. Training at Facebook
  7. Overview of Training at Facebook
  8. Scalability & Challenges
  9. Interconnection Networks
  10. Zion Scale-up Training
  11. Overview
  12. Accelerator Fabric Topology
  13. Communication Pattern
  14. Hardware Design Flexibility
  15. Discussion on Scale-out Training
  16. ...and 5 more sections

Figures (14)

  • Figure 1: (a) Server compute demand for training and (b) number of distributed training workflows across Facebook data centers.
  • Figure 2: High-level overview of DLRM.
  • Figure 3: Increase in model complexity over time.
  • Figure 4: Communication patterns that are common in (a) data- and (b) model-parallelism. Both communication patterns need to be supported in DLRM.
  • Figure 5: Performance as the number of trainers are increased.
  • ...and 9 more figures