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

Maxim Naumov, John Kim+13 more|arXiv (Cornell University)|Mar 20, 2020

Advanced Graph Neural Networks34 references54 citations

TL;DR

The paper presents Zion, Facebook’s scale-up training platform with 8 CPU sockets and 8 accelerators, and discusses design considerations for future scale-out training of deep learning recommendation models (DLRMs) using data- and model-parallelism, along with topology-aware communication and RDMA transports.

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.

Motivation & Objective

• Motivate the growing compute and memory demands of deep learning training in data centers, especially for DLRMs.
• Describe the Zion hardware platform and its memory/compute characteristics.
• Identify design requirements for future scale-out training systems and their impact on interconnects and accelerators.

Proposed method

• Describe DLRM workloads and their mix of dense and sparse features with embedding tables.
• Explain data- and model-parallel training mappings and the use of allreduce and alltoall primitives.
• Present the Zion scale-up design with 8-socket CPUs and 8 accelerators and the Open Accelerator Module (OAM) for vendor-agnostic accelerators.
• Discuss interconnection network choices (CPU fabric, accelerator fabric, PCIe) and their impact on communication patterns.
• Provide an analytical comparison of ring versus fully-connected topologies for allreduce and alltoall to illustrate topology effects on performance.

Experimental results

Research questions

• RQ1 How do data- and model-parallel strategies map to allreduce and alltoall communications in DLRMs?
• RQ2 What are the design implications of a scale-up platform (Zion) for future scale-out training systems?
• RQ3 How do interconnect topology and transport affect synchronization primitives (allreduce/alltoall) in distributed DLRM training?
• RQ4 What is the role of vendor-agnostic accelerator form factors (OAM) in scalable training system design?

Key findings

• DLRMs require both data- and model-parallelism due to embedding tables and dense MLPs; embeddings drive alltoall communication while dense layers drive allreduce.
• Zion provides 8 CPU sockets and 8 accelerators, highlighting the memory/computation trade-offs between CPUs and accelerators.
• Asynchronous training scales throughput nearly linearly with the number of hosts in a job, but synchronous options become necessary as trainer count grows.
• Topology and transport choices (ring vs fully-connected, GDR/RDMA) significantly impact allreduce and alltoall performance, with greater benefits for smaller messages and alltoall bandwidth becoming critical for scale-out.
• The Open Accelerator Module (OAM) enables vendor-agnostic accelerator integration, supporting a flexible, scalable training architecture.