Primus: Unified Training System for Large-Scale Deep Learning Recommendation Models

The scale of deep learning recommendation models (DLRM) continues to grow, demanding increasingly vast computing and storage resources. In production environments, improving training efficiency and effectiveness has become the primary goal to meet the needs of numerous model training jobs under resource limitations. We introduce Primus, a unified training system that unifies the training resources, data, and paradigms to support high-performance DLRM training at ByteDance. Specifically, ① Primus provides a unified abstraction of resources and interoperates with multiple scheduling systems, achieving a consistent training experience with horizontal and vertical dynamic scaling strategies across resource pools. ② Primus offers a unified three-tier data definition and employs a data task graph generation approach to support data orchestration of multi-source training samples composed of batch and stream data. ③ Primus devises a new hybrid training paradigm for DLRMs that ensures high model timeliness by controlling parameter updates and applying fine-grained prioritization of mixed batch and stream data.

Primus has demonstrated its efficiency and effectiveness in handling large-scale, enterprise-grade DLRM training over five years of deployment at ByteDance. Evaluations show Primus's optimizations of resources, data, and paradigms. Firstly, dynamic scaling reduces training cost by 17.1% at the cluster level and increases CPU utilization from 50% to 80% per job. Secondly, data orchestration accelerates task generation by 23× and achieves higher training throughput. Lastly, after applying the hybrid training paradigm with 4 different DLRMs, advertising revenue increases by 0.4%-2.4%.