Disaggregated Multi-Tower: Topology-aware Modeling Technique for Efficient Large-Scale Recommendation
This addresses performance bottlenecks for large-scale recommendation systems in data centers, representing a novel method for a known bottleneck.
The paper tackles inefficiencies in large-scale recommendation systems due to mismatches between model architecture and data center topology, proposing Disaggregated Multi-Tower (DMT) to achieve up to 1.9x speedup without accuracy loss.
We study a mismatch between the deep learning recommendation models' flat architecture, common distributed training paradigm and hierarchical data center topology. To address the associated inefficiencies, we propose Disaggregated Multi-Tower (DMT), a modeling technique that consists of (1) Semantic-preserving Tower Transform (SPTT), a novel training paradigm that decomposes the monolithic global embedding lookup process into disjoint towers to exploit data center locality; (2) Tower Module (TM), a synergistic dense component attached to each tower to reduce model complexity and communication volume through hierarchical feature interaction; and (3) Tower Partitioner (TP), a feature partitioner to systematically create towers with meaningful feature interactions and load balanced assignments to preserve model quality and training throughput via learned embeddings. We show that DMT can achieve up to 1.9x speedup compared to the state-of-the-art baselines without losing accuracy across multiple generations of hardware at large data center scales.