Research map / Optimization / Distributed and compressed
GBA: A Tuning-free Approach to Switch between Synchronous and Asynchronous Training for Recommendation Models
Wenbo Su, Yuanxing Zhang, Yufeng Cai, Kaixu Ren, Pengjie Wang, Huimin Yi, Yue Song, Jing Chen, Hongbo Deng, Jian Xu, Lin Qu, Bo Zheng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GBA instead of the best baseline (Hop-BW) when switching from synchronous training | GBA · Criteo, Alimama, Private | average AUC: at least 0.2% on average | Improved |
| Use GBA instead of other modes when switching to synchronous training | GBA · three tasks | AUC: at least 0.1% | Improved |
| Use GBA instead of synchronous training under strained hardware | GBA · three tasks | training speed: at least 2.4x | Improved |
| Use GBA instead of synchronous AR training | GBA · Criteo | global QPS: 1,436K($\pm$224K) -> 3,240K($\pm$97K) | Improved |
| Use GBA instead of synchronous AR training | GBA · Private | global QPS: 43K($\pm$21K) -> 87K($\pm$19K) | Improved |
| Use GBA instead of canonical asynchronous training | GBA · Criteo | global QPS: 3,253K($\pm$84K) -> 3,240K($\pm$97K) | No clear effect |
| Use GBA instead of Hop-BS | GBA · Private | global QPS: 29K($\pm$11K) -> 87K($\pm$19K) | Improved |
| Use GBA instead of Hop-BW (backup workers) | GBA · Private | excluded batches: 300K -> 786 | Improved |
| Use GBA instead of BSP for gradient aggregation | GBA · Private | avg gradient staleness: 2.61 (12) -> 0.21 (11) | Improved |
| Run GBA in different cluster periods of the day | GBA · Private | AUC: 0.7864 -> 0.7866 | No clear effect |
| Vary number of workers from 100 to 800 while keeping global batch size (ablation) | GBA · Private | AUC and QPS: AUC difference less than 10^-4 | Mixed |
| Use a global batch size different from synchronous training (400 workers, vary local batch size) (ablation) | GBA · Private | AUC after switching | Hurt |
Closest papers by meaning
- Asynchronous Stochastic Gradient Descent for Extreme-Scale Recommender Systems (AAAI 2021 · similarity 0.88)
- Infinite Recommendation Networks: A Data-Centric Approach (NeurIPS 2022 · similarity 0.85)
- KuaiRecV2: Benchmarking Large-Scale Continual Learning for Diversified and Multi-task Recommendation (NeurIPS 2026 · similarity 0.85)
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.85)
- ConvRec-R1: Training LLM-based Conversational Recommender Systems with Reinforcement Learning (ICLR 2026 · similarity 0.83)
- LIME: Link-based User-item Interaction Modeling with Decoupled XOR Attention for Efficient Test Time Scaling (NeurIPS 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
