Research map / Reinforcement learning / Visual and representation RL
DreamShard: Generalizable Embedding Table Placement for Recommender Systems
Daochen Zha, Louis Feng, Qiaoyu Tan, Zirui Liu, Kwei-Herng Lai, Bhargav Bhushanam, Yuandong Tian, Arun Kejariwal, Xia Hu
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 |
|---|---|---|---|
| remove pooling factor feature (ablation) | DreamShard · DLRM-50 (4) test | overall cost (ms): 40.4 -> 47.2 | Hurt |
| remove cost features from state (ablation) | DreamShard · DLRM-50 (4) test | overall cost (ms): 40.4 -> 46.3 | Hurt |
| remove dimension feature (ablation) | DreamShard · DLRM-50 (4) test | overall cost (ms): 40.4 -> 40.9 | Hurt |
| remove hash size feature (ablation) | DreamShard · DLRM-50 (4) test | overall cost (ms): 40.4 -> 40.6 | No clear effect |
| remove table size feature (ablation) | DreamShard · DLRM-50 (4) test | overall cost (ms): 40.4 -> 40.6 | No clear effect |
| remove distribution feature (ablation) | DreamShard · DLRM-50 (4) test | overall cost (ms): 40.4 -> 40.5 | No clear effect |
| add RNN on device representation in policy (ablation) | DreamShard · DLRM-50 (4) test | overall cost (ms): 40.4 -> 40.5 | No clear effect |
| apply DreamShard trained on DLRM-20 (4) to DLRM-100 (4) without fine-tuning | DreamShard · DLRM-100 (4) test | overall cost (ms): 77.8 -> 77.9 | No clear effect |
| apply DreamShard trained on DLRM-100 (4) to DLRM-20 (4) | DreamShard · DLRM-20 (4) test | overall cost (ms): 17.6 -> 17.7 | No clear effect |
| apply DreamShard trained on DLRM-20 (2) to DLRM-20 (4) | DreamShard · DLRM-20 (4) test | overall cost (ms): 17.6 -> 17.8 | No clear effect |
| use DreamShard instead of Lookup-based expert on large 8-GPU task | placement · DLRM-200 (8) test | overall cost (ms): 38.0 -> 36.4 | Improved |
| use DreamShard instead of Dim-based expert on hardest Prod task | placement · Prod-80 (8) test | overall cost (ms): 41.9 -> 35.2 | Improved |
Closest papers by meaning
- A Structure-Aware Framework for Learning Device Placements on Computation Graphs (NeurIPS 2024 · similarity 0.85)
- Graph-enhanced Optimizers for Structure-aware Recommendation Embedding Evolution (NeurIPS 2024 · similarity 0.85)
- On the Unexpected Effectiveness of Reinforcement Learning for Sequential Recommendation (ICML 2024 · similarity 0.81)
- Learning Representations for Hierarchies with Minimal Support (NeurIPS 2024 · similarity 0.81)
- Towards characterizing the value of edge embeddings in Graph Neural Networks (ICML 2025 · similarity 0.81)
- SiT: Symmetry-invariant Transformers for Generalisation in Reinforcement Learning (ICML 2024 · similarity 0.81)
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.
