Research map / Graph learning / Graph algorithms and clustering
DOPPLER: Dual-Policy Learning for Device Assignment in Asynchronous Dataflow Graphs
Xinyu Yao, Daniel Bourgeois, Abhinav Jain, Yuxin Tang, Jiawen Yao, Zhimin Ding, Arlei Silva, Chris Jermaine
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 |
|---|---|---|---|
| add both SEL and PLC learned policies (full Doppler-sys) instead of replacing one with a Critical Path heuristic (ablation) | Doppler · FFNN (4 GPUs) | execution time (ms): 63.2 -> 47.4 (vs PLC), 59.1 -> 47.4 (vs SEL) | Improved |
| add both learned policies instead of replacing one with heuristic (ablation) | Doppler · ChainMM (4 GPUs) | execution time (ms): 121.6 (PLC) vs 123.4 (SYS) | No clear effect |
| add both learned policies instead of replacing one with heuristic (ablation) | Doppler · Llama-layer (4 GPUs) | execution time (ms): 159.5 -> 150.6, 161.7 -> 150.6 | Improved |
| add Stage III real-system RL on top of imitation + simulation stages (ablation) | Doppler · Llama-block (4 GPUs) | execution time (ms): 191.5 -> 160.3 | Improved |
| add Stage III real-system RL on top of imitation + simulation stages (ablation) | Doppler · ChainMM (4 GPUs) | execution time (ms): 122.5 -> 123.4 | No clear effect |
| add Stage III real-system RL on top of imitation + simulation stages (ablation) | Doppler · Llama-layer (4 GPUs) | execution time (ms): 167.0 -> 150.6 | Improved |
| fine-tune a transferred policy for 4k episodes instead of zero-shot (ablation) | Doppler transfer FFNN->Llama-block · Llama-block (4 GPUs) | execution time (ms): 251.0 -> 159.4 | Improved |
| fine-tune a transferred policy for 4k episodes instead of zero-shot (ablation) | Doppler transfer ChainMM->Llama-layer · Llama-layer (4 GPUs) | execution time (ms): 338.2 -> 156.4 | Improved |
| message passing once per episode instead of once per MDP step (ablation) | Doppler-sim · ChainMM (simulator) | execution time (ms): 121.7 -> 122.5 | No clear effect |
| pretrain with imitation learning and simulation before real-system RL (ablation) | Doppler-sys · Llama-layer | execution time convergence | Improved |
Closest papers by meaning
- A Structure-Aware Framework for Learning Device Placements on Computation Graphs (NeurIPS 2024 · similarity 0.84)
- SustainDC: Benchmarking for Sustainable Data Center Control (NeurIPS 2024 · similarity 0.83)
- Learning Multimodal One-step Flow Policy via Value-weighted Optimal Transport (NeurIPS 2026 · similarity 0.82)
- DORA: A Scalable Asynchronous Reinforcement Learning System for Language Model Training (NeurIPS 2026 · similarity 0.81)
- FASTER: Rethinking Real-Time Flow VLAs (NeurIPS 2026 · similarity 0.81)
- Beyond Low-rank Decomposition: A Shortcut Approach for Efficient On-Device Learning (ICML 2025 · 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.
