Research map / Graph learning / Graph algorithms and clustering
Efficient Algorithms for Device Placement of DNN Graph Operators
Jakub M Tarnawski, Amar Phanishayee, Nikhil Devanur, Divya Mahajan, Fanny Nina Paravecino
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 |
|---|---|---|---|
| Allow non-contiguous rather than contiguous splits (ablation) | Pipelined inference and training device placement · Evaluated DNN workloads | Throughput: average ∼10%; BERT-3, 20-27% | Improved |
| Allow non-contiguous rather than contiguous splits (ablation) | BERT-3 operator graph, pipelined inference · BERT-3 operator graph, pipelined inference | TPS (time-per-sample): 27.92 -> 21.91 | Improved |
| Allow non-contiguous rather than contiguous splits (ablation) | BERT-3 operator graph, pipelined training · BERT-3 operator graph, pipelined training | TPS (time-per-sample): 65.30 -> 54.21 | Improved |
| Allow non-contiguous rather than contiguous splits (ablation) | ResNet50 operator graph, pipelined inference · ResNet50 operator graph, pipelined inference | TPS (time-per-sample): 124.35 -> 124.35 | No clear effect |
| Use linearization in DP to restrict the search space (ablation) | DPL · BERT-6 operator-graph training | Throughput: 9% loss | Hurt |
| Use linearization in DP to restrict the search space (ablation) | DPL · InceptionV3 layer-graph training | Throughput: 1% loss | Hurt |
| Optimize the original operator graph rather than its contracted layer graph (ablation) | DP contiguous · BERT-12 operator graph, pipelined inference | TPS (time-per-sample): 159.43 -> 147.48; 8% | Improved |
| Optimize latency directly rather than using max-load DP splits (ablation) | Latency IP · Single-query inference workloads | Latency: up to 42% lower; 17% on average | Improved |
Closest papers by meaning
- A Hierarchical Model for Device Placement (ICLR 2018 · similarity 0.88)
- A Structure-Aware Framework for Learning Device Placements on Computation Graphs (NeurIPS 2024 · similarity 0.87)
- Practical Performance Guarantees for Pipelined DNN Inference (ICML 2024 · similarity 0.84)
- Auto-reconfiguration for Latency Minimization in CPU-based DNN Serving (ICML 2025 · similarity 0.84)
- Integrated Hardware Architecture and Device Placement Search (ICML 2024 · similarity 0.84)
- Pfeife: Automatic Pipeline Parallelism for PyTorch (ICML 2025 · 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.
