DNNSplit: Latency and Cost-Efficient Split Point Identification for Multi-Tier DNN Partitioning
Bibliographic record
Abstract
Due to the high computational demands inherent in Deep Neural Network (DNN) executions, multi-tier environments have emerged as preferred platforms for DNN inference tasks. Previous research on partitioning strategies for DNN models typically involved leveraging all layers of the DNN to identify optimal splits aimed at reducing latency or cost. However, due to their computational complexity, these approaches face scalability issues, particularly with models containing hundreds of layers. The novelty of ourwork lies in uniquely identifying specific split points within variousDNNmodels that consistently lead to efficient latency or cost partitioning. Under the assumption that per unit computing cost decreases in higher tiers and that bandwidth is not free, we show that only these specific split points need to be considered to optimize latency or cost. Importantly, these split points are independent of different infrastructure configurations and bandwidth variations. The key contribution of our work is the significant reduction in the computational complexity of DNN partitioning, making our strategy applicable to models with a large number of layers. IntroducingDNNSplit, an adaptive strategy, enables dynamic split decisions in varying conditions with the least complexity. Evaluated across nine DNN models varying in size and architecture,DNNSplitexhibits exceptional effectiveness in optimizing latency and cost. Even for a more substantial model containing 517 layers, it identifies only 5 points as potential split points, thereby reducing the partitioning complexity by more than 100x. This makes DNNSplit especially advantageous for managing larger models.DNNSplitalso demonstrates significant improvements for multi-tier deployments compared to single-tier execution, including up to 15x latency speedup, 20x cost reduction, and 5x throughput enhancement.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.004 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.005 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".