Research Article
An Adaptive Task Migration Scheduling Approach for Edge-Cloud Collaborative Inference
Table 1
The total inference delay of several typical DNN models under different scheduling methods (seconds).
| DNN | 3G | 4G | C-O1 | E-O2 | S-C3 | S-Q4 | C-O1 | E-O2 | S-C3 | S-Q4 |
| AlexNet | 78.3 | 88.9 | 43.7 | 43.7 | 28.6 | 88.9 | 27.1 | 24.7 | Tiny-YOLO | 200.8 | 159.7 | 59.9 | 59.9 | 84.2 | 159.7 | 58.6 | 58.6 | DarkNet19 | 77.8 | 180.4 | 54.6 | 54.6 | 34.9 | 180.4 | 34.9 | 32.6 | AlexNet-Parallel | 70.2 | 44.9 | 40.6 | — | 27.9 | 44.9 | 27.9 | — | ResNet18 | 84.3 | 81.7 | 32.0 | — | 34.4 | 81.7 | 25.5 | — | GoogLeNet | 91.2 | 121.1 | 63.3 | — | 43.2 | 121.1 | 43.2 | — |
|
|
1Cloud-only, 2edge-only, 3Single Cutting, and 4Scheduling with Queuing.
|