LaLaRAND: Flexible Layer-by-Layer CPU/GPU Scheduling for Real-Time DNN Tasks
Author: Woosung Kang, Kilho Lee, Jinkyu Lee, Insik Shin, Hoon Sung Chwa
In 42nd IEEE Real-Time Systems Symposium (RTSS 2021) Dortmund, Germany, December 2021
CUDA: >= 10.2
cuDNN: >=8.0.2
PyTorch: 1.4.0
Python: >= 3.6
CMake: >= 3.10.2
- Install PyTorch with version 1.4.0
- Go to installation directory (probably /home/{username}/.local/lib/python{version}/site-packages/torch}
- Replace directory nn, quantization.
- Run scheduler before DNN tasks
- Provide resource configuration of DNN tasks by txt file (current: /tmp/{pid of task}.txt)
- Before inference code,
- Call {model}.set_rt() to set rt-priority of task
- Call {model}.hetero() to use heterogeous resource allocation
- hetero() requires inference function and sample inputs for input calibaration