Optional

Experiments for percom

Batch

Variables:

• Memory: [2G, 1G, 512M]
• Number of arrivals: 150

Experiments:

• 3 networks all small (AgeNet, GenderNet, tiny_yolo)
• 3 networks mixed in size (AgeNet, FaceNet, EmotionNet)
• 3 networks all large (ObjectNet, EmotionNet, MemNet)
• All available networks (AgeNet, GenderNet, tiny_yolo, FaceNet, ObjectNet, EmotionNet, MemNet, SceneNet, SoS_AlexNet)
• 5 networks mixed (realistic scenario, lifelogging): (tiny_yolo, EmotionNet, MemNet, SceneNet, SoS_AlexNet)

Stochastic

Variables:

• Memory: [2G, 1G, 512M]
• Number of arrivals: 150
• Traffic intensity: [Light , normal, Heavy]

Experiments:

• realistic scenario
• Full random arrival
• 3 networks with random compositions

Building on RPI

mkdir -p cmake-build-release
cd ~/edgecaffe/cmake-build-release
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j4
cpack

Installing on RPI

• install EdgeCaffe-1.x.x-Linux-EdgeCaffeApp sudo apt-get install EdgeCaffe-1.3.0-Linux-EdgeCaffeApp
• Make sure the (splitted) models are copied to the pi. scrips/copy_models.sh can be used.
• Make sure that a swapfile exists:

sudo swapoff -a
sudo dd if=/dev/zero of=/swapfile bs=1G count=8
sudo mkswap /swapfile
sudo swapon /swapfile
sudo swapon -a

• Make sure the folders are created and the permissions are correct.

  sudo mkdir -p /opt/analysis/thermal
  sudo chown -R $USER:$USER /opt/edgecaffe
  sudo chown -R $USER:$USER /opt/analysis
  sudo chmod 755  /opt/edgecaffe/thermal/record.sh
  sudo systemctl daemon-reload 

Running experiments on RPI

• Make sure the swapfile is enabled:

sudo swapoff -a && sudo mkswap /swapfile && sudo swapon /swapfile && sudo swapon -a

• Run the experiments:

cd /opt/edgecaffe 
sudo nohup python3 scripts/percom/exp_runner.py --mem-limit 512M|1G|2G --logging true &