GS-LinkHQ

GS-LinkHQ is a reinforcement learning based agent that generates policy for both computation offloading and distributed caching in the GEdge-Platform. The policy supports horizontal (between cloud edge and neighbor cloud edge) and vertical (between cloud edge and core cloud) collaboration.

The policy for computaion offloading

• Providing offloading service developers (or service providers) with edge resource allocation policies and capabilities required to implement policies for offloading service development
  • Supporting offloading service provisioning: Optimized resource allocation
• Resource allocation policy
  • Where to deploy
  • How to deploy
  • Collaborative deployment or Noncollaborative deployment
    • Horizontal, vertical and horizontal + vertical collaboration
• Policy provide point
  • Service development point
  • When quality of service decreases → When scale out or scale up
  • Choosing load balancing method

The information considered to decide the policy is as follows.

• Offloading service definition
  1. CPU specification
  2. GPU usage and specification
  3. Storage/cache specification
  4. Service properties (network delay sensitive, availability sensitive, etc.)
• Cloud-edge (include neighbor edge's information) system resource status information
  1. CPU usage
  2. GPU usage
  3. Storage/cache usage
  4. Deployed offloading services quality

Components

GS-LinkHQ consists of three components

• vedge: Develop proposed methods first with simulator
  • Collaborative multi-cluster environment
  • Include resource map, network performance map, and offloading service(task) definition
• agent: Reinforcement learning agent developed with the above two projects
  • Real-world (GEdge-Platform) environment

Each componets communicates through RESTful API

Requirements

• Nvidia Docker
• Docker compose

Usage

1. Clone repository

git clone https://github.com/gedge-platform/gs-linkhq.git

2. Edit hyperparameters (Optional)

# docker-compose.yml
version: "3.9"
services:
  ...
  agent:
    ...
    environment:
      ENV_ADDRESS: vedge
      ENV_PORT: 80
      LOG_LEVEL: ${LOG_LEVEL:-INFO}
      MEM_CAPACITY: 100000
      BATCH_SIZE: 1000 
      GAMMA: 0.99
      EPS_START: 0.9
      EPS_END: 0.00001
      EPS_DECAY: 20000
      TAU: 0.005
      LR: 0.0001
    ...

3. Build images and run containers (Training Model)

docker compose up --build

4. Start Recommendation Server

cd srv
docker build --tag linkhq:4.0 .
docker run -p 80:80 linkhq:4.0