Dynamic Service Placement at the Edge

Edge computing has emerged as a promising solution for delivering services that demand low latency, high bandwidth, and stringent privacy requirements in numerous data- and compute-intensive applications, such as those in Smart Cities. Heterogeneity in edge computing resources and diverse application requirements demand adaptive optimization techniques, such as service placement, to conform to changing conditions. A service placement model must optimize the selection of edge nodes for deploying and executing services, thereby improving application QoS and maximizing resource utilization. Numerous optimization techniques for adaptive service placement problems based on heuristics, meta-heuristics, and reinforcement learning have been proposed, but often evaluated in limited scenarios in terms of scale, complexity, and mobility of the user and the edge servers.  We have three different projects to expand the work in various dimensions by: 

  • Introducing Mobile Edge (ME) servers alongside Static Edge (SE) servers. ME servers can be installed on public transportation such as buses, trains, and trams to offload computation tasks for users within a city, whether they are stationary or on the move. 
  • Using the Multi-Agent Reinforcement Learning (MARL) in decentralized edge computing environment, with multiple edge nodes that operate independently but need to collaborate to optimize overall system performance and the use of available resources.  
  • Exploring the use of Transfer Reinforcement Learning (TRL) that involves transferring the knowledge from a source domain (e.g., a trained model for a specific edge computing scenario) to a target domain (e.g., a new scenario with different conditions) to reduce the time and data required for training in the new environment. This can be particularly useful when deploying service placement strategies across different edge networks or regions. 

The work involves implementing the methodology and conducting comprehensive evaluations through simulations to assess the viability of the model. 

Contact me at Siobhan.Clarke@tcd.ie