AttentionFunc: Balancing FaaS Compute across Edge-Cloud Continuum with Reinforcement Learning

Abstract

Serverless computing is emerging as a promising paradigm to manage compute in Edge-Cloud continuum. However, distributing and balancing the computational load (serverless functions) across the continuum remains a significant challenge. In this paper, we introduce AttentionFunc-a novel framework for decentralized and efficient function offloading and computation balancing in the Edge-Cloud continuum. The AttentionFunc framework strives to introduce a fully decentralized decision-making model that accounts for the multi-objective nature of serverless workflows, the limitations of shared resources in the Edge-Cloud environment, and the dynamic behaviors such as resource contentions or cooperations among serverless functions. In addition, AttentionFunc incorporates an innovative multi-Agent offloading model based on the Markov Decision Process (MDP), designed to minimize functions' execution time and costs. The application of MDP allows the framework to efficiently address these issues using deep reinforcement learning approaches, with an aim to significantly improve function completion latency. Furthermore, AttentionFunc pioneers an attention-based optimization mechanism for multi-Agent deep reinforcement learning. This mechanism permits DRL agents to reach a consensus with minimal coordination information, leading to substantial reductions in communication and computation overhead. We evaluate AttentionFunc and compare it against select relevant state-of-The-Art approaches. Our experiments and simulations show that AttentionFunc outperforms state-of-The-Art approaches in terms of 1) the completion latency (up to 44.2% reduction), 2) the function success rate (up to 43.3% increase). Additionally, we provide the results of many experiments with different MEC scenarios to highlight the components of our approach that influence the results. We conclude that our approach reduces the low-latency challenge faced by most offloading models and improves the successful completion rate of the function.