Microservice Allocation With RL-Guided DVFS for Edge Latency Budgets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for optimizing energy consumption and latency in multi-access edge computing environments with microservices are inefficient, particularly due to the NP-Hard nature of integer linear programming problems and the lack of effective solutions for server-DVFS allocation that satisfy energy and latency budgets over varying time scales.
Innovation Solution
The use of reinforcement learning to overapproximate or underapproximate parameter bounds, combined with integer linear programming and dynamic voltage frequency scaling, to determine optimal server-DVFS allocations for microservices, while employing a reinforcement learning agent to assign rewards for feasible solutions and iteratively improve the allocation process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by stationary object
If integer linear programming is used to optimize server-DVFS allocation for microservices, then energy consumption can be minimized and latency can be controlled, but the computational complexity becomes NP-Hard making it inefficient for real-time decisions
Solution Approach 1:
The patent transforms the NP-Hard integer linear programming problem into a polynomial-time solvable linear programming problem by relaxing the integer constraints on DVFS scaling factors. This allows continuous parameter optimization while maintaining energy minimization and latency control objectives, resolving the contradiction between optimization effectiveness and computational efficiency
Solution Approach 2:
The patent employs reinforcement learning agents that make lightweight, approximate allocation decisions instead of solving complex ILP problems repeatedly. These agents use learned policies to quickly determine server-DVFS allocations without exhaustive computation, sacrificing some optimality for real-time responsiveness and reduced computational burden
2Reliability
If traditional optimization methods are used for microservice allocation, then solutions can satisfy energy and latency budgets, but the methods are inefficient and cannot handle varying time scales effectively
Solution Approach 1:
The patent implements dynamic allocation strategies where reinforcement learning agents continuously adapt server-DVFS assignments based on changing workload conditions and time scales. The system transitions from static ILP solutions to dynamic, learned policies that respond to real-time variations in microservice demands while maintaining budget constraints
Solution Approach 2:
The patent incorporates feedback mechanisms where reinforcement learning agents receive rewards or penalties based on whether allocation decisions satisfy energy and latency budgets. This feedback loop enables iterative improvement of allocation policies, ensuring budget satisfaction while enhancing optimization efficiency through learned experience
Data Source
AI summary
A plurality of edge computing nodes are provided in a multi-access edge computing environment. Operations are performed to ensure that energy consumption of edge computing nodes is minimized and a latency of serving requests is lower than a threshold by overapproximating or underapproximating parameter bounds or budgets; and by using reinforcement learning discrete actions to determine whether to overapproximate or underapproximate in an integer linear programming (ILP) solution.