[Paper Review] Optimizing AI Service Placement and Resource Allocation in Mobile Edge Intelligence Systems

Zehong Lin, Suzhi Bi+1 more|arXiv (Cornell University)|Nov 11, 2020

IoT and Edge/Fog Computing39 references4 citations

TL;DR

This paper proposes an optimization framework for AI service placement and resource allocation in mobile edge intelligence systems, jointly minimizing computation time and energy by strategically placing AI programs on users and allocating CPU, bandwidth, and edge computing resources. The method uses analytical resource allocation and scalable ADMM-based decomposition, achieving near-optimal performance with linear complexity in large networks.

ABSTRACT

Leveraging recent advances on mobile edge computing (MEC), edge intelligence has emerged as a promising paradigm to support mobile artificial intelligence (AI) applications at the network edge. In this paper, we consider the AI service placement problem in a multi-user MEC system, where the access point (AP) places the most up-to-date AI program at user devices to enable local computing/task execution at the user side. To fully utilize the stringent wireless spectrum and edge computing resources, the AP sends the AI service program to a user only when enabling local computing at the user yields a better system performance. We formulate a mixed-integer non-linear programming (MINLP) problem to minimize the total computation time and energy consumption of all users by jointly optimizing the service placement (i.e., which users to receive the program) and resource allocation (on local CPU frequencies, uplink bandwidth, and edge CPU frequency). To tackle the MINLP problem, we derive analytical expressions to calculate the optimal resource allocation decisions with low complexity. This allows us to efficiently obtain the optimal service placement solution by search-based algorithms such as meta-heuristic or greedy search algorithms. To enhance the algorithm scalability in large-sized networks, we further propose an ADMM (alternating direction method of multipliers) based method to decompose the optimization problem into parallel tractable MINLP subproblems. The ADMM method eliminates the need of searching in a high-dimensional space for service placement decisions and thus has a low computational complexity that grows linearly with the number of users. Simulation results show that the proposed algorithms perform extremely close to the optimum and significantly outperform the other representative benchmark algorithms.

Motivation & Objective

• Address the challenge of dynamic AI model updates in mobile edge computing, where outdated models degrade inference performance due to concept drift.
• Minimize total computation time and energy consumption in a multi-user MEC system by jointly optimizing service placement and resource allocation.
• Enable local AI inference on user devices by selectively placing updated AI programs only when beneficial, avoiding unnecessary offloading.
• Design a scalable solution for large networks by decomposing the high-dimensional service placement problem using ADMM.
• Overcome the limitations of prior works that assume pre-existing AI programs at all devices, by modeling program dissemination as part of the optimization.

Proposed method

• Formulates a mixed-integer non-linear programming (MINLP) problem to jointly optimize service placement, local CPU frequency, uplink bandwidth, and edge CPU frequency.
• Derives analytical expressions for optimal resource allocation under a given service placement, enabling fast evaluation of candidate placements.
• Employs search-based algorithms (e.g., greedy or meta-heuristic) to find optimal service placement decisions using the derived analytical solutions.
• Proposes an ADMM-based decomposition method to split the problem into per-user subproblems, reducing computational complexity and enabling parallel processing.
• Uses ADMM to decouple the high-dimensional service placement decision from resource allocation, achieving linear growth in complexity with network size.
• Introduces a heuristic scheme based on uplink channel gains for homogeneous networks, showing near-optimal performance in special cases.

Experimental results

Research questions

• RQ1 How can AI service placement and resource allocation be jointly optimized to minimize total computation time and energy in a multi-user MEC system?
• RQ2 What is the impact of dynamic AI model updates and concept drift on system performance, and how can it be mitigated through intelligent program placement?
• RQ3 How can the computational complexity of service placement be reduced in large-scale networks without sacrificing optimality?
• RQ4 In what scenarios is local AI inference on user devices more efficient than full offloading to the edge server?
• RQ5 How does the system performance vary with network size, task data size, and workload, and what are the scalability trade-offs?

Key findings

• The proposed algorithms achieve performance extremely close to the optimal solution, significantly outperforming benchmark methods in simulation.
• The ADMM-based method exhibits linear computational complexity growth with the number of users, making it highly scalable for large networks.
• The greedy search method is effective for small networks, while the ADMM-based approach scales much better due to parallelizable subproblems.
• In homogeneous networks with varying channel gains, placing the AI program on users with poor uplink channels improves spectrum efficiency and performance.
• The uplink-based heuristic scheme achieves near-optimal performance in special cases where users differ only in channel quality.
• Performance gains are most significant when task data size, computation time weighting factor, or number of users increases, but diminish as workload becomes very high.