Cost Optimization for Serverless Edge Computing with Budget Constraints using Deep Reinforcement Learning

Chen Chen 1 , Peiyuan Guan 2 , Ziru Chen 3 , Amir Taherkordi 2 , Fen Hou 4 and Lin X. Cai 3 · 2025

Read Cost Optimization for Serverless Edge Computing with Budget Constraints using Deep Reinforcement Learning by Chen Chen 1 , Peiyuan Guan 2 , Ziru Chen 3 , Amir Taherkordi 2 , Fen Hou 4 and Lin X. Cai 3 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-12783.txt.

Introduction

Chen Chen and co-authors cut the cost of running serverless functions at the network edge under a fixed budget, in a 2025 paper. Serverless computing charges per use and runs short tasks in containers when events fire, which saves money. But the pay-as-you-go model has no spending cap, so total cost is hard to predict, and many organisations cannot manage their budgets that way.

The scheduling problem, placing functions on varied edge devices under a budget, is too tangled for exact answers in reasonable time. The team propose two online scheduling algorithms based on reinforcement learning, built around traits of serverless functions such as run time and memory. In simulations, their methods matched a commercial exact solver within about three percent, while deciding about one hundred thousand times faster. The authors argue the approach makes budget-capped serverless edge computing practical for real deployments.

Catalog record

CREATORChen Chen 1 , Peiyuan Guan 2 , Ziru Chen 3 , Amir Taherkordi 2 , Fen Hou 4 and Lin X. Cai 3
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE25 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-12783

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Cost Optimization for Serverless Edge Computing with Budget Constraints using Deep Reinforcement Learning Chen Chen 1 , Peiyuan Guan 2 , Ziru Chen 3 , Amir Taherkordi 2 , Fen Hou 4 and Lin X. Cai 3 Abstract Serverless computing adopts a pay-as-you-go billing model where applications are executed in stateless and short-lived containers triggered by events, resulting in a reduction of monetary costs and resource utilization. However, existing platforms do not provide an upper bound for the billing model which makes the overall cost unpredictable, precluding many organizations from managing their budgets. Due to the diverse ranges of serverless functions and the heterogeneous capacity of edge devices, it is challenging to receive near-optimal solutions for deployment cost in a polynomial time. In this paper, we investigated the function scheduling problem with a budget constraint for serverless computing in wireless networks. Users and IoT devices are sending requests to edge nodes, improving the latency perceived by users. We propose two online scheduling algorithms based on reinforcement learning, incorporating several important characteristics of serverless functions. Via extensive simulations, we justify the superiority of the proposed algorithm by comparing with an ILP solver (Midaco). Our results indicate that the proposed algorithms efficiently approximate the results of Midaco within a factor of 1.03 while our decision-making time is 5 orders of magnitude less than that of Midaco. I Introduction The fast advance of artificial intelligence (AI) applications and Internet of Things (IoT) has received significant attention in recent years. However, developers need to spend a large amount of time for the management of virtual machines and servers. Serverless computing, also known as Function-as-a-Service (FaaS) [ 1 ] , has gained great popularity due to its elasticity and simplicity of management. Applications can be divided into multiple functions and deployed as a serverless workflow. The FaaS model, coupled with cloud and edge computing, is beneficial for wireless networks by deploying lightweight functions in the base stations for localized and efficient response to events, and communicating with users and IoT devices [ 2 ] . In this context, edge devices can not only collect raw data but also process data [ 3 ] , improving the response time and Quality of Service by reducing latency. Serverless computing also raises concerns about several performance overheads. Prior work focused some of the major ones, such as cold-start issues [ 4 , 5 , 1 ] , inter-function communication [ 6 ] and state management [ 7 , 8 ] . However, one aspect has been barely investigated, namely the budget of the pay-as-you-use model adopted by serverless computing. For many users, a significant disadvantage of a pure pay-as-you-use billing model is that the overall cost is unpredictable, which precludes many organizations from managing their budgets efficiently. When approving their budget, organizations need to understand how much cost will be incurred by serverless services over a period of time. This raises a legitimate concern, where cloud providers need to cap the overall price similar to the way phone companies offer monthly plans with a capped amount of usage. Our insight is that as serverless computing is increasingly adopted by organizations, it is critical to consider users’ budgets while minimizing the overall deployment cost. To address this shortcoming, this work starts with a thorough formulation of the deployment cost with budget constraints. In particular, we formulate a function scheduling problem, aiming to optimize the deployment cost while offering a budget option to configure in serverless computing. Furthermore, we considered the heterogeneity and connectivity of edge networks, making the problem even more complex. After that, we prove the proposed problem is NP-hard. Due to the computational complexity, we carefully devised two reinforcement learning approaches based on Deep Q Networks ( DQN ) [ 9 ] and Proximal Policy Optimization ( PPO ) [ 10 ] , incorporating several important characteristics of serverless computing which contributes to the state of the art. Our main contributions can be listed as follows. • We formulate a request scheduling problem in edge networks as an Integer Linear Programming (ILP) problem, jointly considering network topology, deployment cost and budget. We proved its NP-hardness. • We carefully design two reinforcement learning-based scheduling policies, incorporating several important features of serverless computing to minimize the deployment cost while meeting fine-grained budget constraints. • We conducted extensive simulations using real-word topology and traces. Experimental results justified that our algorithms approximate the results of an ILP solver Midaco [ 11 ] within a factor of 1.03 while the decision-making time is 5 orders of magnitude less than that of Midaco. The rest of this paper is organized as follows: Section II provides a detailed review of related work, while Section III introduces the system model. In Section IV, we propose two Deep Reinforcement Learning solutions, followed by a performance evaluation in Section V. Finally, Section VI offers concluding remarks. II Related Work Existing work investigates the function scheduling to optimize a number of different objects, including latency, deployment cost and cold starts. Xiao et al. [ 12 ] investigate how to reduce the system cost incurred by caching functions and selecting routes in serverless computing. An online Lazy Caching algorithm is proposed with a worst-case competitive ratio. Li et al. [ 5 ] suggests re-purposing a warm but idle container from another function, aiming to mitigate the cold-start issue. A container management scheme is proposed to schedule intra-function sharing without introducing new security issues. COSE [ 13 ] proposes a statistical learning approach to predict the cost and execution time of serverless functions without configuration information. Thus, COSE optimizes the configuration for running a serverless function. These articles mainly focus on the optimization of deployment cost but overlook the budget constraint and heterogeneity of edge nodes. Chen et al. [ 14 ] uses a probabilistic function caching algorithm inspired by Greedy-Dual-Caching, optimizing the deployment cost in serverless edge computing. Some past work has also focused on data-intensive applications in serverless computing [ 15 , 16 ] . Our work differs from this past work in that we introduce a novel budget-aware problem while fully considering the heterogeneity of edge networks and the nature of serverless computing. Some other work [ 17 , 18 ] minimizes the startup process of serverless containers, resulting in a decrease of end-to-end latency. However, this approach reduces the startup latency but leaves the cold-start mitigation halfway. RainbowCake [ 1 ] uses layer-wise container caching and sharing to not only reduce the startup time but also the occurrence of cold-start issues. Unlike above works only focusing on latency, our work introduces the budget model for organizations to manage their budgets. There are also some work [ 19 , 20 ] focusing on the scheduling of serverless functions while considering the heterogeneity of edge nodes. Nevertheless, these works do not consider the budget model and hence cannot directly apply to our problem. To summarize, this work introduces a fine-grained budget model for serverless computing and considers the heterogeneity and connectivity of edge networks which set our work apart from existing works. III System Model and Problem Formulation In this paper, the deployment cost for serverless invocations consists of three components: the function switching cost, function running cost

Read the complete text