PhD Defense: “Joint Optimization of Pricing and Resource Allocation in Serverless Edge Computing: A Game-Theoretic Perspective”

TECoSA PhD student Feridun Tütüncüoğlu will defend his thesis at F3, KTH Campus. Contact György Dán (gyuri@kth.se) if you are interested in attending.

Abstract: The rapid advancement of Internet of Things (IoT), Augmented Reality (AR), autonomous systems, and intelligent automation is transforming daily life and revolutionizing industrial processes. These technologies demand significant computational resources while also imposing stringent latency requirements. A common approach to meet computational resource demand is leveraging Cloud Computing (CC), which offers scalable processing capabilities through centralized data centers. Nonetheless, this centralized approach often fails to meet stringent latency requirements due to communication delays caused by the geographical distance between cloud servers and end users. This limitation has led to the emergence of the novel paradigm of Edge Computing (EC), which addresses the latency issue by placing compute units closer to end users.

EC server clusters are expected to be smaller in scale and geographically more dispersed compared to CC data centers. This introduces new challenges, including limited computational and storage capacity, making efficient resource allocation crucial. Additionally, the network operator managing edge infrastructure must maintain the financial sustainability under different workload characteristics and application requirements of users, necessitating joint and adaptable resource management and pricing strategies. Function-as-a-Service (FaaS) offers a promising approach in this regard, as its pay-as-you-go pricing model allows users to pay only for the resources they consume, while also enabling dynamic resource management by shifting the entire responsibility of application deployment to the operator. However, this flexibility also makes the choice of computation, memory, and bandwidth resources price-dependent, further complicating resource management and pricing.

In the first part of the thesis, we consider a setting where Wireless Devices (WDs) minimize energy and the monetary costs of computing tasks, while the operator maximizes revenue by optimizing pricing and application caching under memory constraints. We consider a dynamic setting where the operator has no prior knowledge of the varying availability of WDs over time. We model this interaction as a Stackelberg Game (SG) and demonstrate the existence of an equilibrium. To address information asymmetry, we use Bayesian optimization to learn pricing strategies, establish an upper bound on its asymptotic regret, and propose a greedy approximation algorithm for application caching. We then investigate the joint optimization of compute, communication, and memory resources in a static network setting, where WDs minimize the costs of executing the tasks of their applications, including monetary and energy expenses. We model this interaction as a SG, show the existence of an equilibrium, and prove that computing an equilibrium is NP-hard. We propose an efficient approximation algorithm with a bounded approximation ratio. An interesting feature of our solution is that the operator’s revenue is maximized when the WDs maximize their energy savings through computation offloading. Furthermore, we investigate the rate-adaptation problem, where WDs adjust their offloading rates based on available compute resources and pricing. We model the interaction as a SG and propose a Stackelberg gradient play algorithm that computes the operator’s implicit revenue function with respect to rate selection of the WDs.

The second part of this thesis explores a dynamic network and pricing setting where WDs arrive at the edge cell according to a non-homogeneous stochastic process, and the operator sets prices based on the availability of WDs and their heterogeneous workload characteristics. We formulate the problem of maximizing the revenue of the operator as a sequential decision making problem under uncertainty, where the operator’s price can be piecewise linear or non-linear and could vary over time. In a Markovian steady-state setting, we derive analytical results for the optimal pricing strategy, which also serve as a heuristic for the general case. To address the general case, we introduce a Generalized Hidden Parameter Markov Decision Process and propose a dual Bayesian neural network approximator that approximates the state transitions and the revenue to accelerate the learning of the optimal pricing policy. This approach enables pre-training on synthetic traces while adapting quickly to unseen workload patterns.

The third part addresses overlooked computational challenges by examining the impact of server contention on both operator revenue and application latency constraints. To address this, we propose a contention model validated through experiments across applications with varying compute demands, including L1/L2/L3 caches, I/O, and memory bus usage. We develop a novel model-based Bayesian optimization algorithm to maximize operator revenue while ensuring that latency and resource capacity constraints are met.

The algorithmic contributions in the area of pricing and resource management are intended to serve as efficient, deployable, and scalable solutions that enhance the robustness and efficiency of resource allocation and pricing in EC.

Details of the panel are shown below:

Respondent: Feridun Tütüncüoğlu

Supervisors:  György Dan, James Gross