[Paper Review] Optimal Response to Burstable Billing under Demand Uncertainty

Yong Zhan, Mahdi Ghamkhari+3 more|arXiv (Cornell University)|Mar 18, 2016

Supply Chain and Inventory Management17 references3 citations

TL;DR

This paper proposes an optimization-based method for users to minimize bandwidth costs and maximize surplus under burstable billing with demand uncertainty. By deriving a tractable model for 95th percentile usage, it formulates a stochastic bandwidth allocation problem and demonstrates a 26% cost reduction and 23% surplus increase over on-demand allocation using real workload traces.

ABSTRACT

Burstable billing is widely adopted in practice, e.g., by colocation data center providers, to charge for their users, e.g., data centers, for data transferring. However, there is still a lack of research on what the best way is for a user to manage its workload in response to burstable billing. To overcome this shortcoming, we propose a novel method to optimally respond to burstable billing under demand uncertainty. First, we develop a tractable mathematical expression to calculate the 95th percentile usage of a user, who is charged by provider via burstable billing for bandwidth usage. This model is then used to formulate a new bandwidth allocation problem to maximize the user's surplus, i.e., its net utility minus cost. Additionally, we examine different non-convex solution methods for the formulated stochastic optimization problem. We also extend our design to the case where a user can receive service from multiple providers, who all employ burstable billing. Using real-world workload traces, we show that our proposed method can reduce user's bandwidth cost by 26% and increase its total surplus by 23%, compared to the current practice of allocating bandwidth on-demand.

Motivation & Objective

• To address the lack of systematic methods for users to optimally manage bandwidth under burstable billing with uncertain demand.
• To develop a tractable mathematical model for calculating 95th percentile bandwidth usage under arbitrary demand distributions.
• To formulate a surplus-maximization problem that balances user utility and bandwidth cost under uncertainty.
• To extend the model to multi-provider scenarios, enabling workload distribution across providers using burstable billing.
• To evaluate the method using real-world Wikipedia workload traces and demonstrate significant cost and surplus improvements.

Proposed method

• Derives a closed-form expression for 95th percentile bandwidth usage to enable analytical optimization under arbitrary demand distributions.
• Formulates a stochastic optimization problem to maximize user surplus, defined as net utility minus bandwidth cost, under demand uncertainty.
• Applies nonlinear mixed-integer programming techniques to solve the non-convex optimization problem, evaluating both deterministic and stochastic prediction-based solution methods.
• Extends the framework to multi-provider environments, jointly optimizing bandwidth allocation and workload distribution across providers.
• Uses real-world Wikipedia page view traces for workload forecasting and performance evaluation under varying price and utility parameters.
• Employs normalization and comparative simulation across baseline, deterministic, and stochastic methods to validate performance gains.

Experimental results

Research questions

• RQ1 What is the optimal bandwidth allocation strategy for a user under burstable billing when future demand is uncertain?
• RQ2 How can a user maximize its surplus—defined as net utility minus cost—under burstable billing with arbitrary demand distributions?
• RQ3 How does the performance of the proposed optimization method compare to on-demand bandwidth allocation in real-world workloads?
• RQ4 What are the benefits of utilizing multiple providers under burstable billing in terms of cost reduction and surplus improvement?
• RQ5 How do bandwidth price and user utility sensitivity affect the performance of the proposed optimization framework?

Key findings

• The proposed method reduces user bandwidth costs by 26% compared to on-demand allocation using real-world Wikipedia workload traces.
• The method increases user surplus by 23% relative to current on-demand practices, especially under high bandwidth prices or high price sensitivity.
• The stochastic optimization approach outperforms both baseline and deterministic methods, particularly when demand uncertainty is high.
• Using multiple providers further reduces costs and increases surplus, with the multi-provider stochastic method outperforming single-provider strategies.
• Users with lower utility factors (more price-sensitive) benefit more from the optimization framework, showing greater surplus improvement.
• The framework remains effective even with simple forecasting methods, indicating robustness and practical applicability.