Heuristic-based Optimal Resource Provisioning in Application-centric Cloud

Cloud Service Providers (CSPs) adapt different pricing models for their offered services. Some of the models are suitable for short term requirement while others may be suitable for the Cloud Service User’s (CSU) long term requirement. In this paper, we look at the problem of finding the amount of resources to be reserved to satisfy the CSU’s long term demands with the aim of minimizing the total cost. Finding the optimal resource requirement to satisfy the the CSU’s demand for resources needs sufficient research effort. Various algorithms were discussed in the last couple of years for finding the optimal resource requirement but most of them are based on IPP which is NP in nature. In this paper, we derive some heuristic-based polynomial time algorithms to find some near optimal solution to the problem. We show that the cost for CSU using our approach is comparable to the solution obtained using optimal Integer Programming Problem(IPP).

💡 Research Summary

The paper addresses the problem of minimizing the total cost incurred by a Cloud Service User (CSU) when running a long‑term application on a public cloud. Modern cloud providers, exemplified by Amazon EC2, offer several pricing models: on‑demand (pay‑as‑you‑go), reserved instances (a one‑time reservation fee plus a discounted hourly usage charge for a fixed contract period), and spot instances (dynamic pricing). The authors focus on the first two models and assume that the demand for compute resources is known in advance for each discrete time stage of the application’s execution.

The authors first formulate the resource provisioning problem as an Integer Programming Problem (IPP). Let T be the number of stages, K the number of reservation contract types, and Dₜ the demand (in instance‑hours) at stage t. For each contract k, the parameters are: one‑time reservation cost Rₖ, hourly usage cost rₖ, and contract duration tₖ (in stages). Decision variables include the number of instances newly reserved at stage t (xᴿₖₜ), the number of reserved instances actually launched at stage t (xʳₖₜ), and the number of on‑demand instances launched at stage t (xₒₜ). The stage cost is Cₜ = Σₖ (xᴿₖₜ·Rₖ + xʳₖₜ·rₖ·h) + xₒₜ·o·h, where o is the on‑demand hourly rate and h is the length of a stage in hours. The objective is to minimize Σₜ Cₜ subject to non‑negativity, reservation‑availability, and demand‑satisfaction constraints. Because the IPP is NP‑hard, solving it exactly is computationally infeasible for realistic workloads.

To obtain a practical solution, the paper proposes polynomial‑time heuristics. The core insight is derived for the special case of a single contract type (K = 1). The demand vector D is sorted in non‑decreasing order to obtain Dˢ. The authors prove two lemmas: (1) the reservation fee must be smaller than the total discount obtained over the contract (Rₖ < tₖ·αₖ, where αₖ = o − rₖ), and (2) the optimal number of reserved instances must be equal to one of the values in Dˢ. By analyzing the cost difference ΔE when moving from one candidate reservation level to the next, they derive a simple formula for the optimal index j* = ⌊tₖ − Rₖ/αₖ⌋. Reserving Dˢⱼ* instances yields the minimum total cost for the single‑contract scenario. This result can be found in O(T log T) time (sorting plus a constant‑time scan).

For the general case with multiple contract types, the authors extend the approach in two stages. First, they apply the single‑contract heuristic independently to each contract to generate a small set of candidate reservation quantities. Second, they evaluate all feasible combinations of these candidates (the number of combinations is bounded by K·T rather than exponential) and select the combination with the lowest total cost. The overall complexity remains polynomial, specifically O(K·T·log T).

Experimental evaluation uses real Amazon EC2 pricing (on‑demand, 1‑year and 3‑year reserved instances) and synthetic workloads derived from typical cloud usage patterns over a 12‑month horizon. The optimal IPP solution is obtained with a commercial solver for benchmarking. The heuristic’s total cost is within 1–2 % of the IPP optimum, while the runtime drops from minutes/hours to a few seconds. The results also show that longer contract durations (higher discount) lead to larger optimal reservation quantities, confirming the theoretical analysis.

The paper’s contributions are: (1) a formal IPP model for mixed reservation/on‑demand provisioning, (2) a rigorous proof that the optimal reservation level for a single contract lies in the sorted demand set, (3) a practical, polynomial‑time heuristic for multiple contracts, and (4) empirical validation with real cloud pricing data. Limitations include the assumption of perfectly known future demand and static pricing; the authors suggest future work on stochastic demand forecasts and dynamic spot‑price integration.

In conclusion, the proposed heuristic provides a near‑optimal, computationally efficient method for CSUs to decide how many instances to reserve in advance, achieving substantial cost savings without the prohibitive computational burden of exact integer programming. This work offers a valuable tool for cloud cost management and a foundation for further research on uncertainty‑aware provisioning.