Demand forecast-based resource allocation is a key element of both application and network service management and resource cost optimization in cloud environments. A central mechanism called scaling or auto-scaling is responsible for adjusting the allocated resources dynamically to current or predicted demands. Most of today’s solutions support short-term optimization, where the long-term effect of scaling actions is not considered, therefore the overall operational cost can be sub-optimal. In this paper, we address this issue and provide a long and short-term scaling optimization method called LSSO. Our contribution is threefold. First, we introduce the formal description of the LSSO scaling problem motivated by real cloud native applications. Second, we transform the problem into a tractable graph representation and show that the optimal solution of the original problem emerges as a shortest path problem in the graph. As a result, the transformation and the optimal solution can be computed in polynomial time. Third, we evaluate and validate the proposed algorithm by numerical analysis on multiple datasets using a cost model which considers the running and scaling costs. Results suggest that if the dynamics of the traffic is fairly predictable and the scaling cost is not negligible, LSSO is able to reduce the operation costs significantly. The exact cost gain depends on the ratio of the running and scaling costs, but in realistic operation regimes it can reach even 18% and the method can also tolerate some inaccuracies in the forecast.
Fodor Balázs, Sonkoly Balázs, Toka László, Berkovics Fanni, Körtvélyi Nikolett, Mihályi Balázs, Koflanovics Kristóf, Szögi Péter
2023-04-24
Támogató: Ericsson