Using Multiple Seasonal Holt-Winters Exponential Smoothing to Predict Cloud Resource Provisioning

Elasticity is one of the key features of cloud computing that attracts many SaaS providers to minimize their services’ cost. Cost is minimized by automatically provision and release computational resources depend on actual computational needs. However, delay of starting up new virtual resources can cause Service Level Agreement violation. Consequently, predicting cloud resources provisioning gains a lot of attention to scale computational resources in advance. However, most of current approaches do not consider multi-seasonality in cloud workloads. This paper proposes cloud resource provisioning prediction algorithm based on Holt-Winters exponential smoothing method. The proposed algorithm extends Holt-Winters exponential smoothing method to model cloud workload with multi-seasonal cycles. Prediction accuracy of the proposed algorithm has been improved by employing Artificial Bee Colony algorithm to optimize its parameters. Performance of the proposed algorithm has been evaluated and compared with double and triple exponential smoothing methods. Our results have shown that the proposed algorithm outperforms other methods.

💡 Research Summary

The paper addresses a critical challenge in cloud computing: predicting resource demand in advance to avoid Service Level Agreement (SLA) violations caused by the latency of provisioning new virtual machines or containers. While many existing forecasting approaches—such as ARIMA, single‑season Holt‑Winters, or machine‑learning regressors—have been applied to cloud workloads, they typically assume a single dominant seasonal pattern. Real‑world cloud traffic, however, exhibits multiple overlapping cycles (e.g., daily, weekly, monthly), which limits the accuracy of those methods.

To overcome this limitation, the authors extend the classic Holt‑Winters exponential smoothing model to accommodate multiple seasonal components. In the traditional formulation, the forecast at time (t+1) is a function of a level component (L_t), a trend component (T_t), and a single seasonal factor (S_t). The proposed Multi‑Seasonal Holt‑Winters (MSHW) introduces (K) seasonal indices (S^{(k)}_t) (for (k = 1,\dots,K)), each associated with its own period (p_k) (e.g., 24 h, 168 h, 720 h) and smoothing coefficient (\gamma_k). The forecast equation becomes:

\