Identifying the most critical components and maximizing their availability subject to limited cost in cooling subsystems

Abstract

Cooling plays an important role on data center (DC) availability, mitigating the Technology of Information (IT) components’ overheating. Although several works evaluate the performance of cooling subsystem in a DC, a few studies consider the significant relationship between cooling and IT subsystems. Moreover, a DC provider has limited tools in order to choose its IT and cooling components to obtain a desired availability subject to limited cost. This work provides scalable models (using Stochastic Petri Nets - SPN) to represent a cooling subsystem and to analyze its failures’ impact concerning financial costs and service downtime. This study also identifies the components that most impact on DC availability, as well as proposes a strategy to maximize the DC availability with a limited budget. Notwithstanding, the optimization process to maximize availability becomes very costly when used the proposed DC SPN models due to time-to-solve, which leads to the application of cheaper models, however, efficient, called surrogate models. In order to apply the most accurate surrogate model for optimization tasks, this work compares three surrogate models strategies. In the optimization, based on solutions obtained in the chosen surrogate model, there is a three-algorithm comparison to choose one with best results. Results show that a more redundant cooling architecture reduces costs in 70%. Cooling components’ analysis identified the chiller as the most impactful component concerning availability. Regarding surrogate models based on DC model, Gaussian Process (GP) obtained more confident results. Finally, Differential Evolution (DE) had the best results on availability’s maximization in a DC.