Page 85 - Layout 1

Department of Engineering, ICT and Technologies for Energy and Transport Patent Title Method for the real-time control of working frequencies of large-scale infrastructures with heterogeneous hardware requirements. Ref. CNR 10336 Assignee(s): CNR – University of Cagliari CNR Institute: ISSIA Main Inventor: Luca Caviglione Countries: EP Priority date: 26/06/2014 Abstract Nowadays, Online Social Networks are more and more complex and serve million of user simultaneously; for this reason, they require Internet-scale infrastructures, characterized in a homogeneous aggregates of machineries. So, the energetic requirement for these systems is often a challenge, both in terms of power efficiency and in terms of different types of consumption (e.g. dissipation of produced heat). The object of the present invention permits to: i) define a performance index for the service provided; ii) compute “virtual” networking frequencies (that is “technology independent”) at high level ; iii) allow the real-time dynamic variation of working frequencies in order to reduce energy consumption. Background To optimize the consumptions of aggregates of machines (e.g., a group of devices or network appliances), we take also advantage of results obtained in the literature dealing with Dynamic Voltage-Frequency Scaling (DVFS). However, DVFS does not consider global performance indexes and it could lack of scalability. Technology The system consists of: i) a non-linear controller in charge of computing high-level frequencies in an efficient manner for homogeneous aggregates of machineries taking into account a performance parameter; ii) a set of local mappers (one per worker) to compute per-host frequency assignments in real-time and in a fast and computationally feasible way (i.e., via random optimization methods). Advantages and Applications Thanks to a 2-level architecture (controller and mappers), the system is “technology-independent”. The framework is then used to adjust the frequencies of every devices/machinery in order to not consume unneeded power. A typical application scenario is the control of physical machines of a cloud infrastructure or nodes of Internet-scale systems. Development stage The method has been implemented via tools for rapid software prototyping and it is proved to effectively work in many simulated environments. 75