Please use this identifier to cite or link to this item:
http://hdl.handle.net/11375/25465
Title: | THERMAL-AWARE COOLING CONTROL AND WORKLOAD ASSIGNMENT FOR HETEROGENEOUS DATA CENTERS |
Authors: | Mirhoseininejad, Seyedmorteza |
Advisor: | Down, Douglas |
Department: | Computing and Software |
Publication Date: | 2020 |
Abstract: | Data centers are struggling with inefficient power usage, which is one of their crucial challenges. Information technology (IT) and cooling infrastructure are the major contributors to power consumption in data centers. Server over-cooling, inefficient power management of cooling units, and thermal-oblivious assignment of server workload are significant contributors to the considerable power wastage in data centers. These issues can be addressed by recognizing cooling units' ability to dissipate heat from different locations inside a data center (cooling heterogeneity) and thermal properties of individual servers (server heterogeneity). This problem has not been studied thoroughly in the literature. This dissertation consists of five phases aiming to exploit the correlation between IT and cooling units in data centers for the efficient use of power. The study begins with exploiting thermal differences between servers and ends with implementing a complete holistic thermal-aware control system. The first phase identifies server differences due to their cooling requirements and power consumption. Hence, the problem of distributing workload to minimize power consumption while respecting the thermal differences between servers (server heterogeneity) is considered. The resulting optimization problem is addressed using an effective heuristic. This heuristic distributes workload among servers in a way that minimizes their cooling requirements and sets the cooling set-point accordingly. The second phase investigates data center cooling heterogeneity, exploring the thermal differences among servers and between server locations that need to be cooled by cooling units. A physics-based thermal model is used to calculate the inlet temperatures of servers based on cooling and IT settings. It is shown that both the assignment of workload and the adjustment of cooling parameters affect the cooling cost, revealing a possible trade-off that can be optimized. Potential power-savings obtained by optimal assignment of workload and choices of cooling unit operational variables are explored. Due to their complexity (both operationally and in their development), exploring synergies between IT and cooling units using physics-based thermal models is challenging. So, during the third phase of this work, an adaptive data-driven thermal model using time series prediction methods is developed. This thermal model, to a great extent, solves the problem of temperature predictions in data centers. This learning-based thermal model is fast, adapts to thermal changes in a data center, and does not require prior knowledge of heat transfer rules. \textit{Holistic thermal-aware workload management and infrastructure control for heterogeneous data centers using machine learning} is the subject of the next phase, which considers the problem of workload assignment and cooling control, combining the main aspects of thermal heterogeneity in data centers. It assumes thermal differences between servers and between locations in a data center using the thermal model constructed in the previous phase. The results show a potential to save a considerable amount of power as a result of leveraging synergies between the workload scheduler and control of the cooling unit. Finally, a real-time control system is implemented which jointly controls cooling units and workload assignment in a data center. The controller in this system considers the thermal differences of servers to generate the expected thermal requirements corresponding to servers using a temperature requirement map. The capability of the cooling unit and also the thermal effects of servers are accounted for in this map. The system determines the operational variables of the cooling units using model predictive control to minimize the cooling power while satisfying the required temperatures given by the temperature map. |
URI: | http://hdl.handle.net/11375/25465 |
Appears in Collections: | Open Access Dissertations and Theses |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
mirhoseininejad_seyedmoereza_2020May_phd.pdf | 4.98 MB | Adobe PDF | View/Open |
Items in MacSphere are protected by copyright, with all rights reserved, unless otherwise indicated.