Optimization in Data Centres Jayantha Siriwardana and Saman K. Halgamuge Department of Mechanical Engineering Melbourne School of Engineering
Outline Background and Motivation Overview and scale of data centre energy consumption Problems of data centres in long term Possible strategies to tackle problems Conclusions 1
Outline Background and Motivation Overview and scale of data centre energy consumption Current data centre designs Problems of data centres in long term Possible strategies to tackle problems Conclusions 2
Background and Motivation (1/2) What are data centres Server farms, IT centres, Computer rooms, etc. Why they are important Centralized management Powerful computation capabilities Backbone of internet infrastructure Why optimisation of thermal performance is important Improve reliability Reduce system down time Save energy cost (more than 40% are cooling cost) 3
Background and Motivation (2/2) During the 15 year or more lifetime of the data centre, it receives series of equipment upgrades. Smaller server size and greater processing speeds mean more power per m 2 in the data centre. The cooling system needs to be optimised. CFD Modelling of data centre to ensure efficient cooling 4
Outline Background and Motivation Overview and scale of data centre energy consumption Problems of data centres in long term Possible strategies to tackle problems Conclusions 5
Overview of data centre energy consumption Energy consumed by data centres grow 12% annually. Accounted for 1.5% of the US total electricity consumption (2006) Only half of energy is used for powering servers Power conversion & distribution 7% Servers 51% Air conditioning 42% (Source: US Department of Energy Data Center Energy Efficiency Program Publications) Typical data centre energy usage 6
Power trends in data centre equipment Graph based on data centre equipment manufacturers forecast of energy consumption of their own products. (Copyright : The Uptime Institute) 7
Data centre efficiency improvement targets Data centre Infrastructure Efficiency DCiE < 0. 5 Both power conversion and cooling systems can be optimized to save energy 10% energy saving target in all US data centre means: 10.7 billion kwh Energy for IT Equipment = Total Energy for Data Center Equivalent to electricity consumed by 1 million typical households a year. Reducing green house emissions by 6.5 million metric tons per year 8
Current data centre design standards (1/2) Features Raised tile floor Hot aisle, cold aisle server rack placement Cold air floor rate optimised to reduce hot air recirculation 9
Current data centre design standards (2/2) Other data centre architectures Raised floor supply / Ceiling return Raised floor supply / ceiling supply Non raised floor / ceiling supply 10
Outline Background and Motivation Overview and scale of data centre energy consumption Problems of data centres in long term Possible strategies to tackle problems Conclusions 11
Problems of data centres in long term (1/3) Heat Recirculation Hot exhaust air recirculated to cold aisles Recirculating air flow is extremely complex Causes unexpected server rack inlet temperature rises Occurs due to Uneven distribution of workload Uneven thermal dissipation of data centre equipment Non optimised air flow rates of CRACs 12
Problems of data centres in long term (2/3) Cold air short circuiting Cold air flows directly back to CRAC Causes unexpected temperature rise at server inlets Occurs due to Poor air flow within the data centre Non optimised air flow rates of CRACs 13
Problems of data centres in long term (3/3) Hot spots Unusually hot locations in data centres Occurs due to Poor air flow within the data centre Uneven distribution of workload Forces data centre managers to overcool the facility 14
Outline Background and Motivation Overview and scale of data centre energy consumption Problems of data centres in long term Possible strategies to tackle problems Conclusions 15
Strategies to tackle problems Use data centre equipment as long as possible. Competition, new services and technological advancements limit the equipment lifetime. Optimisation of thermal performance Maintain maximum possible server inlet temperature needed for cooling Schedule tasks for servers to optimize thermal dissipation (avoid hot spots) Schedule tasks for servers to minimize heat recirculation Thermal Aware Task Scheduling 16
Optimisation of thermal performance (1/2) Thermal aware task scheduling How to divide a total task C among N servers to finish it with minimal energy cost? Workload of each server is calculated to minimize its thermal interference to other servers. Thereby, minimize the hot air recirculation making it possible to increase cooler air temperature. Task {100} Data centre task scheduler 20 15 20 25 30 40 15 5 17
Optimisation of thermal performance (2/2) Thermal aware task scheduling T T + in = sup DP Server power consumption Server inlet air temperature Supplied cold air temperature Heat recirculation coefficient matrix Proportional to server work load Outlet air flow i T out Server node Can be calculated using CFD T + d P sup i i n Inlet airflow (mixture of supplied cold air and recirculated hot air) 18
Our approach to the problem Modelling thermal energy distribution in data centres (passive versus active strategies) Thermal aware task scheduling for heterogeneous data centres. Optimise multiple online and offline tasks Optimise heterogeneous data centre configurations Optimisation of data centre upgrading for minimum possible energy usage Thermal interference minimisation among server racks Looking into alternative cooling technologies (Liquid cooling) 19
Conclusions Data centres attract growing concerns on their large power consumption. Reasonable efforts are made for data centres to be energy efficient at initial design stage. However, existing data centre designs can fail to perform energy efficiently over time Thermal performance is a key concern in upgrading Thermal aware task scheduling could be used to keep cooling costs low in long term 20