Utility-Function-Driven EnergyEfficient Cooling in Data Centers Authors: Rajarshi Das, Jeffrey Kephart, Jonathan Lenchner, Hendrik Hamamn IBM Thomas J. Watson Research Center Presented by: Shivashis Saha University of Nebraska-Lincoln Outline • • • • Introduction Related Work Data Center Energy Balance Utility Functions – Multiplicative utility functions – Additive utility functions • Experiments • Conclusion 2 Introduction • Data center energy management – “50% of existing data centers will have insufficient power and cooling within two years” – “Power is the second-highest operating cost in 70% of all data centers” – “Data centers are responsible for the tens of millions of metric tons of carbon dioxide emissions annually --- more than 5% of the total global emissions” 3 Introduction • Why use autonomic computing? – Large, difficult to manage, complex – Management problem is both qualitatively similar to and quantitatively harder than that of managing IT alone. 4 Contributions • Apply utility functions to save energy – Tradeoff between energy and temperature – Control parameters: • Fan speed • On/off states of individual Computer Room Air Conditioning (CRAC) • Proposed model show 12% reduction in energy without violating temperature contraints 5 Related Work • Saving more energy is not good if administrator does not want that! – Proposed model is flexible • Apply computational fluid dynamics modeling to complex data center environments • Temperature aware workload placement based on inlet temperature or heat recirculation 6 Data Center Energy Balance • PDC, power to run data center is split using switch gear equipment into: – Path to power the IT equipments – Path to power the supporting equipments 7 Data Center Energy Balance • The support path may include – Power for pumping coolant to and from CRACs to the chiller and to and from the chiller to the cooling tower • Power path for IT equipments include – Conversion loss due to the uninterruptible power supply (UPS) systems – Losses associated with the power distribution PPDU – The UPS systems are located outside the raised floor area 8 Data Center Energy Balance • The total power on the floor: • PIT is the power consumed by the IT equipments • Total CRAC fan power and CDU pump power: • The relation between fan power PCRACi and relative fan speed Θi 9 Data Center Energy Balance • Under steady state condition, the total raised floor power equal to the total cooling power – The reduced fan speed reduces the air flow: 10 Data Center Energy Balance • All raised floor power needs to be cooled by the chilling system, which required power for refrigeration – COP: the coefficient of performance of the chiller system (assume, average COP = 4.5) 11 Data Center Energy Balance • Reducing CRAC fan speeds, the fan power is reduced • This reduces both the raised floor power and the power needed from chiller system • However, reducing fan speed also increases the server inlet temperature A tradeoff between energy consumption and the temperature!!! 12 Utility Functions • Data center operators responsible for the physical environment tend not to be concerned about application level performance, e.g. performance, availability, or security • They are more concerned about cost, energy, temperature, and hardware lifetimes • There are two CRAC units, whose fan speeds are Θ1 and Θ2 13 Utility Functions • Multiplicative utility functions 14 Utility Functions The previous utility function is very harsh! 15 Utility Functions • Additive utility functions 16 Experiments 17 Experiments 18 Experiments 19 Experiments Each CRAC was: 1. Turned off 2. Turned on at lowest speed (60%) 3. Turned on at max speed (100%) 20 Experiments 21 Experiments • Snorkels were placed 22 Experiments 23 Conclusion • Use of utility functions in data centers • Total reduction of energy consumption by 14% • Dynamic aspects of utility functions are not yet considered • Investigation of techniques combining dynamic workload scheduling with dynamic workload migration 24 Thanks!