Energy Efficiency Optimization Of Datacenters With Liquid Cooling Systems

With the rapid development of cloud computing,the energy consumption of datacenters is rising sharply.In addition to the IT equipment,a large proportion of the total energy consumption of datacenters can be attributed to the cooling system,which impels the revolution of datacenters to improve the energy efficiency by reducing energy consumption of cooling.Compared with the traditional air cooling,liquid cooling improves cooling efficiency,in line with the trend of increasingly computation-intensive datacenters.The conservative cold water cooling strategy designed for the worst-case scenario causes large energy waste,and is now gradually replaced by warm water cooling.Warm water cooling can not only reduce cooling energy by less use of the chiller,but also bring the opportunity to harvest and recycle the waste heat by raising its quality.Currently,however,most of the waste heat produced by datacenters is ejected into the environment without being recycled.Besides,district heating,the most commonly used method in industry has its limitations,while by being turned into electricity,waste heat can play a wider range of uses in datacenters.Heat to Power(H2P),a practical,safe,economical and energy-recycling system,is proposed for warm water-cooled datacenters.In H2 P,thermoelectric generators are integrated into the existing liquid cooling system of datacenters and generate electricity by the temperature difference between warm coolant from servers and cold natural water.To verify the feasibility of the design,a real hardware prototype is built and a series of measurements are conducted on it.Based on these measurement results,some software optimization methods are proposed.Specifically,a water circulation optimization design considers how many servers should be arranged in a water circulation.To maximize the power generation and meanwhile guarantee the CPU’s safety,a CPU-utilization-adaptive cooling control is developed,in which the optimal cooling setting(i.e.,the coolant temperature and the flow rate)is dynamically adjusted according to the maximum CPU utilization when a CPU safe operating temperature is set as a constraint.In the workload balancing strategy,the cooling setting is determined by the average CPU utilization,so that the power generation can be optimized.The evaluation based on the prototype and cluster traces from Google and Alibaba verifies that the proposed H2 P system effectively improves the energy efficiency and saves the total cost of liquid-cooled datacenters.The results show that about 4.177 W of electricity can be generated on one CPU with 12 thermoelectric generators.H2 P can achieve a power reusing efficiency of 12.56% and 14.23% after optimization.It also can reduce the total cost of ownership by 0.49%～0.57%.