Comparative Study of Energy Performance between Chip and Inlet Temperature-Aware Workload Allocation in Air-Cooled Data Center

Abstract

Improving the energy efficiency of data center has become a research focus in recent years. Previous works commonly adopted the inlet temperature constraint to optimize the thermal environment in the data center. However, the inlet temperature-aware method cannot prevent the servers from over-cooling. To cope with this issue, we propose a thermal-aware workload allocation strategy with respect to the chip temperature constraint. In this paper, we conducted a comparative evaluation of the performance between the chip and inlet temperature-aware workload allocation strategies. The workload allocation strategies adopt a POD-based heat recirculation model to characterize the thermal environment in data center. The contribution of the temperature-dependent leakage power to server power consumption is also considered. We adopted a sample data center under constant-flow and variable-flow cooling air supply to evaluate the performance of these two different workload allocation strategies. The comparison results show that the chip temperature-aware workload allocation strategy prevents the servers from over-cooling and significantly improves the energy efficiency of data center, especially for the case of variable-flow cooling air supply.

Data center power consumption prediction based on principal component analysis and DeepAR

“…The literature [9] obtains the total power consumption of the server by modeling the power consumption of CPU, memory, hard disk, fan and other components separately and then accumulating them, each of the above different components has its own power consumption model. The literature [10] used the product of server utilization and chip temperature to fit the server power consumption. The literature [11] proposes a server power prediction model based on performance counters.…”

Energy, exergy and computing efficiency based data center workload and cooling management

“…𝑁 𝑟 𝑖=1 𝑁 𝑧 𝑗=1 , (11) where the number of racks 𝑁 𝑟 = 5, number of servers per zone 𝑁 𝑠 = 3, and the number of zones in a rack 𝑁 𝑧 = 5. The server power model constants 𝐴 1 = 223.4 and 𝐴 2 = 154.5 are obtained from the datasheet of an HP ProLiant DL360 G5 server with two Intel Xeon E5-2697 v3 processors [43].…”

“…It is worth noting that 𝐶𝑃𝑅 𝑚 only captures the effects of computing performance degradation at elevated chip temperature (see Section 2.4). This approach is somewhat similar to CPU temperature-aware workload scheduling and cooling control, where the idea is to keep the CPU die temperature below a certain value [11]. The objective function 𝐶𝑃𝑅 𝑚 is represented as a function of the mean CPU temperatures of each zone 𝑇 𝑖,𝑗 𝐶𝑃𝑈 (see Eqs.…”

“…The thermal reliability of computing chips is degraded because of (1) transient temperature fluctuations that diminish their lifetimes and (2) steady-state operating temperatures that rise above the critical allowable chip temperature and induce thermal shutdown. These thermal issues can be addressed by utilizing chip temperature-aware and cold aisle temperature-aware workload management approaches [11]. Other workload scheduling approaches address thermal guidelines and operational expenditure by minimizing heat recirculation across the cold and hot air streams in a DC [12], diminishing hot spots among neighboring servers [13], and lowering temperature nonuniformities in server exhausts [14,15] (see Table 1).…”