應用深度強化學習於氣冷式伺服器散熱系統之節能研究

Investigation of energy-saving for air-cooled server system via deep reinforcement learning

連允玄(Yun-Hsuan Lien)

指導教授 ： 王啟川

國立交通大學/工學院/機械工程系所/碩士(2019年)

摘要

氣冷卻是目前應用最廣的伺服器散熱方案，由於伺服器內部存在熱源彼此間會交互影響散熱狀態的問題，並且不同的伺服器配置的風扇系統性能亦並非相同。故本研究嘗試以深度強化學習演算法，透過蒐集伺服器內部的散熱特徵的大數據資料，來使演算法辨識進而達成統御的控制方法。基於這樣的研究方法，本研究以過往的散熱器性能經驗式研究為基礎發展伺服器的暫態熱傳模型。並以過往的相關研究資料修正其對密集型平板鰭片性能預測失準的問題，同時也擴大了經驗式的適用範圍。所得到的修正係數可對141筆實驗資料的紐森數平均預測誤差為21.5 %，而對159筆實驗資料的無因次化壓降平均預測誤差為20.4 %。最後以實驗的方式驗證環境模型計算的正確性，以本研究列舉計算案例而言，有效熱傳溫差的誤差最大為17.4 %、最小為0.2 %。而基於這樣的環境模型，本研究以統計的方式描述伺服器內的熱源狀態，並提出描述風扇性能的特徵參數作為演算法辨識的觀測特徵。最終的訓練結果，深度確定性策略梯度演算法(Deep Deterministic Policy Gradient，DDPG)在本研究的伺服器配置測試範圍內，可控制伺服器內最大過熱度熱源在設定的溫度上限附近進行散熱。基於這樣的散熱控制可利用較高的有效熱傳溫差進行散熱，減少依靠風扇來提升熱傳效果，使散熱系統的能耗得以降低。本研究歸內散熱控制上的行為與性能，總結最大過熱度熱源是節能散熱控制上的限制，後進的研究者可以此想法為核心改善演算法的認知特徵品質，以獲得更好的節能成效。

關鍵字

資料中心 ； 氣冷式散熱 ； 深度強化學習

並列摘要

Air-cooling is currently the most widely used server cooling scheme, there is an issue that the heat sources in the server that will influence each cooling condition, and the performance of the fan system in the different server is not the same, either. Therefore, this study attempts to use the deep reinforcement learning algorithm to let the agent can identify the cooling features of the server system from the collecting big data, and achieve the governing control method about this kind of system. Based on this research method, the transient heat transfer model of the server is developed based on the heat sink performance correlation of the previous research. And the previous research experiment data were used to correct the misprediction of the performance of the compact plate fin, and the application range of the correlation is also extended. The average predicted error of Nusselt number from 141 experimental data is 21.5%, and the average predicted error of dimensionless pressure drop from 159 experimental data is 20.4%, Finally, use experimental data to verify the correctness of the transient server model. For example in this study, the maximum error of effective heat transfer temperature difference is 17.4%, and the minimum was 0.2%. Based on such an environment model, this study describes the situation of the heat source in the server in a statistical way and proposes the characteristic parameters to describe the fan performance as the observed features identified by the algorithm. In the final training result of the algorithm, within the range of the server configuration in this study, the DDPG method can control the max overheated heat sources in the server near the temperature limitation set during cooling. Based on this cooling control, the cooling system can use a higher effective heat transfer temperature difference during cooling, and reduce the reliance from the increasing heat transfer effect cause of the fan, and then can reduce the energy consumption of the cooling system. This study generalizes the behavior and performance of cooling control, concludes that the max overheated heat source is the limitation of energy-saving cooling control. Later researchers can take this point as the main idea to improve the quality of the algorithm’s observed features, so as to obtain better energy-saving results.

並列關鍵字

data center ； air cooling ； deep reinforcement learning

參考文獻

[1] Stats, I. W., 2019, "Internet World Stats : Usage and Population STatistics," https://www.internetworldstats.com/stats.htm.

[2] CoinMarketCap, 2019, "Bitcoin," https://coinmarketcap.com/currencies/bitcoin/.

[3] Herrlin, M. K. J. T.-A. S. o. H. R., and Engineers, A. c., 2005, "Rack cooling effectiveness in data centers and telecom central offices: The rack cooling index (RCI)," 111(2), p. 725.

[4] Deodhar, A., Bhagwat, H., Singh, U., Sankaranarayanan, D., Singh, A. K., and Sivasubramania, A. J. A. T., 2015, "Coordinated real-time management of return-air-temperature-controlled cooling units in data centers," 121(1), pp. 440-454.

[5] Gao, J., 2014, "Machine learning applications for data center optimization."