由於萬物聯網的蓬勃發展，伺服器因此被大量使用，也因為耗能的考量，其散熱問題受到很大的重視。伺服器散熱的指標端看其內部的PCIE匯流排擴充卡(以下簡稱擴充卡)之溫度控制，而擴充卡種類繁多，發熱表現不盡相同，經常受到伺服器內部發熱元件佈局影響，產生不均勻的溫度分布，甚至會出現熱耦合現象。有別於傳統將所有風扇統一控制在相同轉速，本研究藉由深度學習預測模型、貪婪探索演算法以及費用方程式得到預先控制器，取得風扇分組之最佳組合，分別操作不同風扇組於不同轉速，可將擴充卡核心的溫度收斂至設定點之下，且能降低部份風扇轉速，有效地減少伺服器散熱的功耗。本研究開發自動化資料擷取系統，能大量自動蒐集資料；提出實時控制系統，可以即時控制PCIE擴充卡溫度；並設計友善的圖形化介面，以利系統的操作。實驗結果顯示，本研究所提的分組風扇的預先控制器與統一風扇的PID控制器皆可使目標擴充卡核心溫度達到設定點，但比起PID控制器，預先控制器可節省20%至30%左右的風扇轉速。

Due to rapidly growing of AI Internet of Things (AIoT), a huge amount of servers have been used to process the signals of AIoT recently. As a result, the ability of removing heat from servers becomes very crucial. In a server, the most critical heat performance index is the temperature control of its PCIE riser cards. However, there are various types of PCIE riser cards, and their heat generation will be different according to their specifications. The temperatures of PCIE riser cards tend to be unevenly distributed and sometimes heat-coupling may occur. Different from using the same loading for all cooling fans, this thesis proposes a preactive controller, combining deep learning, greedy algorithm, and cost function, which has the capability to obtain optimal grouped fans combination that can achieve the temperature control for PCIE cards. The results of experiments show that the server can run with two less loaded fans, compared to the PID control, resulting in effective reduction of power consumption. Also, an auto-data-acquiring system and a real-time control system, alone with their GUI, are developed by integrating server, PCIE riser cards, fans, and temperature acquiring device. These two systems are made for collecting massive data and for real-time controlling the PCIE temperatures. The results demonstrate that the proposed preactive grouped fan-control can control the temperature of target riser card around the set point, while the fanspeeds are reduced by around 20% to 30%, compared to the conventional unified PID fan-control.

[1] A. Holst, "Public cloud services end-user spending worldwide from 2009 to 2022(in billion U.S. dollars)," 2020.
[2] R. Hintemann and S. Hinterholzer, "Energy Consumption of Data Centers Worldwide - How will the Internet become Green?," 2019.
[3] D. Li, R. Ge, and K. Cameron, "System-Level, Unified In-band and Out-of-band Dynamic Thermal Control," in 2010 39th International Conference on Parallel Processing, 12-16 Sept. 2010, pp. 131-140,
[4] B. Acun, E. K. Lee, Y. Park, and L. V. Kale, "Support for Power Efficient Proactive Cooling Mechanisms," in 2017 IEEE 24th International Conference on High Performance Computing (HiPC), 18-21 Dec. 2017, pp. 94-103,
[5] K. Zhang, A. Guliani, S. Ogrenci-Memik, G. Memik, K. Yoshii, R. Sankaran, and P. Beckman, "Machine Learning-Based Temperature Prediction for Runtime Thermal Management Across System Components," IEEE Transactions on Parallel and Distributed Systems, vol. 29, no. 2, pp. 405-419, 2018.
[6] R. Lucchese, J. Olsson, A.-L. Ljung, W. Garcia-Gabin, and D. Varagnolo, "Energy savings in data centers: A framework for modelling and control of servers’ cooling," IFAC-PapersOnLine, vol. 50, no. 1, pp. 9050-9057, 2017.
[7] C. Lee and R. Chen, "Optimal Self-Tuning PID Controller Based on Low Power Consumption for a Server Fan Cooling System," Sensors, vol. 15, no. 5, pp. 11685-11700, 2015.
[8] 何宗翰, "以類神經網路設計與實現伺服器散熱系統溫度控制器," 碩士, 動力機械工程學系, 國立清華大學, 新竹市, 2014.
[9] 林炘泓, "GRU實現伺服器PCIE溫度估測器及散熱控制系統," 碩士, 動力機械工程學系, 國立清華大學, 新竹市, 2020.
[10] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, "Empirical evaluation of gated recurrent neural networks on sequence modeling," arXiv preprint arXiv:1412.3555, 2014.
[11] Y. Qin, D. Song, H. Chen, W. Cheng, G. Jiang, and G. Cottrell, "A dual-stage attention-based recurrent neural network for time series prediction," arXiv preprint arXiv:1704.02971, 2017.
[12] K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y. Bengio, "Learning phrase representations using RNN encoder-decoder for statistical machine translation," arXiv preprint arXiv:1406.1078, 2014.
[13] R. Hübner, M. Steinhauser, and C. Lehle, "A dual-stage two-phase model of selective attention," (in eng), Psychol Rev, vol. 117, no. 3, pp. 759-84, 2010.
[14] S. Hochreiter and J. Schmidhuber, "Long Short-Term Memory," Neural Computation, vol. 9, no. 8, pp. 1735-1780, 1997.
[15] S. Du, T. Li, Y. Yang, and S.-J. Horng, "Multivariate time series forecasting via attention-based encoder–decoder framework," Neurocomputing, vol. 388, pp. 269-279, 2020.
[16] M. Schuster and K. K. Paliwal, "Bidirectional recurrent neural networks," IEEE Transactions on Signal Processing, vol. 45, no. 11, pp. 2673-2681, 1997.
[17] J. Hu and W. Zheng, "Multistage attention network for multivariate time series prediction," Neurocomputing, vol. 383, pp. 122-137, 2020.
[18] Y. Liang, S. Ke, J. Zhang, X. Yi, and Y. Zheng, "Geoman: Multi-level attention networks for geo-sensory time series prediction," in IJCAI, vol. 2018, pp. 3428-3434,
[19] J. Hu and W. Zheng, "Transformation-gated LSTM: efficient capture of short-term mutation dependencies for multivariate time series prediction tasks," in 2019 International Joint Conference on Neural Networks (IJCNN), 14-19 July 2019, pp. 1-8,
[20] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, "Attention is all you need," in Advances in neural information processing systems, pp. 5998-6008,