基於變異感知多核心系統與高可靠度機器學習引擎之階層式物聯網系統設計

Variation-Aware Many-core System and Robust Machine Learning Engine for Layered IoT Environment

李懷霆(Huai-Ting Li)

指導教授 ： 吳安宇

國立臺灣大學/電機資訊學院/電子工程學研究所/博士(2018年)

https://doi.org/10.6342/NTU201801780

摘要

隨著電子技術的高速發展，物聯網(Internet of Things)帶動了許多革命性的進步與應用。其中，自駕車與醫療救護等應用對於資料傳輸過程中所造成的延遲較沒有忍受能力，需要即時地處理、分析與儲存。為了提升後端資料分析運算的效率及降低資料傳輸的延遲、頻寬有限之問題，霧運算與邊緣運算等階層性的概念被視為可行性高的物聯網設計架構。此架構下，在雲端，通常為以先進製程實現的多核心系統，提供處理海量數據之計算能力。在使用者端，需利用各種低功耗感測器，有效率地擷取各種所需要的資料。而邊緣運算是在使用者端和雲端交界處做運算，以提供即時性服務，然而，邊緣運算設備通常是個人所有的行動裝置，其記憶體容量與硬體計算資源皆為有限。
為提升此架構之效率，可行的霧運算分工如下：雲端藉由事先得到的大量資料，利用高複雜度機器學習演算法建立分析模型；感測端採用近年新興的壓縮感知(Compressive Sensing)技術的各種低功耗感測器，有效率地獲取資料；邊緣運算在進行即時初步分析時，必須直接執行在低維度的壓縮感知測量域達到壓縮學習(Compressed Learning)的目標，以避免執行高複雜度的壓縮感知還原演算法；並只將重要的資料傳往雲端進行詳細的分析，以降低網路流量以及雲端的運算負荷。
然而，可靠度問題在低功耗設計或是先進製程的技術之下皆會更加嚴重，成為設計一個物聯網系統必須仔細考量的議題。隨著製程不斷微縮，多核心系統在單位面積內可放入更多的元件以提供更高的運算能力，但是系統軟錯誤(Soft Error)發生率也隨之提高。除此之外，製程參數變異隨著製程微縮越來越嚴重，且軟錯誤發生率與製程參數呈現指數相關性，進而造成了各個核心之軟錯誤發生率有所差異，如何在此系統特性下開發一個變異感知之多核心運算系統將會是雲端設計的一大挑戰。另一方面，在壓縮學習中，壓縮感知感測器在感測資料時伴隨著硬體不理想性，感測資料將受到熱雜訊參雜，或是干擾產生的訊號飄移，造成訊號品質下降，進而降低分析準確度。因此，如何同時保持壓縮學習演算法之低運算複雜度並解決發生在感測器中的可靠度問題，將會是霧端設計最重要的議題。
為了克服上述挑戰，在雲端之多核心系統設計上，本論文利用多核心系統中本身就具有的核心冗餘特性，讓各核心可動態地組成具有N級冗餘(N-Modular Redundancy)運算以及錯誤回復的機制，並設計一適當的任務配置機制以實現低能源成本的高可靠度運算系統。本設計將考慮多種變異性，包含執行緒之間的執行時間與重要性差異，以及核心之間因製程變異所造成的錯誤率差異性，再依照這些資訊給予每個執行緒最適當的冗餘核心數量和配置，以利減少冗餘核心使用量與增加系統可靠性。另外，在霧端之壓縮學習設計上，本論文開發一種整體式(Ensemble)極限學習機(Extreme Learning Machine)，在有限的運算資源下，達到初步分析所需要的準確度。並利用主成份分析在事先以部分還原的訓練資料中找出字典，壓縮訊號便可不經過還原步驟直接轉換到特徵空間，轉換的過程中盡量保留訊號相關資訊、移除雜訊成分以達到高可靠度的目標。除此之外，為了將此壓縮學習引擎實現在邊緣運算裝置上，本論文提出適合硬體實作的架構設計，採用硬體資源共用的方式，並在台積電90奈米製程環境下布局與繞線，藉此驗證此引擎能夠達到較高的能源與晶片面積使用效率。
本論文針對於階層式物聯網架構提出多項前瞻性的高可靠度設計方法，並希望能夠成為未來智慧型物聯網的關鍵技術。

關鍵字

可靠度設計 ； 多核心系統 ； 機器學習

並列摘要

With the advance of the electronics technology, the Internet of Things (IoT) speeds up many revolutionary progress and applications. Among them, self-driving cars and healthcare are less tolerant to latency caused by data transmission because they need to provide real-time analytics. To reduce the amount of data transmission and increase the computation efficiency of data analytics, Fog Computing or Edge Computing, a hierarchical layered structure, is considered as a feasible way to build the IoT infrastructure. In this layered structure, we usually utilize the many-core systems to achieve high performance computing in the cloud layer. In the sensing layer, data are collected effectively by different low-power sensors. As for the edge computing, the computation is performed in the layers between end users and cloud to provide real-time service. However, the memory and computation hardware resources on edge devices become limited compared to the cloud.
To improve the efficiency, a feasible framework of fog computing is as follows: In the cloud layer, by performing high computational-complexity machine learning training algorithm with off-line collected data, we can obtain the inference models for analytics. In the sensing layer, data are collected effectively by different low-power sensors based on emerging Compressive Sensing (CS) technique. With the off-line trained inference models, edge computing can achieve real-time preliminary screening in the on-line phase. To avoid huge computation cost for signal reconstruction, edge computing analyzes signals directly in compressed measurement domain and achieves the goal of Compressed Learning (CL). In addition, by transmitting only abnormal data to the cloud layer for detailed analytics, it can further reduce the energy consumption.
However, reliability issues are more severe in low-power design and advanced process technology, and are elevating to dominating concerns in the design of IoT systems. Technology scaling, which has enabled high density core integration, also has a detrimental effect on the many-core system reliability due to increasing susceptibility to soft errors. Furthermore, variations in process parameters raise more complications to the reliability issues. Because of the exponential dependence of soft error rate (SER) on the process parameters, the variations for SER are significant. Such characteristic provides new challenges for designing reliable many-core systems in the cloud layer. On the other hand, in the CL framework, hardware non-idealities, such as thermal noise and interference, in CS sensors result in the degradation in learning performance. Therefore, we aim to solve the reliability issues caused by CS sensors while maintaining low complexity in the edge layers.
This dissertation presents a variation-aware core-level redundancy (VACLR) scheme for the cloud layer. By exploiting the inherent core redundancy of many-core systems, proposed VACLR scheme implicitly implements N-Modular Redundant sub-systems to achieve area-efficient fault-tolerant computing. We consider the impacts of variations in soft error rate, task vulnerability, and task significance on the reliability. This scheme dynamically allocates each replicated task to a proper core with variation-aware mapping algorithms to optimize system reliability under core resources constraint. On the other hand, this dissertation presents the ensemble of Extreme Learning Machine (ELM) framework that can achieve the accuracy for preliminary screening in resource constrained edge layers. By using principal components analysis, we develop the eigenspace transformation for compressed noisy data, and further utilize a subspace-based dictionary to remove the interferences directly in the compressed domain. In addition, we implement this engine in TSMC 90 nm technology with the proposed hardware sharing architecture. The postlayout results show that the proposed CL engine can provide competitive area- and energy-efficiency.
In summary, this dissertation presents several advanced reliable designs for the hierarchical layered IoT system. We expect the proposed frameworks to be the key techniques for the future intelligent IoT system.

並列關鍵字

Robust System Design ； Many-Core System ； Machine Learning

參考文獻

[1] M. S. Mahdavinejad, M. Rezvan, M. Barekatain, P. Adibi, P. Barnaghi, and A. P. Sheth, "Machine learning for Internet of Things data analysis: A survey" in Digit. Commun. Netw., 2017.

[2] G. Chrysos, "Intel® Xeon Phi coprocessor (codename Knights Corner)," in Proc. IEEE Hot Chips 24 Symposium (HCS), Cupertino, CA, 2012, pp. 1-31.

[3] Y. K. Chen, J. Chhugani, P. Dubey, C. J. Hughes, D. Kim, S. Kumar, V. W. Lee, A. D. Nguyen, and M. Smelyanskiy, "Convergence of recognition, mining, and synthesis workloads and its implications," Proc. IEEE, vol. 96, no. 5, pp. 790-807, May 2008.

[4] H. Sunami, "Dimension increase in metal-oxide-semiconductor memories and transistors" in Advanced in Solid State Circuit Technologies, Croatia, Rijeka:InTech, pp. 307-332, 2010.

[5] "Cisco Fog Computing Solutions: Unleash the Power of the Internet of Things," http://www.cisco.com/web/solutions/trends/iot/docs/computing-solutions.pdf, 2015.