基於深度學習於容器叢集自動擴展機制之分散式物件儲存服務系統

Deep Learning Based Auto-Scaling Load Balancing Mechanism for Distributed Software-Defined Storage Service

李麒傑(Ci-Jie Li)

指導教授 ： 陳弘明

共同指導教授 ： 陳世穎(Shih-Ying Chen)

國立臺中科技大學/資訊與流通學院/資訊工程系碩士班/碩士(2018年)

摘要

隨著雲端技術的普及化，使得資料中心的資料量也相對成長快速，許多公司紛紛推出自家公有雲服務，常見的公有雲像是Amazon Web Services (AWS)、Microsoft Azure、Google Cloud Platform (GCP)等。但若企業使用公有雲，不僅每年需支付一筆可觀的維運費用，對於企業內部敏感性較高的資料也存在相對較高的風險，因此使用私有雲儲存系統也是企業重要選項之一。除此之外，依照公司的規模需考量雲端伺服器系統如何同時承載極大流量的問題。典型的企業皆採用預算式方案，建置多臺高階伺服器叢集處理大量請求，但是仍無法準確預估未來系統可能的使用者流量，達到準確估算所需的運算資源。除此之外，在原有架構中加入更多的硬體設備對於企業不僅需支付更昂貴的設備費用，基於架構之上的應用服務也缺乏水平擴展的特性，同時無法提供有效的負載平衡機制解決大流量問題。為改善此一問題本研究採用開放原始碼系統並建置一套基於Ceph分散式軟體定義儲存系統之S3相容物件儲存服務，該系統利用AWS 儲存服務S3(Simple Storage Services)相容APIs開發儲存管理服務系統並建置於私有雲之上且基於Ceph分散式軟體定義物件儲存之服務，利用分散式儲存提升整體可靠性並且結合開源Kubernetes容器叢集調度技術與開源TensorFlow框架設計一套具深度學習模型之動態存取服務優化自動擴展機制，對於未來的使用者流量進行評估，進而讓服務副本數量與使用者流量達到最吻合狀態，以最少的服務副本數量來提供高可靠的服務品質與系統穩定性。

關鍵字

自動擴展 ； 負載平衡 ； 物件儲存 ； Amazon S3 ； Kubernetes ； Tensorflow

並列摘要

With the popularization of cloud technologies, the data of the data center is also growing relatively quickly. Each cloud provider start to provide its own public cloud service. The common cloud provider such as Amazon Web Services (AWS), Microsoft Azure, Google Cloud Platform (GCP) etc. These public clouds, however, also come with a handsome fee every year, as well as the security risks of sensitive internal data leakage. Therefore, using a private cloud storage system is also one of the important options for enterprises. In addition, according to the size of the company, it is necessary to consider how the cloud server system will bear the problem of receiving large amounts of traffic at the same time. General companies use budget plans that build multiple high-level server clusters to handle a large number of requests. However, it is still impossible to accurately predict the possible user traffic and required computing resources of future. In addition, the traditional high-level server clusters also require great delays and large amounts of human resources for service startup. Therefore, this research uses an open source system and builds an AWS Simple Storage Services (S3) based on distributed software-defined object storing system Ceph, whose distributed storing system increases reliability where a corrupted datum can be recovered from the backup. And combined with open source Kubernetes container cluster scheduling technology and open source TensorFlow machine learning framework, designing an auto-scaled load balancing mechanism with deep learning techniques. Evaluate the user's request in the future, so that the number of replicas and user traffic are in the best match. Provide highly reliable service quality and system stability with a minimum number of replicas.

並列關鍵字

auto-scaling ； load balance ； object storage ； Amazon S3 ； Kubernetes ； Tensorflow

參考文獻

[1] Mell, Peter, and Tim Grance. "The NIST definition of cloud computing."(2011)

[2] 2020年全球資料量將成長至44ZB. (2018). Retrieved from https://www.ithome.com.tw/article/87190

[3] 鐘嘉德、高天助、楊嘉栩(2010)：雲端運算與產業發展。研考雙月刊，34(4)，20-31.

[4] Barroso, L. A., Clidaras, J., & Hölzle, U. (2013). The datacenter as a computer: An introduction to the design of warehouse-scale machines. Synthesis lectures on computer architecture, 8(3), 1-154.

[5] Microservices. (2018). Retrieved from https://en.wikipedia.org/wiki/Microservices