一般商業上在使用智慧計算雲端平台時，通常會利用智慧型平台管理介面 (IPMI) 來監控實體機器層級的運作，並在實體機器上安裝多層中介軟體(如虛擬機器)支援計算需求，並降低雲端平台的管理難度。根據不同的商業用途，這些平台可能會由用於各種用途的異質性計算節點(如一般伺服器與GPU伺服器)構成，以此來滿足各種不同的計算需求。商業上非常重視會造成商業損失的服務中斷時間(Service Downtime)，所以對於可減少中斷時間的高可靠度(High Availability)技術非常感興趣，此技術透過軟體自動偵測會造成服務中斷的故障後再透過IPMI等工具來復原服務，如此可在沒有人為的介入下持續支持服務的可用性。然而，現今多以Heartbeating機制作為高可靠度解決方案。這種機制將服務所在環境視為黑箱並只對服務來進行故障偵測，而根據偵測時間長短，此機制可能會有誤判或偵測時間過久的問題。本計畫主要針對由異質性計算節點構成的智慧計算雲端平台的高可靠度進行理論探討與技術研發，並針對其多階層特性與異質性提供高可靠度的解決方案。本計畫為三年期的整合型計畫「高安全等級智慧型行動應用惡意行為分析架構之技術研發」中的子計畫(子計畫三)「支援保護機制客製化與自動復原之雲端平台服務」。其它兩個子計畫主要目的在提供惡意行動程式檢測與智慧裝置使用者認證服務，而這兩者都需要使用具高可靠度的異質性智慧計算雲端平台(使用實體機器機器或虛擬機器)來支持服務的可用性。本計畫在第一年中會利用雲端平台多階層且階層相依的特性來設計保護機制客製化的功能，這種功能主要是基於異質運算節點仍有大部分相似的層級 ，只要抽換部分的故障種類即可完成兩故障集之間的切換，而且不影響階層式特性，所以預期只需透過一高可靠度系統即可保護具異質性的智慧計算雲端平台。第二年我們預計將高可靠度保護延伸到虛擬機器層級，所以我們會研發虛擬機層級的故障偵測與自動復原方法。第三年我們會研發階層分群故障診斷方法，而這是一種更有效率的故障診斷方法，它會將原有的偵錯階層分群，預期可以達到部份故障平行化偵測並減少故障診斷時間。 ;High availability (HA) is a critical mechanism to protect users' services/applications from unexpected failures. Supporting HA on a multi-layered, heterogeneous cloud platform is a challenging issue in both the academia and the industry. Existing failure protection solutions usually treat each compute node as a single-layered system, and rely on a heartbeating mechanism to detect the liveness of critical services/applications. The major problem of this approach is that, the heartbeating mechanism cannot identify the root cause of a failure, and thus prohibits fine-grained failure management. In addition, heartbeating is very sensitive to a predefined heartbeat waiting time, which may result in long failover time or wrong decisions. In this proposal, we aim to develop a novel customizable fault detection and recovery mechanism, which provides fine-grained HA protection to critical cloud services/applications running on a multi-layered, heterogeneous cloud platform. We plan to operate this project in a three-year period. In the first year, we will develop a customizable fault detection and recovery mechanism for multi-layered, heterogeneous cloud platforms, in particular for the compute nodes with different layers (ex. with/without VM support). We will also implement the new design on OpenStack. In the second year, we will apply our new HA design to the virtualization layer of the platform, in order to support finer-grained fault detection and recovery. In the third year, we will develop and implement a novel layer-grouping fault detection mechanism that can concurrently detect the root cause of a failure in a more efficient way.
