支援雲端應用效能管理之基於事件知識的需求預測

Abstract

在網路應用服務的效能控管上最常面臨到的問題就是如何在面臨突如其來的大量需求下，依然可以保證效能目標。由於雲端技術日趨成熟，讓運算資源隨需求取得的公用運算(Utility Computing)有實現的可能。其中獲益最大的就是需求變動劇烈的網路服務，例如:線上球賽轉播或者訂票服務系統都是這類型的服務，我們預期特定的事件會導致系統使用量大幅上升。透過雲端的具備的彈性(elasticity)以及突發性(burstability)能夠讓這類型的網路服務在高度變動的需求下仍舊提供其所需的資源，然而提供資源需要花費時間(resizing time)及系統重新配置的成本及風險，因此在動態資源分配的機制下，預測需求是相當重要的環節。 本篇論文提出一套結合學習方法以及即時資訊的預測演算法，根據過去事件的變化行為利用學習方法了解造成應用程式需求變動的事件知識，加上即時的需求測量結果，預測未來一段控制區間中的需求變化，以達成預測服務需求的目標。同時為了避免低估需求導致目標效能違反地的情形，我們考慮需求本身存在的變動性(varation)，提出並比較不同的安全邊際(Safety Margin)做法，找出其中有效降低低估發生機率的做法。 在實驗結果中顯示我們的方法比起傳統的回歸預測更有效的提高預測的準確度，並且在考慮安全邊際後大幅度的降低發生需求低估的機率。另外我們的方法能夠應用在需要耗費時間進行資源分配的虛擬環境中，增加了預測方法本身的實用性。

In Application Performance Management (APM), the most common problem encountered by the administrators of many network service providers is how to sustain performance targets for the applications. Cloud computing offers the possibility to provide the resources on demand as Utility Computing. The internet services with highly changing demand can benefit from the implementation of cloud services rapid elasticity technology as live sport game broadcasting or online ticket booking system. We expect that specific events will cause significant increase the demand of applications. Cloud computing is elasticity and burstability, so it can help the internet services acquire necessary resources under the heavy variation of demand. However, system reallocation will take a resizing time and bring some cost and risk. The external demand prediction is an important component in dynamic resource allocation to provide target performance guarantees in cloud. In this work, we propose a learning-based prediction model and the real-time algorithm to forecast the external demand for this class of cloud applications. We use the learning method to understand the event knowledge based on the behaviors of historical events and consider the online measurements to predict the trend of external demand in next control period. We also develop safety margin-based prediction schemes to avoid the under-estimation errors of prediction. The experimental results show that our prediction method has more accurate prediction results than the traditional simple linear prediction methods. The use of safety margin only incurs a very small probability of under-estimation.