基於雙指數平滑法預測系統資源使用量

Abstract

雲端運算議題近年來相當熱門，各家科技公司也推出了自己的雲端服務，例如Google Compute Engine、Amazon Web Services、Microsoft Azure，由於傳統的實體機器無法即時調整系統架構，不少企業紛紛投向雲端懷抱，絕大部分的網路服務都採用多層式架構，搭配雲端服務可以彈性選擇要擴展或是刪減特定層數的資源，即時的處理能避免資源不足或浪費。 適當的資源配置在基礎設施即服務（Infrastructure as a Service）中非常重要，一般都採用自動擴展（Auto Scaling）來增減系統資源，系統負載變高的時候就自動給予更多資源，缺點是由於新增虛擬機器需要暖機時間，會有一定程度的延遲，這段期間可能會因為負載持續增加而導致當機，本論文採用時間序列法預測系統資源使用量，包含簡單移動平均法、加權移動平均法、雙指數平滑法，我們以TPC-W基準來做測試，蒐集資料加以分析後決定使用雙指數平滑法作為預測演算法，並設計一套從監控伺服器、預測使用量、雲端資源配置全自動化的系統。

Cloud computing has become popular in recent years. Google, Amazon, Microsoft, etc. has release their own cloud service. Many companies around the world started to use cloud service as the system architecture of physical machines cannot be adjusted instantly. The other reason is that most web service providers use a multi-tiered architecture, but with elastic cloud service, increasing or reducing the certain tier’s resources can avoid a waste of resources. Resources allocation is very important in IaaS (Infrastructure as a Service), generally it uses auto scaling to deploy system resources. If the system workload is heavy, then it will add more resources to it. However, adding virtual machines cost warm-up time and there will be some degree of delay, and during this period the system might be corrupted. In this thesis, we use time-series algorithms which included simple moving average, weighted moving average and double exponential smoothing to predict system utilization, and we also performed simulations on TPC-W benchmark. After analyzing collected data, we chose double exponential smoothing as final prediction algorithm, then we designed a fully automatic system that able to monitor server, forecast usage and allocate resources.