提供擴展建議的自動瓶頸偵測系統

Abstract

瓶頸偵測一直以來都是系統服務上線前很重要的一環，其功能是為了保障服務品質，讓使用者能有更好的體驗，然而大部分瓶頸偵測是相當麻煩的，其所消耗的時間長且相當不方便。 現在的網路服務越來越複雜，其中大多數是採取多層伺服器的架構。在多層伺服器之中，因為不同的伺服器之間可能會互相影響效能，所以瓶頸的判斷相當不易。然而大部分多層伺服器的瓶頸偵測演算法是需要根據服務水準協議(Service Level Agreement)來做判斷的，因此，如何定義服務標準的門檻值也成為了服務提供者的麻煩。 鑒於以上的問題，本研究提出一個自動化的瓶頸偵測系統，在測試、搜集資料與分析瓶頸期間完全自動。我們也改進了他人的多層伺服器瓶頸偵測演算法，讓我們的系統能夠適用於各種不同種類的服務，管理者也可以不用定義服務標準的門檻值，僅透過資源使用率的變化就能得知伺服器中的瓶頸。在實驗中，我們使用 Benchmark TPC-W 來驗證我們的系統，我們的系統能夠成功地找到確實的瓶頸，在我們增加瓶頸資源之後，伺服 器的表現也有更進一步的提升。

Bottleneck detection is one of the most important things before service on-line. It ensures the quality of service. However, most bottleneck detections take long time and are inconvenient for service providers. In recent years, web services become more and more complex. Most of them are based on n-tier architecture. Because of different servers influencing each other are possible, bottleneck detection is difficult. Besides, most bottleneck detection algorithms are based on Service Level Agreement (SLA), how to define the threshold of SLO (Service Level Objective) causes difficulty for service providers. To solve the problems we mentioned above, we propose an automated bottleneck detection system which is fully automatic during the testing, data gathering and analysis. Service providers do not require much technical knowledge, only needs to input some test parameters then wait for the bottleneck results. We also improve the algorithm which is proposed before. The improved algorithm is applicable for different types of services and intuitive. We can find out the bottleneck in servers by the changes of resource utilization and without defining the thresholds. In the experiment, we use benchmark named TPC-W to verify our system. Our purposed system detects the bottleneck in the system and gives a scaling recommendation. After we scale up the metric which we determine as the bottleneck, the performance of service is improved. It is convenient for service providers and trusted to use.