摘要
隨著歐洲電信標準協會提出網路功能虛擬化的架構，電信業一直嘗試解決高昂硬體設備成本與維運費用的痛點。然而新的架構伴隨著新的挑戰，即雲端運算資源如何隨著負載變化而進行調整，以維持一定的服務品質。此問題雖已有不少研究，但針對網路功能虛擬化所需之資源擴展，仍有一些問題待釐清，例如資源供給與效能間的邊際效用問題未被考慮，很少文獻利用實際系統驗證，鮮少同時考量水平與垂直擴展等。有鑑於此，本文基於網路功能虛擬化架構之真實雲端系統，以虛擬IP多媒體子系統作為使用場景，建構一個擴展預測模型。此模型由實測數據建構之廻歸所建立，其擴展建議不僅可取得當前條件下的最佳接通率，也如實反應系統在特定資源與情境下之行爲，並且接通率評估乃同時考量水平與垂直擴展之結果。鑒於資料非線性特性，建模工具則選用廣義可加性模型。模型訓練完成後之R平方值可達約0.8，顯示其模擬的預測能力可在接受範圍內，而當實際系統需要擴展時，預測模型可依當前條件，預測較佳成功接通率的資源組合與擴展方向，以利系統在維持高成功接通率下，承接可變動的負載。實驗結果顯示本模型在輸入預期負載和當前資源組合，同時評估水平與垂直擴展後接通率後，提供可獲得較佳成功接通率的擴展建議，其成功接通率之預測值與實際值的誤差小於1.8%，擴展建議方向也與實際情境相符。此外，本模型亦提供系統超負荷警示功能，其結果顯示可成功預測系統即將超負荷，相關資訊並可做爲服務供應商營運時的重要參考。
並列摘要
As per European Telecommunications Standards Institute (ETSI) proposed a reference framework for network function virtualization (NFV), the telecom industry looks forward to alleviating the pain points of the hardware and operating costs. The new architecture brings new challenge for cloud computing to allocate resources with dynamically changed loadings while maintaining certain quality level. In literature, fewer studies consider both horizontal and vertical scaling under a system validation. This paper takes the virtual IP multimedia subsystem, a network function virtualization architecture, as a NFV application to model the scaling prediction in the real cloud system. The proposed scaling model not only achieves the best successful call rate with the consideration of both horizontal and vertical scaling, but also reflects the system behavior under specific resources and situations. A generalized additive model is used as the modeling tool because of the non-linear nature of empirical results. The R-squared value of the model after training can approach to 0.8, which shows the predictive capability of the model is acceptable for a real application. When the practical system has to scale the resources, the model can predict the combination of resources and the scaling strategy for a better successful call rate. As a result, the cloud system provision sufficient resources to adapt the varied loading while retaining a high successful call rate. Experimental results show that the model can provide scaling recommendations for achieving an optimal successful call rate with the given information about the expected loading and the combination of resources at present. The error between the predicted and actual successful call rates is less than 1.8%, and the proposed scaling strategy is consistent with a practical scenario. The model also provides a warning function for a system overload. The test results illustrate that the model can predict the system overloaded in advance for the service operation.