近年來排隊理論廣泛被應用於通訊、計算、網路等領域。由於科技日新月異，許多的網路系統已具有「可資源共享」的特性。此特性使得系統更加複雜化，這也使系統的各種表現值(performance)不易分析。本文考慮一可資源共享的非循環(acyclic)網路系統並針對系統延時(delay)之最佳化提出一動態控制策略。此策略以連續層級的流體控制(fluid control)問題描述系統延時之最佳化資源配置(allocation)問題，由於此最佳化問題需仰賴該時間點的工作輸入流量強度(traffic intensity)以解出最佳資源配置，故衍生出工作輸入流量強度估計問題。而本文採指數加權移動平均來估計此流量強度，並引入多變量指數加權移動平均控制圖進行流量強度變動的偵測，以解決速率變動頻繁導致最佳化的計算與伺服器配置切換之成本(cost)問題。而文末分別以獨立的卜瓦松輸入過程與相關的再生(renewal)輸入過程進行模擬以探討延時等表現值，並由模擬結果驗證此策略在這些輸入過程中的某些情況下可在大幅降低伺服器重新配置次數的同時維持一定水平之平均等待時間。 In this paper we consider a general model framework for acyclic stochastic processing networks with shared resources.Motivated by a fluid control problem of the model,we propose a measurement based control policy that improves QoS performance with respect to delay metrics.The policy employs statistical techniques involving estimation and shift-monitoring of traffic intensities so that it is responsive to traffic fluctuations and minimizes the number of switchings between service configurations.The performance of the policy is illustrated on a number of queueing systems with different types of input traffic.
