에너지 절감형 서버 클러스터 환경에서 QoS 향상을 위한 소비전력 예측

Abstract

In an energy saving server cluster environment, the power modes of servers are controlled according to load situation, that is, by makingON only minimum number of servers needed to handle current load while making the other servers OFF. This algorithm works well undernormal circumstances, but does not guarantee QoS under abnormal circumstances such as sharply rising or falling loads. This is because thenumber of ON servers cannot be increased immediately due to the time delay for servers to turn ON from OFF. In this paper, we propose anew prediction algorithm of the power consumption for improving QoS under not only normal but also abnormal circumstances. The proposedprediction algorithm consists of two parts: prediction based on the conventional time series analysis and prediction adjustment based on trendanalysis. We performed experiments using 15 PCs and compared performance for 4 types of conventional time series based prediction methodsand their modified methods with our prediction algorithm. Experimental results show that Exponential Smoothing with Trend Adjusted (ESTA)and its modified ESTA (MESTA) proposed in this paper are outperforming among 4 types of prediction methods in terms of normalized QoSand number of good reponses per power consumed, and QoS of MESTA proposed in this paper is 7.5% and 3.3% better than that ofconventional ESTA for artificial load pattern and real load pattern, respectively.