Resource Prediction Based on Double Exponential Smoothing in Cloud Computing Authors: Jinhui Huang, Chunlin Li, Jie Yu The International Conference on Consumer Electronics, Communications and Networks (CECNet 2012) Speaker: Tzu-Chung Su

Outline Introduction Method Introduce Prediction Model and Scheduling Algorithm Experiment Result Conclusion Comment 2

Outline Introduction Method Introduce Prediction Model and Scheduling Algorithm Experiment Result Conclusion Comment 3

Introduction Cloud computing is an emerging paradigm that aims at streamlining the on-demand provisioning of software, hard-ware, and data as services, providing end-users with flexible and scalable services accessible through the Internet. How to predict the amount of resources is important! With higher accuracy, the idle rate and cost of system may be significantly reduced. Not only the current state but also the history records of resources should be considered. 4

Outline Introduction Method Introduce Double Exponential Smoothing Determine Smoothing Factor and Initial Values Prediction Model and Scheduling Algorithm Experiment Result Conclusion Comment 5

Double Exponential Smoothing The situation of time series is of stability and regularity. That is to say, the previous statements will continue into the near future. “Simple mean method” use all of the history records. “Moving average method” do not consider forward data. “Weighted moving average method” give the recent data greater weight. 6

Double Exponential Smoothing This paper uses Brown’s Double Exponential Smoothing, i.e. another smoothing of single exponential smoothing. And the prediction formula as below:, where 7 v.s. s t : prediction value of time t x t : actual value of time t α: smoothing factor(0≤ α ≤1)

Determine the smoothing factor α In general, the value of α should be greater when the data is prone to big swings which will increase the impact of the recent data to the result. There are two methods we can choose. decided by experience E.g. if the time series data is stable, α should be small( ). If data is prone to small swing, α can be a little greater( ). When data prone to bigger swing, α can be much greater( ). α should be (0.6-1) when data is of rising or declining trend. trial method E.g. try several α to calculate predict standard error then pick the one with the smallest standard error. 8

Determine the initial values In this paper, the authors choose the first data of the time series as the initial value when the item number is great(n>15). Inversely, they set the initial value as average of the previous few numbers when its item number is small(n<15). There are also some more accurate methods, such as choose several data(3 to 5) at both ends of the series. And then calculate the mean number of them. 9

Outline Introduction Method Introduce Prediction Model and Scheduling Algorithm Experiment Result Conclusion Comment 10

Prediction Model Two assumptions about prediction model are proposed to simplify the building of modeling. a) Each resource performance characteristic can be measured quantifiably, and described by a stream of measurements. b) We can monitor and gather the performance measurements non- intrusively. Two factors of the performance, CPU and memory, are considered. 11

Scheduling Algorithm Description Resource Management Log(RML) A database, keeps the resource information the customer used before Once the customer completes a job scheduling, the resources used this time will be added to RML. Below is the pseudo code of the algorithm: 12

Scheduling Algorithm Description 13 If the item number of time series greater than 15, set the initial value as the first data. Else, set the initial value of exponential smoothing as average of the previous few numbers. Set the smoothing factor by the standard deviation of actual value adjacent of the resource.

Outline Introduction Method Introduce Prediction Model and Scheduling Algorithm Experiment Result Conclusion Comment 14

Experiment and Result The experiment is on the CloudSim cloud simulator. In order to confirm the accuracy of double exponential smoothing prediction algorithm, we compare the predicting value with the actual value. Also, the authors compare this method with simple mean based method and weighted moving average method, which show the advantage of double exponential smoothing method. 15

Outline Introduction Method Introduce Prediction Model and Scheduling Algorithm Experiment Result Conclusion Comment 16

Conclusion The double exponential smoothing based resource prediction model fully consider the current data and the history records. Experiment proved that the model has a better performance comparing with the other two models. 17

Outline Introduction Method Introduce Prediction Model and Scheduling Algorithm Experiment Result Conclusion Comment 18

Comment This paper introduces the Brown’s DES in detail. Also, it tells the determine methods of smoothing factor α. However, we still have no idea on how the variable prediction period, i.e. m, be determined. A data record for each session type can be constructed. Question: Does session has a linear “trend”? 19