A flaming act in the Internet, or a hostile and insulting interaction between the Internet users, caused by an inappropriate personal behavior results in a large number of accesses, thus, high load generated on the part of servers as well as network equipment hosting the Internet services, and often causes a denial of service situa- tion. Automatically provisioning server resources is a solution for mitigating such situation, and one such example is Amazon Web Service (AWS) Auto-Scaling. However, we posit that frequent fluctuation of workload (so called noise) around the threshold for scaling may incur extra cost in hosting services caused by superficial provisioning in the current auto-scaling methods. In order to solve this problem, it is necessary to automatically predict such flaming acts via collective knowledge on people’s behavior in the Internet, e.g., on Social Network Services (SNS). Therefore, in this paper, we propose a method for automatically detecting the Internet flaming acts in Twitter, and changing the algorithm of auto-scaling from the traditional server workload basis to workload fitting one so that the provisioning endures noise. In more detail, we collect tweets including words such as ”Enjyo” or ”Bakatter”, classify these tweets using machine learning techniques and attempt to detect the Internet flaming acts. Once such flaming acts are detected, we start filtering noise using EWMA in measured values of CPU utilization of the servers to detect the optimal point of provisioning resources. We evaluate how effectively our proposed method infers the Internet flaming acts, and also prevents extra cost caused by noise dealing with server load efficiently.

田原俊一, 中尾彰宏. “SNSによるアクセス集中予測を利用したサーバの自動スケール制御手法”. 信学技報, vol. 115, no. 483, NS2015-254, pp. 499-504, 2016年3月. copyright©2016 IEICE