A Case Study of Container Behavior Analysis with Cloud Resource Profile

In the field of cloud computing, a container is a lightweight abstraction of an operating system and the software application running for a particular service. To select a machine to deploy the container, the scheduling component of a cloud computing system usually considers the behavior of the container for decision. Several recent studies have shown that the prediction of the future behavior of a container can provide helpful guidance for scheduling. In this paper, we exploit the Alibaba cloud scheduling trace data which was generated on the year 2018 and named cluster-trace-v2018, to analyse containers behavior. For cluster-trace-v2018, we report our method for the prepossessing, container behavior feature engineering, and behavior classification. Specifically, we developed a processing procedure to filter out incomplete records. We designed a method for building behavior feature vector for containers on 12-hours long and 24-hours long periods. And, based on these features, we utilize the K-Means method to cluster the containers into a specified number of clusters. Experimenters on varying length of container feature vector show that the containers fits well into 10 clusters.