Icebreaker: A Platform for Service Dependency Detection

Numerous enterprises are migrating their applications onto cloud platforms to gain benefit from cloud computing technology. These applications are composed of several service components with complicated dependency structures, especially for network applications which commonly take serial collaboration of service components. We call a group of VMs belonging to the same application a VM family. Finding out the family relationship will make cloud vendors more capable on VM consolidation and traffic pressure reduction. In this paper, we present a platform named icebreaker to assemble VMs belonging to the same family without sniffing tenants´ privacy. Icebreaker characterizes each VM with basic resource consuming information and proposes an iEntropy distance to distinct VM nodes. We adopt adaptive Affinity Propagation (AP) clustering algorithm and maximum Silhouette index to adaptively identify the number of VM families and categorize them. Experiments are conducted on RUBiS, Hadoop and Apache Bench clusters with 169 VMs. Testing results show that icebreaker can get 96.45% accuracy and make distinct performance improvement on RUBiS cluster.