Mécanisme de surveillance dynamique du trafic et de ressources dans des systèmes IMS virtualisés et distribués

Abstract

Cloud computing exploits computing power and storage in mainframe through a network. With this technology, infrastructure managers have the opportunity to create flexible systems that adapt quickly to their customers’ needs regardless of their location. The problem is to successfully predict the moment when more resources will be required by the client. This is why infrastructure managers need to constantly predict adequate amount of resources required by theirs clients. This prediction aims to deploy enough resources for the client to respect the level agreement signed between the two parties, while the utilization of these resources remain optimized in order to save energy and money. During this project, we have focused on the dynamic management of a virtualized IP Multimedia Subsystem (IMS) system’s resources. With this goal, we have developed a tool adapting the number of useful cores for processors of S-CSCF nodes in the system. The system virtualized during the project looks like systems used by operators of landline and mobile services to provide multimedia services for their subscribers. We have tried to find out whether we could predict users’ traffic on the network to be able to dynamically and proactively manage resources of its CSCF (call/session control functions) nodes. Before developing the previously mentioned tool, we have studied the virtualized IMS system. To make this possible, we have detected and analyzed several high and low level metrics for studied nodes. Through simulation of different scenarios, collection of the studied metrics data and analysis of this data, we can better understand the system. We can also note that P-CSCF nodes in the system seem not totally reliable because they stop without explicit reasons. We have also found that the main cause of system shutdown was a lack of CPU for S-CSCF containers. By studying them more closely we have noticed, through simulation of different scenarios, that the workload of their CPU as a function of number of calls per second follows an exponential model. Although we cannot predict the behavior of users, monitoring this resource allows us to allocate more cores at S-CSCF containers before they stop due to overload. This dynamic and proactive CPU allocation lets our system to optimize resources usage.

Date	3 May 2016
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Nadjia Kara (Supervisor)