Section: Application Domains

Capacity Planning in Cluster, Grid and Cloud Computing

Cluster, Grid and more recently Cloud computing platforms aim at delivering large capacities of computing power. These capacities can be used to improve performance (for scientific applications) or availability (e.g., for Internet services hosted by datacenters). These distributed infrastructures consist of a group of coupled computers that work together and may be spread across a LAN (cluster), across a WAN (Grid), and across the Internet (Clouds). Due to their large scale, these architectures require permanent adaptation, from the application to the system level and call for automation of the corresponding adaptation processes. We focus on self-configuration and self-optimization functionalities across the whole software stack: from the lower levels (systems mechanisms such as distributed file systems for instance) to the higher ones (i.e. the applications themselves such as J2EE clustered servers or scientific grid applications).

In 2013, we have confirmed the scalability of the DVMS proposal by conducting experiments on a very large scale involving up to 5K virtual machines (VM)s upon 500 nodes, thus establishing it as one of the most scalable placement algorithm for virtual machines. Moreover, we have extended the SimGrid framework by adding virtualization abstractions for hundreds of thousands of VMs. Finally, we have also provided several results on the energy efficient management of Cloud applications and infrastructures, see Sec.  6.3 .

In the energy field, we have designed a set of techniques, named Optiplace, for cloud management with flexible power models through constraint programming. OptiPlace supports external models, named views. Specifically, we have developed a power view, based on generic server models, to define and reduce the power consumption of a datacenter's physical servers. We have shown that OptiPlace behaves at least as good as our previous system, Entropy, requiring as low as half the time to find a solution for the constrained-based placement of tasks for large datacenters.