Memory management in a virtual environment

Overview & Synopsis

The increasing computation needs of modern applications led to the development of medium to large scale computer infrastructures. Cloud computing became a key solution for resource mutualization. More and more enterprises and scientific institutions set-up their own private cloud facilities. The key technology behind the development of these infrastructures is virtualization. Virtualized infrastructures bring many advantages for resource management, but resource optimization is still a challenge, for ensuring high hardware utilization and low waste. Server consolidation was introduced for optimizing such infrastructures. Its principle is to gather VMs on as less servers as possible, thus allowing unused servers to be suspended. However, consolidation strategies have to take into account many types of resource (CPU, memory, I/O) thus inducing a high complexity. Additionally, these resources may be fluctuating at runtime and consolidation relies on VM migration which is a heavy operation. Therefore, consolidations are performed infrequently. In this thesis, we propose the design of a memory management system which allows mutualizing memory between VMs. This system is orthogonal and complementary to consolidation. The first issue is to design a monitoring system that should track the working set of the VMs at runtime with low intrusiveness. The next important step is to reclaim unused memory from under-loaded VMs and finally grant it to over-loaded VMs to help them to overcome temporary memory difficulties. As a result, it is proposed a both local and global memory mutualization system which allows to monitor the working set size of each VM and to mutualize unused memory, either locally (with the VMs located on the same node) or globally (with the VMs located on neighbor nodes). The solution has been evaluated with modern HPC, data intensive benchmarks as well as scientific and Big Data (Apache Spark) applications. The results demonstrate the effectiveness of our design choices.