Section: New Results

Memory management in system software

Participants : Damien Carver, Jonathan Lejeune, Pierre Sens, Julien Sopena [correspondent] , Gauthier Voron.

Recent years have seen the increasingly widespread use of multicore architectures and virtualized environments. This development has an impact on all parts of the system software. Virtual machine (VM) technology offers both isolation and flexibility but has side effects such as fragmentation of the physical resources, including memory. This fragmentation reduces the amount of available memory a VM can use. Many recent works study that a NUMA (Non Uniform Memory Access) architecture, common in large multi-core processors, highly impacts application performance. We focus on improving the memory and cache management in various virtualized environments such as Xen hypervisor or linux-containers targeting big data applications on multicore architectures.

While virtualization only introduces a small overhead on machines with few cores, this is not the case on larger ones. Most of the overhead on the latter machines is caused by the NUMA architecture they are using. In order to reduce this overhead, in [34] we show how NUMA placement heuristics can be implemented inside Xen. With an evaluation of 29 applications on a 48-core machine, we show that the NUMA placement heuristics can multiply the performance of 9 applications by more than 2.

We also study the memory arbitration between containers. In the Damien Carver’s PhD thesis, we are designing ACDC [23] (Advanced Consolidation for Dynamic Containers), a kernel-level mechanisms that automatically provides more memory to the most active containers.

In the Francis Laniel's PhD thesis, we study a new architecture using Non Volatile RAM NVRAM. Although NVRAM are slower than classical RAM, they have better energetic features. We investigate solutions where RAM and NVRAM coexist in order to balance the energy consumption and performance according to the needs of the system.