OLM: online LLC management for container-based cloud service

Abstract

In contrast to the hypervisor-based virtualization method, the container-based scheme does not incur the overhead required by virtual machines since it requires neither a fully abstract hardware stack nor separate guest operating systems (OSes). In this virtualization method, the host OS controls the accesses of the containers to hardware resources. One container can thus be provided with resources such as CPU, memory and network, expectedly isolated from the others. However, due to the lack of architectural support, the last-level cache (LLC) is not utilized in an isolated manner, and thus, it is shared by all containers in the same cloud infrastructure. If a workload of a container leads to cache pollution, it negatively affects the performance of other workloads. To address this problem, we propose an efficient LLC management scheme. By monitoring the memory access pattern, the indirect LLC usage pattern of a container can be figured out. Then, our proposed scheme makes two groups at runtime without using any offline profiling data on containers. The first group is made up of cache-thrashing containers, which fill up the LLC without any temporal locality of data, and the second one consists of normal ones. For isolation, the two separate groups use different partitions of the LLC by the OS-based page coloring method. Our experimental study suggests that the performance of a normal container can be improved by up to 40% in the case of using our proposed scheme.