Modern cloud platforms rely on containers in order to deploy applications and allocate resources to them. Users of Container-as-a-Service platforms interact with another layer of abstraction, container orchestrators, to set resource allocations. Regarding the CPU allocation, orchestrators can use one of two strategies to apply the specified allocation: (1) the allocation of cores, reserved for one application; or (2) the allocation of quotas, which can be provided by any of the available processors. However current orchestrators only use the quota strategy. We benchmark both, demonstrating that the quota strategy can show up to 68% of degradation in our experiments when compared to the first strategy. We identify that this degradation comes from violating what we call the What You See Is What You Get (WYSIWYG) principle: a container's view of its available resources is wrong under the quota strategy. We state that a better trade-off can be found in combining these two strategies, and we design a hybrid resource allocation algorithm that can be integrated into any container orchestrator. Our evaluations show that it prevents resource management problems that come from allocating cores, while canceling the performance overhead associated with the quota allocation strategy that violates the WYSIWYG principle.
展开▼
