Time-constrained and network-aware containers scheduling in GPU era

Abstract The recent advances on data center management and applications development are reflected by lightweight containers technology and critical Quality-of-Service (QoS) requirements. Tenants encapsulate applications in containers abstracting away details on the infrastructure, and entrust its management framework with the provisioning of network and time QOS requirements. In this paper, we addressed this NP-hard scheduling problem proposing a GPU Accelerated Containers Scheduler (GPUACS). We model the joint allocation of network and containers with QoS requirements as a graph embedding problem. GPUACS innovates by refactoring two Multicriteria Decision Makings (MCDMs) to GPU model, as well as by defining an efficient data structure to speed up the comparison of time-evolving QoS requirements. GPUACS follows a modular and configurable architecture, and the scheduling objective function can be adjusted by selecting the MCDM method and setting the appropriated weights to guide the comparisons. An experimental analysis demonstrated the sensitivity that GPU-tailored MCDM methods have to schedule container requests considering critical time, network, and processing criteria, as well as multiple queueing policies.