Towards LTE physical layer virtualization on a COTS multicore platform with efficient scheduling

This paper explores the runtime behavior of a class of multiprocessor systems in specific contexts associated with cloud radio access network implementation. It specifically deals with task scheduling, run time behavior, and their characterization. It also relates to Network Functions Virtualization (NFVs) and especially with the constraints associated with virtualization of a Long Term Evolution (LTE) stack. To validate the effectiveness of different scheduling algorithms, an emulation of an LTE uplink virtualized stack is made. Experiments are carried out using a runtime system, StarPU, coupled with profiling tools, which allows characterizing the need for dedicated threads or cores to manage tasks within a server. Reported experimental results confirm the feasibility of software scheduling of an LTE uplink stack with two of the six tested algorithms. In addition, this paper explores distributing the scheduling load across multiple computing units, which is more efficient than implementations where the scheduler is centralized on a dedicated processor.