Scheduling Algorithm Design for Virtualized Hardware Accelerator in C-RAN

Abstract

In order to improve energy efficiency of computation-intensive workloads in Cloud Radio Access Network (C-RAN), virtualized hardware accelerators (HA) are proposed in this paper. In C-RAN architecture, the base stations (BS) are running in the virtual machines. Virtualization mechanism of HA makes each BS feel like owning a HA exclusively, but actually BSs sharing the HAs. To tackle the problem of efficiently sharing HA by multiple BSs in the cloud platform, a scheduling algorithm with pipelined first-come-first-service (FCFS) was proposed in this paper, named PL-FCFS. PL-FCFS sets up two queues for tasks aiming to fully utilize HA with pipelining model. With this design, multiple virtual BSs can access the HA simultaneously and efficiently. To evaluate the performance of PL-FCFS, we use HA to accelerate turbo decoding of LTE physical layer(LTE PHY), and implement this scheduling algorithm for virtuaized HA in Xen paravirtualization environment. Experimental results indicate that four virtual BSs running LTE PHY with 5M bandwidth can be supported well by the PL-FCFS based evaluation system, and the performance is improved by 25% compared with FCFS scheduling algorithm.