Scalable and energy-efficient deployment of fully virtualized 10  Gb/s OLTs on multi-node GPU servers

To address the rapid diversification of service requirements in optical access networks while supporting legacy systems, network function virtualization (NFV) of access equipment is essential. While the utilization of virtualized systems comprising control and management layer software and whitebox optical line terminals (OLTs) is beginning to spread, studies are underway to fully virtualize OLTs, including the physical (PHY) layer, in order to further increase flexibility and efficiency. Although previous studies have demonstrated 10G-EPON functionality through parallel implementation on graphic processing units (GPUs), the evaluation was limited to the operation on a single server that did not take into account scaling, and the power consumption was excessive, at approximately 51 W per OLT. This paper proposes an architecture and lightweight resource allocation algorithms to enable scalable and energy-efficient deployment of a large number of fully virtualized 10G-EPON OLTs on multi-node GPU servers. While detailing the control architecture of the OLT processes on multi-node servers, we formulate the problem of efficiently allocating active OLTs to GPUs as a bin packing problem and solve it using heuristic approaches based on first fit and next fit algorithms. Evaluations show that 40,960 users can be accommodated on a 16-node GPU server cluster and when the traffic volume of each OLT is low, the proposed algorithm, which operates in a processing time of less than a few milliseconds, can reduce the GPU power consumption to 0.594 W per OLT by consolidating 320 OLTs into one GPU.