Multi-VNF performance characterization for virtualized network functions

Softwarization of 5G Networks–Implications to Open Platforms and Standardizations

“…Assuming that the issue of optimal function placement can be overcome or at least better understood as the packet core scales to support 5G and beyond, concerns over the packet forwarding capabilities of the elements in the core remain. Raumer et al [102], Kawashima et al [103] and Pitaev et al [104] each discuss the issue of traffic performance and detail potential bottlenecks in current solutions. Raumer et al [102] note that there is little detailed work in this area and go on to show significant performance degradation through chained links between forwarding elements built using both Open vSwitch (OVS) and Linux namespaces.…”

“…Pitaev et al [104], having documented the points in the virtualized system architecture where bottlenecks can occur, extend the performance measurement to include the FD.io Vector Packet Processing (VPP) data-plane which is comparable in performance with OVS-DPDK with slightly higher performance in terms of both throughput and latency. Both OVS-DPDK and FD.io are based on DPDK technologies so this is not a surprising result The results mirror those of Kawashima et al [103] and both agree that at present, in terms of wire-to-VM, the only viable option is SR-IOV for maximum throughput and minimum latency.…”

“…Lab results taken over the past few at the 5G Innovation Centre and presented later in this report agree with and verify these findings. As well as individual performance characteristics, Pitaev et al [104] look at the performance of VMs running concurrently in the same hypervisor connected to the physical infrastructure by a variety of technologies, namely OVS-DPDK, FD.io VPP and SR-IOV. There are several important conclusions to be drawn from this work including implications for multi-socket systems.…”

VNF Placement Optimization at the Edge and Cloud †

“…Thus, the processing delay depends on the type of the VNF that is used and the resource footprint allocated to the VM. In this work, we assume an optimal server configuration, avoiding multiple packet copy operations between NIC, OS, and VM (e.g., by using Single Root Input/Ouput Virtualization for our I/O technology), providing a meticulous Non-Uniform Memory Access (NUMA) socket design, and an efficient CPU memory access [46] that can minimize the effect of those in the overall processing delay. Thus, we assume that we can fully use the total processing capacity of the server.…”

Distributed Search for Exchangeable Service Chain Based on In-Network Guidance

“…Through communication and process via a series of intermediate VNFs from each user to its destination, a specific services are provided. The performance of VNF has been investigated in various environments [7], [8].…”