White Papers

Understanding OPNFV

Issue link: https://read.uberflip.com/i/1344850

Contents of this Issue

Navigation

Page 71 of 144

Understanding OPNFV 71 VPP processes a number of packets in parallel instead of one at a time. This spreads the overhead of lookups and computation across an entire set of packets contributing to the added efficiency. VPP exposes a high-performance low level API. For software stacks that need to use a higher-level mechanism to communicate with VPP, another FD.io technology called Honeycomb translates YANG models exposed via Netconf/Restconf to VPP APIs. A controller which supports Netconf/YANG, such as ODL, can "mount" the Honeycomb Management Agent to communicate with the VPP. Independent testing has shown FD.io throughput to be about 5x better than OVS with DPDK (see next section) when forwarding to 2,000 IPv4 addresses and about 39x better when forwarding to 20,000 IPv4 addresses. FD.io, you guessed it, is also a Linux Foundation project and the Platinum members of FD.io include Cisco, Ericsson and Intel, with 12 additional Gold and Silver members. Data Plane Acceleration Given the importance of performance, as measured by packets per second, latency or throughput, special technologies are required to optimize data plane performance. This requirement is unique to NFV and not typically shared by enterprise use cases that are generally compute intensive. OPNFV integrates three primary data plane acceleration technologies. The OpenStack Ocata release with Nova's Placement API will enable better control over the placement of virtual machines associated with specific data plane acceleration resources. DPDK and Other Performance/ Scaling Features DPDK, another Linux Foundation project, is a set of libraries that bypass the kernel and provide polling mechanisms, instead of interrupt based operations, to speed up packet processing. In a recent study, OVS with DPDK showed a 75% improvement in throughput over plain OVS. The technology is available on multiple processors and is being promoted by Intel® as part of a broader portfolio of Enhanced Platform Awareness (EPA) technologies aimed at accelerating the data plane. The main technologies in EPA other than DPDK are huge pages, NUMA pinning and SR-IOV. Huge pages improve VNF efficiency by reducing page lookups, NUMA pinning ensures that the workload uses memory local to the processor, and SR-IOV enables network traffic to bypass the hypervisor and go directly to the virtual machine.

Articles in this issue

Links on this page

view archives of White Papers - Understanding OPNFV