Ron-Renwick-blog.original.jpg

Netronome and Qualcomm: Together Enabling Operators to Realize the Cost and Efficiency Promises of NFV

By Ron Renwick | Nov 08, 2017

The promise of NFV has always been that large, powerful, network-centric applications can be moved from purpose-built appliances and onto commodity servers. This approach would provide many benefits including:

  • Reduced costs by moving to commodity infrastructure
  • Improved flexibility to scale and migrate workloads
  • Rapid deployment of new features and services.

Making this happen is one of the challenges that NFV has faced to date. Many virtual network functions (VNFs) are compute intensive, requiring many CPU cores, while at the same time having high network I/O requirements.  These two things compete with each other for server resources, impeding server efficiency and precluding operators from putting multiple VNFs on the same server. 

The NFVi layer, and in particular virtual switches such as OVS, are typically a drag on server efficiency and becomes a bottleneck that limits top-end performance. Using enhanced virtual switching in the form of OVS-DPDK has provided some improvement for operators. OVS-DPDK can greatly improve server virtual switching performance, but it still burns CPU cycles to make that happen. Back to the earlier challenge – VNFs need a lot of CPU cores.

Another solution would be to increase the CPU core count. If VNFs need a lot of CPU cores and you give them more cores, then it stands to reason that an operator could run more VNFs.

Qualcomm has done just that. The Centriq 2400 is a 48-core ARM processor-based server. By significantly increasing CPU core count, Qualcomm has made it possible for operators to run more VNFs on each server. That goes a long way, but now the larger number of VNFs need even more data, and the virtual switch is still the bottleneck, ultimately limiting performance yet again.  

Netronome Agilio SmartNICs are the solution to this problem. By offloading the virtual switching layer to the SmartNIC, the server CPU resources are freed from having to perform this processing.  This gives back more CPU resources to the VNFs, while at the same time dramatically improving network I/O throughput via acceleration of the virtual switching function on the SmartNIC hardware.  

Netronome and Qualcomm have done a comparison demo of a Qualcomm Centriq server with Agilio OVS against a traditional x86 server and NIC running OVS-DPDK. You can see the demo at our booth at the Centriq launch event, and if you can’t make it to the event, you can view a video of the demo here.   

As you can see from the example below, Centriq + Agilio can run twice as many VNFs per server, while improving the network I/O throughput by 4X, and reducing CPU utilization by almost half, when compared to a traditional x86 + XL710 NIC running OVS-DPDK.

Function                              x86 w/ OVS-DPDK           Centriq w/ Agilio SmartNIC

VNF/Server                                          3                                           6

Server Throughput                           6Gb/s                                 24Gb/s

CPU Usage                                        85-90%                              50-60%

More detailed performance testing was done in our lab and can be found in the following solution brief.

In the end, the comparison is simple. With Qualcomm Centriq and Netronome, operators will be able to deploy more VNFs with better network throughput to each VNF than if they were to use traditional servers.  This will finally allow operators to realize the cost and efficiency promises of NFV and make it a deployable reality.