Performance for VNFs

VNFs require high performance and deterministic behaviour, particularly in packet processing to reduce latency and jitter. Upstream communities like Openstack, OPNFV, libvirt and others are adding features specific for the NFV use-case. These features give service providers and their vendors tight control about performance and predictability of their NFV workloads, even during congestion or high-load situations.

Non-Uniform Memory Architecture (NUMA) aware scheduling allows applications and management systems to allocate workloads to processor cores considering their internal topology. This way, they can optimize for latency and throughput of memory and network interface accesses and avoid interference from other workloads. Combined with other tuning, we are, for example, able to reach >95% of baremetal efficiency processing small (64 bytes) packets in a DPDK-enabled application within a VM using SR-IOV. This represents between 10 to 100 times better performance than non-optimized private clouds 

Screenshot from 2016-03-16 23-48-40.png

The latest Linux kernel offers Real-Time KVM extensions that enable NFV workloads with the well-bounded, low-latency (<10μs) response times required for signal processing use cases like Cloud-based Radio Access Networks. It is also important to realize that KVM has recently improved its PCI and SR-IOV management (VFIO) to allow zero-overhead passthrough of devices to VMs

Open vSwitch has also received improvements, and further extensions with Intel’s Data-Plane Development Kit (DPDK), which enables higher packet throughput between Virtualized Network Functions and network interface cards as well as between Network Functions. Figures as high as 213 Million Packets per second for 64-byte packets on a x86 COTS server, representing the worst case scenario, proves that the technology has less than 1% performance overhead on the worst cases, and a even smaller impact on most real-world cases.

Additional performance measures are available at the operating-system, with tools like Tuned, with predefined profiles to change the way interruptions are handling, or the prioritization of the process scheduler.


One thought on “Performance for VNFs

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s