Virtualized high performance (DPDK) packet processing

This proposal has been accepted as a session.


One Line Summary

Moving from 500Kpps to low latency and high packet rate - 2Mpps to 200Mpps - requires some evolutions of vNICs


vhost/virtio solution is the de facto standard for high performance with packet processing when Qemu is used. VMware/ESX has its vmxnet3 equivalent solution.

As it was demonstrated using DPDK, CPUs allow sustaining faster packet rate. The arrival of 2×40Gbps and soon 100Gbps network interfaces will add even more pressure to remove any overheads.

These solutions do not scale when low latency (sub 1us) and when high packet rate are being used (2Mpps and more).

Moroever, there are some new needs to create some direct VM to VM communications without any processing by the host/hypervisor.

The purpose will be to review some existing options:
– vhost-user
– virtio: both kernel and userland mode DPDK drivers (PMD)
– ivshmem


virtio, DPDK, vhost-user, ivshmem


  • Idvincentjardin


    Vincent JARDIN is 6WIND’s CTO since 2006. He leads the design of 6WINDGate, the company’s flagship embedded networking software solution for multicore processors, which was rapidly adopted beginning with 3G/4G core infrastructure equipment, then with high performance network appliances and now with high performance NFV. It provides a seamless path to bare metal and virtualized architectures.

    Vincent is also in charge of developing the company’s patent portfolio. He co-founded the Quagga project in early 2002, the well-known open source project for routing, and remains one of the contributors. He is also a major architect and founding member of, a major open source community that enables high performance network applications such as NFV.

    His main skills are in designing high performance packet processing solutions for open systems.