09-01

Building efficient network dataplanes in software

In this talk we give a survey of solutions -- and especially, discuss the underlying design principles -- that we developed in recent years to achieve extremely high packet processing rates in commodity operating systems, for both bare metal and virtual machines.

By using simple abstractions, and resisting the temptation to design systems around performant but constraining assumptions, we have been able to build a very flexible framework that addresses the speed and latency communication requirements of both bare metal and virtual machines up to the 100 Gbit/s range. Our goal is not to provide the fastest framework in the universe, but one that is easy to use and rich of features (and still, amazingly fast), to taking away concerns on the dataplane's speed from (reasonable) network applications.

Our NETMAP framework, opensource and BSD licensed, runs on 3 OSes (FreeBSD, Linux, Windows); provides access to NICs, host stack, virtual switches and point-to-point channels (netmap pipes), running between 20 and over 100Mpps on software ports, and saturating NICs with just a single core (reported up to 45 Mpps on recent 40G nics); achieves bare-metal speed on VMs thanks to a virtual passthrough mode (available for Qemu and bhyve); and can be used with no modifications by libpcap clients.

Luigi Rizzo is an associate professor at the Universita` di Pisa. He has worked on network emulation, high performance networking, packet scheduling, multicast and reliable multicast.  He is a long time contributor to FreeBSD, for which he has developed several subsystems including the dummynet network emulator, the ipfw firewall, and the netmap framework.  He has been program committe member for for sigcomm, conext, infocom, nsdi, Usenix ATC, ANCS and other conferences, as well as PC chair for Sigcomm 2009 and Conext 2014, ANCS 2016, and general chair for Sigcomm 2006. Luigi has been a frequent visiting researcher at various institutions including ICSI/UC Berkeley, Google Mountain View, Intel Research Cambridge, Intel Research Berkeley.