Netmap - a framework for fast packet I/O
	VALE -  a Virtual Local Ethernet using the netmap API
========================================================================

NETMAP is a framework for very fast packet I/O from userspace.
VALE is an equally fast in-kernel software switch using the netmap API.
Both are implemented as a single kernel module for FreeBSD, Linux and
since summer 2015, also for Windows.
Netmap/VALE can handle tens of millions of packets per second, matching
the speed of 10G and 40G ports even with minimum sized frames.
See details at

	http://info.iet.unipi.it/~luigi/netmap/

This repository, hosted at https://github.com/luigirizzo/netmap , contains
source code (BSD-Copyright) for FreeBSD, Linux and Windows.
Note that recent FreeBSD distributions already include both NETMAP and VALE.


What is this good for
---------------------
Netmap is mostly useful for userspace applications that must deal with raw
packets: traffic generators, sinks, monitors, loggers, software switches
and routers, generic middleboxes, interconnection of virtual machines.

The example/ directory includes pkt-gen.c (a fast traffic generator/receiver)
and bridge.c, a simple bidirectional interconnect between two ports.
The kernel module itself implements a learning ethernet bridge.

More resources are hosted on other repositories. For example

  https://github.com/luigirizzo/netmap-libpcap

    contains a netmap-enabled version of libpcap (which is also
    included in FreeBSD distribution) so you can run any libpcap client
    on top of netmap at much higher speeds than using bpf.

  https://github.com/luigirizzo/netmap-ipfw

    is a userspace version of ipfw and dummynet which can handle several
    million packets per second in a single thread

Qemu/kvm has native netmap support, so it can interconnect VMs at high speed
through netmap ports. There is experimental netmap support in the FreeBSD's
bhyve hypervisor.


Netmap alone DOES NOT accelerate your TCP. For that you need to implement
your own tcp/ip stack probably using some of the techniques indicated
below to reduce the processing costs.

Architecture
------------
netmap uses a number of techniques to establish a fast and efficient path
between applications and the network. In order of importance:

	1. I/O batching
	2. efficient device drivers
	3. pre-allocated tx/rx buffers
	4. memory mapped buffers

Despite the name, memory mapping is NOT the key feature for netmap's
speed; systems that do not apply all these techniques do not achieve
the same speed _and_ efficiency.

Netmap clients use a select()-able file descriptor to synchronize
with the network card/software switch, and exchange multiple packets
per system call through device-independent memory mapped buffers and
descriptors. Device drivers are completely in the kernel, and the system
does not rely on IOMMU or other special mechanisms.


Installation instructions 
-------------------------
A single kernel module implements the core NETMAP functions, including
the VALE switch and access to physical NICS using unmodified device drivers
(at the price of much lower performance than netmap-aware drivers).

Netmap-aware device drivers are needed to use netmap at high speed
on ethernet ports.  To date, we have support for Intel ixgbe (10G),
ixl (10/40G), e1000/e1000e/igb (1G), Realtek 8169 (1G) and Nvidia (1G).
FreeBSD has also native netmap support in the Chelsio 10/40G cards.

  FreeBSD
  -------
  Since recent FreeBSD distributions already include netmap, you only
  need build the new kernel or modules as below:

  + add 'device netmap' to your kernel config file and rebuild a kernel.
    This will include the netmap module and netmap support in the device
    drivers.  Alternatively, you can build standalone modules
    (netmap, ixgbe, em, lem, re, igb)
  + sample applications are in the examples/ directory in this archive,
    or in src/tools/tools/netmap/ in FreeBSD distributions

  Linux
  -----
  On Linux, netmap is an out-of-tree module, so you need to compile it
  from these sources. The Makefile in the LINUX/ directory will also
  let you patch device driver sources and build some netmap-enabled
  device drivers.
  + make sure you have kernel sources matching your installed kernel
    (headers only suffice, if you want NETMAP/VALE but no drivers)

  + build kernel modules and sample applications:
    If kernel sources are in /foo//linux-A.B.C/ , then you should do

	cd netmap/LINUX
	# build kernel modules
	make NODRIVERS=1 KSRC=/foo/linux-A.B.C/	# only netmap
	make KSRC=/a/b/c/linux-A.B.C/		# netmap+device drivers
	# build sample applications
	make KSRC=/a/b/c/linux-A.B.C/ apps	# builds sample applications

    You can omit KSRC if your kernel sources are in a standard place.

  + if you use distribution packages, source may not contain headers (e.g., on
    debian systems). Use

        make SRC=/a/b/c/linux-sources-A.B/ KSRC=/a/b/c/linux-headers-A.B/
  
   WINDOWS
   -------
   Netmap has been ported to Windows in summer 2015 by Alessio Faina as part of
   his Master thesis. Please look at WINDOWS/README.txt for details.

Applications
------------
The directory examples/ contains some programs that use the netmap API

    pkt-gen.c	a packet generator/receiver working at line rate at 10Gbit/s
    vale-cfg.c	utility to configure ports of a VALE switch
    bridge.c	a utility that bridges two interfaces or one interface
		with the host stack

For libpcap and other applications look at the extra/ directory.

Testing
-------
pkt-gen is a generic test program which can act as a sender or receiver.
It has a large number of options, but the simplest form is:

    pkt-gen -i ix0 -f rx	# receive and print stats
    pkt-gen -i ix0 -f tx -l 60	# send a stream of 60-byte packets

(replace ix0 with the name of the interface or VALE port).
This should be able to work at line rate (up to 14.88 Mpps on 10
Gbit/interfaces, even higher on VALE) but note the following

OPERATING SPEED
---------------
Netmap is able to send packets at very high rates, and for simple
packet transmission and reception, speed generally not limited by
the CPU but by other factors (link speed, bus or NIC hw limitations).

For a physical link, the maximum numer of packets per second can
be computed with the formula:

	pps = line_rate / (672 + 8 * pkt_size)

where "line_rate" is the nominal link rate (e.g 10 Gbit/s) and
pkt_size is the actual packet size including MAC headers and CRC.
The following table summarizes some results

			LINE RATE
    pkt_size \	100M	1G	10G	40G

	  64	.1488	1.488	14.88	59.52
	 128	.0589	0.589	 5.89	23.58
	 256	.0367	0.367	 3.67	14.70
	 512	.0209	0.209	 2.09	 8.38
	1024	.0113	0.113	 1.13	 4.51
	1518	.0078	0.078	 0.78	 3.12

On VALE ports, there is no physical link and the throughput is
limited by CPU or memory depending on the packet size.

COMMON PROBLEMS
---------------
Before reporting slow send or receive speed on a physical interface,
check ALL of the following:

CANNOT SET THE DEVICE IN NETMAP MODE:
  + make sure that the netmap module and drivers are correctly
    loaded and can allocate all the memory they need (check into
    /var/log/messages or equivalent)
  + check permissions on /dev/netmap
  + make sure the interface is up before invoking pkt-gen

SENDER DOES NOT TRANSMIT
  + some switches/interfaces take a long time to (re)negotiate
    the link after starting pkt-gen; in case, use the -w N option
    to increase the initial delay to N seconds;

    	This may cause inability to transmit, or lost packets for
	the first few seconds of transmission

RECEIVER DOES NOT RECEIVE
  + make sure traffic uses a broadcast MAC addresses, or the UNICAST
    address of the receiving interface, or the receiving interface is in
    promiscuous mode (this must be done with ifconfig; pkt-gen does not
    change the operating mode)

LOWER SPEED THAN LINE RATE
  + check that your CPUs are running at the maximum clock rate
    and are not throttled down by the governor/powerd.

  + make sure that the sender/receiver interfaces and switch have
    flow control (FC) disabled (either via sysctl or ethtool).

        If FC is enabled and the receiving end is unable to cope
	with the traffic, the driver will try to slow down transmission,
	sometimes to very low rates.

  + a lot of hardware is not able to sustain line rate. For instance,
    ixgbe has problems with receiving frames that are not multiple
    of 64 bytes (with/without CRC depending on the driver); also on
    transmissions, ixgbe tops at about 12.5 Mpps unless the driver
    prefetches tx descriptors. igb does line rate in all configurations.
    e1000/e1000e vary between 1.15 and 1.32 Mpps. re/r8169 is
    extremely slow in sending (max 4-500 Kpps)


Credits
-------
NETMAP and VALE are projects of the Universita` di Pisa,
partially supported by various entities including:
Intel Research Berkeley, EU FP7 projects CHANGE and OPENLAB,
Netapp/Silicon Valley Community Foundation, ICSI, 

Author:		Luigi Rizzo
Contributors:
		Giuseppe Lettieri
		Michio Honda
		Marta Carbone
		Gaetano Catalli
		Matteo Landi
		Vincenzo Maffione
		Stefano Garzarella
		Alessio Faina

References
----------
There are a few academic papers describing netmap, VALE and applications.
You can find the papers at http://info.iet.unipi.it/~luigi/research.html

+ Luigi Rizzo,
	netmap: a novel framework for fast packet I/O,
	Usenix ATC'12, Boston, June 2012

+ Luigi Rizzo,
	Revisiting network I/O APIs: the netmap framework,
	Communications of the ACM 55 (3), 45-51, March 2012

+ Luigi Rizzo, Marta Carbone, Gaetano Catalli,
	Transparent acceleration of software packet forwarding using netmap,
	IEEE Infocom 2012, Orlando, March 2012

+ Luigi Rizzo, Giuseppe Lettieri,
	VALE: a switched ethernet for virtual machines,
	ACM Conext 2012, Nice, Dec. 2012

+ Luigi Rizzo, Giuseppe Lettieri, Vincenzo Maffione,
	Speeding up packet I/O in virtual machines,
	IEEE/ACM ANCS 2013, San Jose, Oct. 2013

+ Stefano Garzarella, Giuseppe Lettieri, Luigi Rizzo,
	Virtual device passthrough for high speed VM networking
	IEEE/ACM ANCS 2015, Oakland, May 2015