netdev-dpdk: Add support for vHost dequeue zero copy (experimental)

[ovs.git] / Documentation / intro / install / dpdk.rst

..
      Licensed under the Apache License, Version 2.0 (the "License"); you may
      not use this file except in compliance with the License. You may obtain
      a copy of the License at

          http://www.apache.org/licenses/LICENSE-2.0

      Unless required by applicable law or agreed to in writing, software
      distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
      WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
      License for the specific language governing permissions and limitations
      under the License.

      Convention for heading levels in Open vSwitch documentation:

      =======  Heading 0 (reserved for the title in a document)
      -------  Heading 1
      ~~~~~~~  Heading 2
      +++++++  Heading 3
      '''''''  Heading 4

      Avoid deeper levels because they do not render well.

======================
Open vSwitch with DPDK
======================

This document describes how to build and install Open vSwitch using a DPDK
datapath. Open vSwitch can use the DPDK library to operate entirely in
userspace.

.. seealso::

    The :doc:`releases FAQ </faq/releases>` lists support for the required
    versions of DPDK for each version of Open vSwitch.

Build requirements
------------------

In addition to the requirements described in :doc:`general`, building Open
vSwitch with DPDK will require the following:

- DPDK 17.11

- A `DPDK supported NIC`_

  Only required when physical ports are in use

- A suitable kernel

  On Linux Distros running kernel version >= 3.0, only `IOMMU` needs to enabled
  via the grub cmdline, assuming you are using **VFIO**. For older kernels,
  ensure the kernel is built with ``UIO``, ``HUGETLBFS``,
  ``PROC_PAGE_MONITOR``, ``HPET``, ``HPET_MMAP`` support. If these are not
  present, it will be necessary to upgrade your kernel or build a custom kernel
  with these flags enabled.

Detailed system requirements can be found at `DPDK requirements`_.

.. _DPDK supported NIC: http://dpdk.org/doc/nics
.. _DPDK requirements: http://dpdk.org/doc/guides/linux_gsg/sys_reqs.html

Installing
----------

Install DPDK
~~~~~~~~~~~~

#. Download the `DPDK sources`_, extract the file and set ``DPDK_DIR``::

       $ cd /usr/src/
       $ wget http://fast.dpdk.org/rel/dpdk-17.11.tar.xz
       $ tar xf dpdk-17.11.tar.xz
       $ export DPDK_DIR=/usr/src/dpdk-17.11
       $ cd $DPDK_DIR

#. (Optional) Configure DPDK as a shared library

   DPDK can be built as either a static library or a shared library.  By
   default, it is configured for the former. If you wish to use the latter, set
   ``CONFIG_RTE_BUILD_SHARED_LIB=y`` in ``$DPDK_DIR/config/common_base``.

   .. note::

      Minor performance loss is expected when using OVS with a shared DPDK
      library compared to a static DPDK library.

#. Configure and install DPDK

   Build and install the DPDK library::

       $ export DPDK_TARGET=x86_64-native-linuxapp-gcc
       $ export DPDK_BUILD=$DPDK_DIR/$DPDK_TARGET
       $ make install T=$DPDK_TARGET DESTDIR=install

#. (Optional) Export the DPDK shared library location

   If DPDK was built as a shared library, export the path to this library for
   use when building OVS::

       $ export LD_LIBRARY_PATH=$DPDK_DIR/x86_64-native-linuxapp-gcc/lib

.. _DPDK sources: http://dpdk.org/rel

Install OVS
~~~~~~~~~~~

OVS can be installed using different methods. For OVS to use DPDK datapath, it
has to be configured with DPDK support (``--with-dpdk``).

.. note::
  This section focuses on generic recipe that suits most cases. For
  distribution specific instructions, refer to one of the more relevant guides.

.. _OVS sources: http://openvswitch.org/releases/

#. Ensure the standard OVS requirements, described in
   :ref:`general-build-reqs`, are installed

#. Bootstrap, if required, as described in :ref:`general-bootstrapping`

#. Configure the package using the ``--with-dpdk`` flag::

       $ ./configure --with-dpdk=$DPDK_BUILD

   where ``DPDK_BUILD`` is the path to the built DPDK library. This can be
   skipped if DPDK library is installed in its default location.

   If no path is provided to ``--with-dpdk``, but a pkg-config configuration
   for libdpdk is available the include paths will be generated via an
   equivalent ``pkg-config --cflags libdpdk``.

   .. note::
     While ``--with-dpdk`` is required, you can pass any other configuration
     option described in :ref:`general-configuring`.

#. Build and install OVS, as described in :ref:`general-building`

Additional information can be found in :doc:`general`.

.. note::
  If you are running using the Fedora or Red Hat package, the Open vSwitch
  daemon will run as a non-root user.  This implies that you must have a
  working IOMMU.  Visit the `RHEL README`__ for additional information.

__ https://github.com/openvswitch/ovs/blob/master/rhel/README.RHEL.rst

Setup
-----

Setup Hugepages
~~~~~~~~~~~~~~~

Allocate a number of 2M Huge pages:

-  For persistent allocation of huge pages, write to hugepages.conf file
   in `/etc/sysctl.d`::

       $ echo 'vm.nr_hugepages=2048' > /etc/sysctl.d/hugepages.conf

-  For run-time allocation of huge pages, use the ``sysctl`` utility::

       $ sysctl -w vm.nr_hugepages=N  # where N = No. of 2M huge pages

To verify hugepage configuration::

    $ grep HugePages_ /proc/meminfo

Mount the hugepages, if not already mounted by default::

    $ mount -t hugetlbfs none /dev/hugepages``

.. _dpdk-vfio:

Setup DPDK devices using VFIO
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

VFIO is prefered to the UIO driver when using recent versions of DPDK. VFIO
support required support from both the kernel and BIOS. For the former, kernel
version > 3.6 must be used. For the latter, you must enable VT-d in the BIOS
and ensure this is configured via grub. To ensure VT-d is enabled via the BIOS,
run::

    $ dmesg | grep -e DMAR -e IOMMU

If VT-d is not enabled in the BIOS, enable it now.

To ensure VT-d is enabled in the kernel, run::

    $ cat /proc/cmdline | grep iommu=pt
    $ cat /proc/cmdline | grep intel_iommu=on

If VT-d is not enabled in the kernel, enable it now.

Once VT-d is correctly configured, load the required modules and bind the NIC
to the VFIO driver::

    $ modprobe vfio-pci
    $ /usr/bin/chmod a+x /dev/vfio
    $ /usr/bin/chmod 0666 /dev/vfio/*
    $ $DPDK_DIR/usertools/dpdk-devbind.py --bind=vfio-pci eth1
    $ $DPDK_DIR/usertools/dpdk-devbind.py --status

Setup OVS
~~~~~~~~~

Open vSwitch should be started as described in :doc:`general` with the
exception of ovs-vswitchd, which requires some special configuration to enable
DPDK functionality. DPDK configuration arguments can be passed to ovs-vswitchd
via the ``other_config`` column of the ``Open_vSwitch`` table. At a minimum,
the ``dpdk-init`` option must be set to ``true``. For example::

    $ export PATH=$PATH:/usr/local/share/openvswitch/scripts
    $ export DB_SOCK=/usr/local/var/run/openvswitch/db.sock
    $ ovs-vsctl --no-wait set Open_vSwitch . other_config:dpdk-init=true
    $ ovs-ctl --no-ovsdb-server --db-sock="$DB_SOCK" start

There are many other configuration options, the most important of which are
listed below. Defaults will be provided for all values not explicitly set.

``dpdk-init``
  Specifies whether OVS should initialize and support DPDK ports. This is a
  boolean, and defaults to false.

``dpdk-lcore-mask``
  Specifies the CPU cores on which dpdk lcore threads should be spawned and
  expects hex string (eg '0x123').

``dpdk-socket-mem``
  Comma separated list of memory to pre-allocate from hugepages on specific
  sockets.

``dpdk-hugepage-dir``
  Directory where hugetlbfs is mounted

``vhost-sock-dir``
  Option to set the path to the vhost-user unix socket files.

If allocating more than one GB hugepage, you can configure the
amount of memory used from any given NUMA nodes. For example, to use 1GB from
NUMA node 0 and 0GB for all other NUMA nodes, run::

    $ ovs-vsctl --no-wait set Open_vSwitch . \
        other_config:dpdk-socket-mem="1024,0"

or::

    $ ovs-vsctl --no-wait set Open_vSwitch . \
        other_config:dpdk-socket-mem="1024"

.. note::
  Changing any of these options requires restarting the ovs-vswitchd
  application

See the section ``Performance Tuning`` for important DPDK customizations.

Validating
----------

At this point you can use ovs-vsctl to set up bridges and other Open vSwitch
features. Seeing as we've configured the DPDK datapath, we will use DPDK-type
ports. For example, to create a userspace bridge named ``br0`` and add two
``dpdk`` ports to it, run::

    $ ovs-vsctl add-br br0 -- set bridge br0 datapath_type=netdev
    $ ovs-vsctl add-port br0 myportnameone -- set Interface myportnameone \
        type=dpdk options:dpdk-devargs=0000:06:00.0
    $ ovs-vsctl add-port br0 myportnametwo -- set Interface myportnametwo \
        type=dpdk options:dpdk-devargs=0000:06:00.1

DPDK devices will not be available for use until a valid dpdk-devargs is
specified.

Refer to ovs-vsctl(8) and :doc:`/howto/dpdk` for more details.

Performance Tuning
------------------

To achieve optimal OVS performance, the system can be configured and that
includes BIOS tweaks, Grub cmdline additions, better understanding of NUMA
nodes and apt selection of PCIe slots for NIC placement.

.. note::

   This section is optional. Once installed as described above, OVS with DPDK
   will work out of the box.

Recommended BIOS Settings
~~~~~~~~~~~~~~~~~~~~~~~~~

.. list-table:: Recommended BIOS Settings
   :header-rows: 1

   * - Setting
     - Value
   * - C3 Power State
     - Disabled
   * - C6 Power State
     - Disabled
   * - MLC Streamer
     - Enabled
   * - MLC Spacial Prefetcher
     - Enabled
   * - DCU Data Prefetcher
     - Enabled
   * - DCA
     - Enabled
   * - CPU Power and Performance
     - Performance
   * - Memeory RAS and Performance Config -> NUMA optimized
     - Enabled

PCIe Slot Selection
~~~~~~~~~~~~~~~~~~~

The fastpath performance can be affected by factors related to the placement of
the NIC, such as channel speeds between PCIe slot and CPU or the proximity of
PCIe slot to the CPU cores running the DPDK application. Listed below are the
steps to identify right PCIe slot.

#. Retrieve host details using ``dmidecode``. For example::

       $ dmidecode -t baseboard | grep "Product Name"

#. Download the technical specification for product listed, e.g: S2600WT2

#. Check the Product Architecture Overview on the Riser slot placement, CPU
   sharing info and also PCIe channel speeds

   For example: On S2600WT, CPU1 and CPU2 share Riser Slot 1 with Channel speed
   between CPU1 and Riser Slot1 at 32GB/s, CPU2 and Riser Slot1 at 16GB/s.
   Running DPDK app on CPU1 cores and NIC inserted in to Riser card Slots will
   optimize OVS performance in this case.

#. Check the Riser Card #1 - Root Port mapping information, on the available
   slots and individual bus speeds. In S2600WT slot 1, slot 2 has high bus
   speeds and are potential slots for NIC placement.

Advanced Hugepage Setup
~~~~~~~~~~~~~~~~~~~~~~~

Allocate and mount 1 GB hugepages.

- For persistent allocation of huge pages, add the following options to the
  kernel bootline::

      default_hugepagesz=1GB hugepagesz=1G hugepages=N

  For platforms supporting multiple huge page sizes, add multiple options::

      default_hugepagesz=<size> hugepagesz=<size> hugepages=N

  where:

  ``N``
    number of huge pages requested
  ``size``
    huge page size with an optional suffix ``[kKmMgG]``

- For run-time allocation of huge pages::

      $ echo N > /sys/devices/system/node/nodeX/hugepages/hugepages-1048576kB/nr_hugepages

  where:

  ``N``
    number of huge pages requested
  ``X``
    NUMA Node

  .. note::
    For run-time allocation of 1G huge pages, Contiguous Memory Allocator
    (``CONFIG_CMA``) has to be supported by kernel, check your Linux distro.

Now mount the huge pages, if not already done so::

    $ mount -t hugetlbfs -o pagesize=1G none /dev/hugepages

Isolate Cores
~~~~~~~~~~~~~

The ``isolcpus`` option can be used to isolate cores from the Linux scheduler.
The isolated cores can then be used to dedicatedly run HPC applications or
threads.  This helps in better application performance due to zero context
switching and minimal cache thrashing. To run platform logic on core 0 and
isolate cores between 1 and 19 from scheduler, add  ``isolcpus=1-19`` to GRUB
cmdline.

.. note::
  It has been verified that core isolation has minimal advantage due to mature
  Linux scheduler in some circumstances.

Compiler Optimizations
~~~~~~~~~~~~~~~~~~~~~~

The default compiler optimization level is ``-O2``. Changing this to more
aggressive compiler optimization such as ``-O3 -march=native`` with
gcc (verified on 5.3.1) can produce performance gains though not siginificant.
``-march=native`` will produce optimized code on local machine and should be
used when software compilation is done on Testbed.

Multiple Poll-Mode Driver Threads
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

With pmd multi-threading support, OVS creates one pmd thread for each NUMA node
by default, if there is at least one DPDK interface from that NUMA node added
to OVS. However, in cases where there are multiple ports/rxq's producing
traffic, performance can be improved by creating multiple pmd threads running
on separate cores. These pmd threads can share the workload by each being
responsible for different ports/rxq's. Assignment of ports/rxq's to pmd threads
is done automatically.

A set bit in the mask means a pmd thread is created and pinned to the
corresponding CPU core. For example, to run pmd threads on core 1 and 2::

    $ ovs-vsctl set Open_vSwitch . other_config:pmd-cpu-mask=0x6

When using dpdk and dpdkvhostuser ports in a bi-directional VM loopback as
shown below, spreading the workload over 2 or 4 pmd threads shows significant
improvements as there will be more total CPU occupancy available::

    NIC port0 <-> OVS <-> VM <-> OVS <-> NIC port 1

Refer to ovs-vswitchd.conf.db(5) for additional information on configuration
options.

Affinity
~~~~~~~~

For superior performance, DPDK pmd threads and Qemu vCPU threads needs to be
affinitized accordingly.

- PMD thread Affinity

  A poll mode driver (pmd) thread handles the I/O of all DPDK interfaces
  assigned to it. A pmd thread shall poll the ports for incoming packets,
  switch the packets and send to tx port.  A pmd thread is CPU bound, and needs
  to be affinitized to isolated cores for optimum performance.  Even though a
  PMD thread may exist, the thread only starts consuming CPU cycles if there is
  at least one receive queue assigned to the pmd.

  .. note::
    On NUMA systems, PCI devices are also local to a NUMA node.  Unbound rx
    queues for a PCI device will be assigned to a pmd on it's local NUMA node
    if a non-isolated PMD exists on that NUMA node.  If not, the queue will be
    assigned to a non-isolated pmd on a remote NUMA node.  This will result in
    reduced maximum throughput on that device and possibly on other devices
    assigned to that pmd thread. If such a queue assignment is made a warning
    message will be logged: "There's no available (non-isolated) pmd thread on
    numa node N. Queue Q on port P will be assigned to the pmd on core C
    (numa node N'). Expect reduced performance."

  Binding PMD threads to cores is described in the above section
  ``Multiple Poll-Mode Driver Threads``.

- QEMU vCPU thread Affinity

  A VM performing simple packet forwarding or running complex packet pipelines
  has to ensure that the vCPU threads performing the work has as much CPU
  occupancy as possible.

  For example, on a multicore VM, multiple QEMU vCPU threads shall be spawned.
  When the DPDK ``testpmd`` application that does packet forwarding is invoked,
  the ``taskset`` command should be used to affinitize the vCPU threads to the
  dedicated isolated cores on the host system.

Enable HyperThreading
~~~~~~~~~~~~~~~~~~~~~

With HyperThreading, or SMT, enabled, a physical core appears as two logical
cores. SMT can be utilized to spawn worker threads on logical cores of the same
physical core there by saving additional cores.

With DPDK, when pinning pmd threads to logical cores, care must be taken to set
the correct bits of the ``pmd-cpu-mask`` to ensure that the pmd threads are
pinned to SMT siblings.

Take a sample system configuration, with 2 sockets, 2 * 10 core processors, HT
enabled. This gives us a total of 40 logical cores. To identify the physical
core shared by two logical cores, run::

    $ cat /sys/devices/system/cpu/cpuN/topology/thread_siblings_list

where ``N`` is the logical core number.

In this example, it would show that cores ``1`` and ``21`` share the same
physical core. Logical cores can be specified in pmd-cpu-masks similarly to
physical cores, as described in ``Multiple Poll-Mode Driver Threads``.

NUMA/Cluster-on-Die
~~~~~~~~~~~~~~~~~~~

Ideally inter-NUMA datapaths should be avoided where possible as packets will
go across QPI and there may be a slight performance penalty when compared with
intra NUMA datapaths. On Intel Xeon Processor E5 v3, Cluster On Die is
introduced on models that have 10 cores or more.  This makes it possible to
logically split a socket into two NUMA regions and again it is preferred where
possible to keep critical datapaths within the one cluster.

It is good practice to ensure that threads that are in the datapath are pinned
to cores in the same NUMA area. e.g. pmd threads and QEMU vCPUs responsible for
forwarding. If DPDK is built with ``CONFIG_RTE_LIBRTE_VHOST_NUMA=y``, vHost
User ports automatically detect the NUMA socket of the QEMU vCPUs and will be
serviced by a PMD from the same node provided a core on this node is enabled in
the ``pmd-cpu-mask``. ``libnuma`` packages are required for this feature.

Binding PMD threads is described in the above section
``Multiple Poll-Mode Driver Threads``.

DPDK Physical Port Rx Queues
~~~~~~~~~~~~~~~~~~~~~~~~~~~~

::

    $ ovs-vsctl set Interface <DPDK interface> options:n_rxq=<integer>

The above command sets the number of rx queues for DPDK physical interface.
The rx queues are assigned to pmd threads on the same NUMA node in a
round-robin fashion.

.. _dpdk-queues-sizes:

DPDK Physical Port Queue Sizes
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

::

    $ ovs-vsctl set Interface dpdk0 options:n_rxq_desc=<integer>
    $ ovs-vsctl set Interface dpdk0 options:n_txq_desc=<integer>

The above command sets the number of rx/tx descriptors that the NIC associated
with dpdk0 will be initialised with.

Different ``n_rxq_desc`` and ``n_txq_desc`` configurations yield different
benefits in terms of throughput and latency for different scenarios.
Generally, smaller queue sizes can have a positive impact for latency at the
expense of throughput. The opposite is often true for larger queue sizes.
Note: increasing the number of rx descriptors eg. to 4096  may have a negative
impact on performance due to the fact that non-vectorised DPDK rx functions may
be used. This is dependent on the driver in use, but is true for the commonly
used i40e and ixgbe DPDK drivers.

Exact Match Cache
~~~~~~~~~~~~~~~~~

Each pmd thread contains one Exact Match Cache (EMC). After initial flow setup
in the datapath, the EMC contains a single table and provides the lowest level
(fastest) switching for DPDK ports. If there is a miss in the EMC then the next
level where switching will occur is the datapath classifier.  Missing in the
EMC and looking up in the datapath classifier incurs a significant performance
penalty.  If lookup misses occur in the EMC because it is too small to handle
the number of flows, its size can be increased. The EMC size can be modified by
editing the define ``EM_FLOW_HASH_SHIFT`` in ``lib/dpif-netdev.c``.

As mentioned above, an EMC is per pmd thread. An alternative way of increasing
the aggregate amount of possible flow entries in EMC and avoiding datapath
classifier lookups is to have multiple pmd threads running.

Rx Mergeable Buffers
~~~~~~~~~~~~~~~~~~~~

Rx mergeable buffers is a virtio feature that allows chaining of multiple
virtio descriptors to handle large packet sizes. Large packets are handled by
reserving and chaining multiple free descriptors together. Mergeable buffer
support is negotiated between the virtio driver and virtio device and is
supported by the DPDK vhost library.  This behavior is supported and enabled by
default, however in the case where the user knows that rx mergeable buffers are
not needed i.e. jumbo frames are not needed, it can be forced off by adding
``mrg_rxbuf=off`` to the QEMU command line options. By not reserving multiple
chains of descriptors it will make more individual virtio descriptors available
for rx to the guest using dpdkvhost ports and this can improve performance.

Output Packet Batching
~~~~~~~~~~~~~~~~~~~~~~

To make advantage of batched transmit functions, OVS collects packets in
intermediate queues before sending when processing a batch of received packets.
Even if packets are matched by different flows, OVS uses a single send
operation for all packets destined to the same output port.

Furthermore, OVS is able to buffer packets in these intermediate queues for a
configurable amount of time to reduce the frequency of send bursts at medium
load levels when the packet receive rate is high, but the receive batch size
still very small. This is particularly beneficial for packets transmitted to
VMs using an interrupt-driven virtio driver, where the interrupt overhead is
significant for the OVS PMD, the host operating system and the guest driver.

The ``tx-flush-interval`` parameter can be used to specify the time in
microseconds OVS should wait between two send bursts to a given port (default
is ``0``). When the intermediate queue fills up before that time is over, the
buffered packet batch is sent immediately::

    $ ovs-vsctl set Open_vSwitch . other_config:tx-flush-interval=50

This parameter influences both throughput and latency, depending on the traffic
load on the port. In general lower values decrease latency while higher values
may be useful to achieve higher throughput.

Low traffic (``packet rate < 1 / tx-flush-interval``) should not experience
any significant latency or throughput increase as packets are forwarded
immediately.

At intermediate load levels
(``1 / tx-flush-interval < packet rate < 32 / tx-flush-interval``) traffic
should experience an average latency increase of up to
``1 / 2 * tx-flush-interval`` and a possible throughput improvement.

Very high traffic (``packet rate >> 32 / tx-flush-interval``) should experience
the average latency increase equal to ``32 / (2 * packet rate)``. Most send
batches in this case will contain the maximum number of packets (``32``).

A ``tx-burst-interval`` value of ``50`` microseconds has shown to provide a
good performance increase in a ``PHY-VM-PHY`` scenario on ``x86`` system for
interrupt-driven guests while keeping the latency increase at a reasonable
level:

  https://mail.openvswitch.org/pipermail/ovs-dev/2017-December/341628.html

.. note::
  Throughput impact of this option significantly depends on the scenario and
  the traffic patterns. For example: ``tx-burst-interval`` value of ``50``
  microseconds shows performance degradation in ``PHY-VM-PHY`` with bonded PHY
  scenario while testing with ``256 - 1024`` packet flows:

    https://mail.openvswitch.org/pipermail/ovs-dev/2017-December/341700.html

The average number of packets per output batch can be checked in PMD stats::

    $ ovs-appctl dpif-netdev/pmd-stats-show

Limitations
------------

- Currently DPDK ports does not use HW offload functionality.
- Network Interface Firmware requirements: Each release of DPDK is
  validated against a specific firmware version for a supported Network
  Interface. New firmware versions introduce bug fixes, performance
  improvements and new functionality that DPDK leverages. The validated
  firmware versions are available as part of the release notes for
  DPDK. It is recommended that users update Network Interface firmware
  to match what has been validated for the DPDK release.

  The latest list of validated firmware versions can be found in the `DPDK
  release notes`_.

.. _DPDK release notes: http://dpdk.org/doc/guides/rel_notes/release_17_11.html

- Upper bound MTU: DPDK device drivers differ in how the L2 frame for a
  given MTU value is calculated e.g. i40e driver includes 2 x vlan headers in
  MTU overhead, em driver includes 1 x vlan header, ixgbe driver does not
  include a vlan  header in overhead. Currently it is not possible for OVS
  DPDK to know what upper bound MTU value is supported for a given device.
  As such OVS DPDK must provision for the case where the L2 frame for a given
  MTU includes 2 x vlan headers. This reduces the upper bound MTU value for
  devices that do not include vlan headers in their L2 frames by 8 bytes e.g.
  ixgbe devices upper bound MTU is reduced from 9710 to 9702. This work
  around is temporary and is expected to be removed once a method is provided
  by DPDK to query the upper bound MTU value for a given device.

Reporting Bugs
--------------

Report problems to bugs@openvswitch.org.