1 .. This work is licensed under a Creative Commons Attribution 4.0 International License.
2 .. http://creativecommons.org/licenses/by/4.0
3 .. (c) OPNFV, Intel Corporation, AT&T and others.
5 vSwitchPerf test suites userguide
6 ---------------------------------
11 VSPERF requires a traffic generators to run tests, automated traffic gen
12 support in VSPERF includes:
14 - IXIA traffic generator (IxNetwork hardware) and a machine that runs the IXIA
16 - Spirent traffic generator (TestCenter hardware chassis or TestCenter virtual
17 in a VM) and a VM to run the Spirent Virtual Deployment Service image,
18 formerly known as "Spirent LabServer".
19 - Xena Network traffic generator (Xena hardware chassis) that houses the Xena
20 Traffic generator modules.
21 - Moongen software traffic generator. Requires a separate machine running
22 moongen to execute packet generation.
23 - T-Rex software traffic generator. Requires a separate machine running T-Rex
24 Server to execute packet generation.
26 If you want to use another traffic generator, please select the :ref:`trafficgen-dummy`
32 To see the supported Operating Systems, vSwitches and system requirements,
33 please follow the `installation instructions <vsperf-installation>`.
35 Traffic Generator Setup
36 ^^^^^^^^^^^^^^^^^^^^^^^
38 Follow the `Traffic generator instructions <trafficgen-installation>` to
39 install and configure a suitable traffic generator.
41 Cloning and building src dependencies
42 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
44 In order to run VSPERF, you will need to download DPDK and OVS. You can
45 do this manually and build them in a preferred location, OR you could
46 use vswitchperf/src. The vswitchperf/src directory contains makefiles
47 that will allow you to clone and build the libraries that VSPERF depends
48 on, such as DPDK and OVS. To clone and build simply:
50 .. code-block:: console
55 VSPERF can be used with stock OVS (without DPDK support). When build
56 is finished, the libraries are stored in src_vanilla directory.
58 The 'make' builds all options in src:
61 * OVS with vhost_user as the guest access method (with DPDK support)
63 The vhost_user build will reside in src/ovs/
64 The Vanilla OVS build will reside in vswitchperf/src_vanilla
66 To delete a src subdirectory and its contents to allow you to re-clone simply
69 .. code-block:: console
73 Configure the ``./conf/10_custom.conf`` file
74 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
76 The ``10_custom.conf`` file is the configuration file that overrides
77 default configurations in all the other configuration files in ``./conf``
78 The supplied ``10_custom.conf`` file **MUST** be modified, as it contains
79 configuration items for which there are no reasonable default values.
81 The configuration items that can be added is not limited to the initial
82 contents. Any configuration item mentioned in any .conf file in
83 ``./conf`` directory can be added and that item will be overridden by
84 the custom configuration value.
86 Further details about configuration files evaluation and special behaviour
87 of options with ``GUEST_`` prefix could be found at :ref:`design document
88 <design-configuration>`.
90 Using a custom settings file
91 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
93 If your ``10_custom.conf`` doesn't reside in the ``./conf`` directory
94 or if you want to use an alternative configuration file, the file can
95 be passed to ``vsperf`` via the ``--conf-file`` argument.
97 .. code-block:: console
99 $ ./vsperf --conf-file <path_to_custom_conf> ...
101 Evaluation of configuration parameters
102 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
104 The value of configuration parameter can be specified at various places,
105 e.g. at the test case definition, inside configuration files, by the command
106 line argument, etc. Thus it is important to understand the order of configuration
107 parameter evaluation. This "priority hierarchy" can be described like so
110 1. Testcase definition section ``Parameters``
111 2. Command line arguments (e.g. ``--test-params``, ``--vswitch``, etc.)
112 3. Environment variables (see ``--load-env`` argument)
113 4. Custom configuration file specified via ``--conf-file`` argument
114 5. Standard configuration files, where higher prefix number means higher
117 For example, if the same configuration parameter is defined in custom configuration
118 file (specified via ``--conf-file`` argument), via ``--test-params`` argument
119 and also inside ``Parameters`` section of the testcase definition, then parameter
120 value from the ``Parameters`` section will be used.
122 Further details about order of configuration files evaluation and special behaviour
123 of options with ``GUEST_`` prefix could be found at :ref:`design document
124 <design-configuration>`.
126 .. _overriding-parameters-documentation:
128 Overriding values defined in configuration files
129 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
131 The configuration items can be overridden by command line argument
132 ``--test-params``. In this case, the configuration items and
133 their values should be passed in form of ``item=value`` and separated
140 $ ./vsperf --test-params "TRAFFICGEN_DURATION=10;TRAFFICGEN_PKT_SIZES=(128,);" \
141 "GUEST_LOOPBACK=['testpmd','l2fwd']" pvvp_tput
143 The ``--test-params`` command line argument can also be used to override default
144 configuration values for multiple tests. Providing a list of parameters will apply each
145 element of the list to the test with the same index. If more tests are run than
146 parameters provided the last element of the list will repeat.
150 $ ./vsperf --test-params "['TRAFFICGEN_DURATION=10;TRAFFICGEN_PKT_SIZES=(128,)',"
151 "'TRAFFICGEN_DURATION=10;TRAFFICGEN_PKT_SIZES=(64,)']" \
154 The second option is to override configuration items by ``Parameters`` section
155 of the test case definition. The configuration items can be added into ``Parameters``
156 dictionary with their new values. These values will override values defined in
157 configuration files or specified by ``--test-params`` command line argument.
163 "Parameters" : {'TRAFFICGEN_PKT_SIZES' : (128,),
164 'TRAFFICGEN_DURATION' : 10,
165 'GUEST_LOOPBACK' : ['testpmd','l2fwd'],
168 **NOTE:** In both cases, configuration item names and their values must be specified
169 in the same form as they are defined inside configuration files. Parameter names
170 must be specified in uppercase and data types of original and new value must match.
171 Python syntax rules related to data types and structures must be followed.
172 For example, parameter ``TRAFFICGEN_PKT_SIZES`` above is defined as a tuple
173 with a single value ``128``. In this case trailing comma is mandatory, otherwise
174 value can be wrongly interpreted as a number instead of a tuple and vsperf
175 execution would fail. Please check configuration files for default values and their
176 types and use them as a basis for any customized values. In case of any doubt, please
177 check official python documentation related to data structures like tuples, lists
180 **NOTE:** Vsperf execution will terminate with runtime error in case, that unknown
181 parameter name is passed via ``--test-params`` CLI argument or defined in ``Parameters``
182 section of test case definition. It is also forbidden to redefine a value of
183 ``TEST_PARAMS`` configuration item via CLI or ``Parameters`` section.
185 **NOTE:** The new definition of the dictionary parameter, specified via ``--test-params``
186 or inside ``Parameters`` section, will not override original dictionary values. Instead
187 the original dictionary will be updated with values from the new dictionary definition.
189 Referencing parameter values
190 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
192 It is possible to use a special macro ``#PARAM()`` to refer to the value of
193 another configuration parameter. This reference is evaluated during
194 access of the parameter value (by ``settings.getValue()`` call), so it
195 can refer to parameters created during VSPERF runtime, e.g. NICS dictionary.
196 It can be used to reflect DUT HW details in the testcase definition.
208 # set destination MAC to the MAC of the first
209 # interface from WHITELIST_NICS list
210 'dstmac' : '#PARAM(NICS[0]["mac"])',
218 VSPERF uses a VM image called vloop_vnf for looping traffic in the deployment
219 scenarios involving VMs. The image can be downloaded from
220 `<http://artifacts.opnfv.org/>`__.
222 Please see the installation instructions for information on :ref:`vloop-vnf`
230 A Kernel Module that provides OSI Layer 2 Ipv4 termination or forwarding with
231 support for Destination Network Address Translation (DNAT) for both the MAC and
232 IP addresses. l2fwd can be found in <vswitchperf_dir>/src/l2fwd
234 Additional Tools Setup
235 ^^^^^^^^^^^^^^^^^^^^^^
237 Follow the `Additional tools instructions <additional-tools-configuration>` to
238 install and configure additional tools such as collectors and loadgens.
243 All examples inside these docs assume, that user is inside the VSPERF
244 directory. VSPERF can be executed from any directory.
246 Before running any tests make sure you have root permissions by adding
247 the following line to /etc/sudoers:
249 .. code-block:: console
251 username ALL=(ALL) NOPASSWD: ALL
253 username in the example above should be replaced with a real username.
255 To list the available tests:
257 .. code-block:: console
261 To run a single test:
263 .. code-block:: console
267 Where $TESTNAME is the name of the vsperf test you would like to run.
269 To run a test multiple times, repeat it:
271 .. code-block:: console
273 $ ./vsperf $TESTNAME $TESTNAME $TESTNAME
275 To run a group of tests, for example all tests with a name containing
278 .. code-block:: console
280 $ ./vsperf --conf-file=<path_to_custom_conf>/10_custom.conf --tests="RFC2544"
284 .. code-block:: console
286 $ ./vsperf --conf-file=<path_to_custom_conf>/10_custom.conf
288 Some tests allow for configurable parameters, including test duration
289 (in seconds) as well as packet sizes (in bytes).
293 $ ./vsperf --conf-file user_settings.py \
294 --tests RFC2544Tput \
295 --test-params "TRAFFICGEN_DURATION=10;TRAFFICGEN_PKT_SIZES=(128,)"
297 To specify configurable parameters for multiple tests, use a list of
298 parameters. One element for each test.
302 $ ./vsperf --conf-file user_settings.py \
303 --test-params "['TRAFFICGEN_DURATION=10;TRAFFICGEN_PKT_SIZES=(128,)',"\
304 "'TRAFFICGEN_DURATION=10;TRAFFICGEN_PKT_SIZES=(64,)']" \
305 phy2phy_cont phy2phy_cont
307 If the ``CUMULATIVE_PARAMS`` setting is set to True and there are different parameters
308 provided for each test using ``--test-params``, each test will take the parameters of
309 the previous test before appyling it's own.
310 With ``CUMULATIVE_PARAMS`` set to True the following command will be equivalent to the
315 $ ./vsperf --conf-file user_settings.py \
316 --test-params "['TRAFFICGEN_DURATION=10;TRAFFICGEN_PKT_SIZES=(128,)',"\
317 "'TRAFFICGEN_PKT_SIZES=(64,)']" \
318 phy2phy_cont phy2phy_cont
321 For all available options, check out the help dialog:
323 .. code-block:: console
327 Executing Vanilla OVS tests
328 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
330 1. If needed, recompile src for all OVS variants
332 .. code-block:: console
338 2. Update your ``10_custom.conf`` file to use Vanilla OVS:
340 .. code-block:: python
342 VSWITCH = 'OvsVanilla'
346 .. code-block:: console
348 $ ./vsperf --conf-file=<path_to_custom_conf>
350 Please note if you don't want to configure Vanilla OVS through the
351 configuration file, you can pass it as a CLI argument.
353 .. code-block:: console
355 $ ./vsperf --vswitch OvsVanilla
358 Executing tests with VMs
359 ^^^^^^^^^^^^^^^^^^^^^^^^
361 To run tests using vhost-user as guest access method:
363 1. Set VSWITCH and VNF of your settings file to:
365 .. code-block:: python
367 VSWITCH = 'OvsDpdkVhost'
368 VNF = 'QemuDpdkVhost'
370 2. If needed, recompile src for all OVS variants
372 .. code-block:: console
380 .. code-block:: console
382 $ ./vsperf --conf-file=<path_to_custom_conf>/10_custom.conf
384 **NOTE:** By default vSwitch is acting as a server for dpdk vhost-user sockets.
385 In case, that QEMU should be a server for vhost-user sockets, then parameter
386 ``VSWITCH_VHOSTUSER_SERVER_MODE`` should be set to ``False``.
388 Executing tests with VMs using Vanilla OVS
389 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
391 To run tests using Vanilla OVS:
393 1. Set the following variables:
395 .. code-block:: python
397 VSWITCH = 'OvsVanilla'
398 VNF = 'QemuVirtioNet'
400 VANILLA_TGEN_PORT1_IP = n.n.n.n
401 VANILLA_TGEN_PORT1_MAC = nn:nn:nn:nn:nn:nn
403 VANILLA_TGEN_PORT2_IP = n.n.n.n
404 VANILLA_TGEN_PORT2_MAC = nn:nn:nn:nn:nn:nn
406 VANILLA_BRIDGE_IP = n.n.n.n
408 or use ``--test-params`` option
410 .. code-block:: console
412 $ ./vsperf --conf-file=<path_to_custom_conf>/10_custom.conf \
413 --test-params "VANILLA_TGEN_PORT1_IP=n.n.n.n;" \
414 "VANILLA_TGEN_PORT1_MAC=nn:nn:nn:nn:nn:nn;" \
415 "VANILLA_TGEN_PORT2_IP=n.n.n.n;" \
416 "VANILLA_TGEN_PORT2_MAC=nn:nn:nn:nn:nn:nn"
418 2. If needed, recompile src for all OVS variants
420 .. code-block:: console
428 .. code-block:: console
430 $ ./vsperf --conf-file<path_to_custom_conf>/10_custom.conf
437 Currently it is not possible to use standard scenario deployments for execution of
438 tests with VPP. It means, that deployments ``p2p``, ``pvp``, ``pvvp`` and in general any
439 :ref:`pxp-deployment` won't work with VPP. However it is possible to use VPP in
440 :ref:`step-driven-tests`. A basic set of VPP testcases covering ``phy2phy``, ``pvp``
441 and ``pvvp`` tests are already prepared.
443 List of performance tests with VPP support follows:
445 * phy2phy_tput_vpp: VPP: LTD.Throughput.RFC2544.PacketLossRatio
446 * phy2phy_cont_vpp: VPP: Phy2Phy Continuous Stream
447 * phy2phy_back2back_vpp: VPP: LTD.Throughput.RFC2544.BackToBackFrames
448 * pvp_tput_vpp: VPP: LTD.Throughput.RFC2544.PacketLossRatio
449 * pvp_cont_vpp: VPP: PVP Continuous Stream
450 * pvp_back2back_vpp: VPP: LTD.Throughput.RFC2544.BackToBackFrames
451 * pvvp_tput_vpp: VPP: LTD.Throughput.RFC2544.PacketLossRatio
452 * pvvp_cont_vpp: VPP: PVP Continuous Stream
453 * pvvp_back2back_vpp: VPP: LTD.Throughput.RFC2544.BackToBackFrames
455 In order to execute testcases with VPP it is required to:
457 * install VPP manually, see :ref:`vpp-installation`
458 * configure ``WHITELIST_NICS``, with two physical NICs connected to the traffic generator
459 * configure traffic generator, see :ref:`trafficgen-installation`
461 After that it is possible to execute VPP testcases listed above.
465 .. code-block:: console
467 $ ./vsperf --conf-file=<path_to_custom_conf> phy2phy_tput_vpp
471 Using vfio_pci with DPDK
472 ^^^^^^^^^^^^^^^^^^^^^^^^^
474 To use vfio with DPDK instead of igb_uio add into your custom configuration
475 file the following parameter:
477 .. code-block:: python
479 PATHS['dpdk']['src']['modules'] = ['uio', 'vfio-pci']
482 **NOTE:** In case, that DPDK is installed from binary package, then please
483 set ``PATHS['dpdk']['bin']['modules']`` instead.
485 **NOTE:** Please ensure that Intel VT-d is enabled in BIOS.
487 **NOTE:** Please ensure your boot/grub parameters include
490 **NOTE:** In case of VPP, it is required to explicitly define, that vfio-pci
491 DPDK driver should be used. It means to update dpdk part of VSWITCH_VPP_ARGS
492 dictionary with uio-driver section, e.g. VSWITCH_VPP_ARGS['dpdk'] = 'uio-driver vfio-pci'
494 .. code-block:: console
496 iommu=pt intel_iommu=on
498 To check that IOMMU is enabled on your platform:
500 .. code-block:: console
503 [ 0.000000] Intel-IOMMU: enabled
504 [ 0.139882] dmar: IOMMU 0: reg_base_addr fbffe000 ver 1:0 cap d2078c106f0466 ecap f020de
505 [ 0.139888] dmar: IOMMU 1: reg_base_addr ebffc000 ver 1:0 cap d2078c106f0466 ecap f020de
506 [ 0.139893] IOAPIC id 2 under DRHD base 0xfbffe000 IOMMU 0
507 [ 0.139894] IOAPIC id 0 under DRHD base 0xebffc000 IOMMU 1
508 [ 0.139895] IOAPIC id 1 under DRHD base 0xebffc000 IOMMU 1
509 [ 3.335744] IOMMU: dmar0 using Queued invalidation
510 [ 3.335746] IOMMU: dmar1 using Queued invalidation
518 To use virtual functions of NIC with SRIOV support, use extended form
519 of NIC PCI slot definition:
521 .. code-block:: python
523 WHITELIST_NICS = ['0000:05:00.0|vf0', '0000:05:00.1|vf3']
525 Where 'vf' is an indication of virtual function usage and following
526 number defines a VF to be used. In case that VF usage is detected,
527 then vswitchperf will enable SRIOV support for given card and it will
528 detect PCI slot numbers of selected VFs.
530 So in example above, one VF will be configured for NIC '0000:05:00.0'
531 and four VFs will be configured for NIC '0000:05:00.1'. Vswitchperf
532 will detect PCI addresses of selected VFs and it will use them during
535 At the end of vswitchperf execution, SRIOV support will be disabled.
537 SRIOV support is generic and it can be used in different testing scenarios.
540 * vSwitch tests with DPDK or without DPDK support to verify impact
541 of VF usage on vSwitch performance
542 * tests without vSwitch, where traffic is forwarded directly
543 between VF interfaces by packet forwarder (e.g. testpmd application)
544 * tests without vSwitch, where VM accesses VF interfaces directly
545 by PCI-passthrough_ to measure raw VM throughput performance.
549 Using QEMU with PCI passthrough support
550 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
552 Raw virtual machine throughput performance can be measured by execution of PVP
553 test with direct access to NICs by PCI pass-through. To execute VM with direct
554 access to PCI devices, enable vfio-pci_. In order to use virtual functions,
555 SRIOV-support_ must be enabled.
557 Execution of test with PCI pass-through with vswitch disabled:
559 .. code-block:: console
561 $ ./vsperf --conf-file=<path_to_custom_conf>/10_custom.conf \
562 --vswitch none --vnf QemuPciPassthrough pvp_tput
564 Any of supported guest-loopback-application_ can be used inside VM with
565 PCI pass-through support.
567 Note: Qemu with PCI pass-through support can be used only with PVP test
570 .. _guest-loopback-application:
572 Selection of loopback application for tests with VMs
573 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
575 To select the loopback applications which will forward packets inside VMs,
576 the following parameter should be configured:
578 .. code-block:: python
580 GUEST_LOOPBACK = ['testpmd']
582 or use ``--test-params`` CLI argument:
584 .. code-block:: console
586 $ ./vsperf --conf-file=<path_to_custom_conf>/10_custom.conf \
587 --test-params "GUEST_LOOPBACK=['testpmd']"
589 Supported loopback applications are:
591 .. code-block:: console
593 'testpmd' - testpmd from dpdk will be built and used
594 'l2fwd' - l2fwd module provided by Huawei will be built and used
595 'linux_bridge' - linux bridge will be configured
596 'buildin' - nothing will be configured by vsperf; VM image must
597 ensure traffic forwarding between its interfaces
599 Guest loopback application must be configured, otherwise traffic
600 will not be forwarded by VM and testcases with VM related deployments
601 will fail. Guest loopback application is set to 'testpmd' by default.
603 **NOTE:** In case that only 1 or more than 2 NICs are configured for VM,
604 then 'testpmd' should be used. As it is able to forward traffic between
605 multiple VM NIC pairs.
607 **NOTE:** In case of linux_bridge, all guest NICs are connected to the same
608 bridge inside the guest.
610 Mergable Buffers Options with QEMU
611 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
613 Mergable buffers can be disabled with VSPerf within QEMU. This option can
614 increase performance significantly when not using jumbo frame sized packets.
615 By default VSPerf disables mergable buffers. If you wish to enable it you
616 can modify the setting in the a custom conf file.
618 .. code-block:: python
620 GUEST_NIC_MERGE_BUFFERS_DISABLE = [False]
622 Then execute using the custom conf file.
624 .. code-block:: console
626 $ ./vsperf --conf-file=<path_to_custom_conf>/10_custom.conf
628 Alternatively you can just pass the param during execution.
630 .. code-block:: console
632 $ ./vsperf --test-params "GUEST_NIC_MERGE_BUFFERS_DISABLE=[False]"
635 Selection of dpdk binding driver for tests with VMs
636 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
638 To select dpdk binding driver, which will specify which driver the vm NICs will
639 use for dpdk bind, the following configuration parameter should be configured:
641 .. code-block:: console
643 GUEST_DPDK_BIND_DRIVER = ['igb_uio_from_src']
645 The supported dpdk guest bind drivers are:
647 .. code-block:: console
649 'uio_pci_generic' - Use uio_pci_generic driver
650 'igb_uio_from_src' - Build and use the igb_uio driver from the dpdk src
652 'vfio_no_iommu' - Use vfio with no iommu option. This requires custom
653 guest images that support this option. The default
654 vloop image does not support this driver.
656 Note: uio_pci_generic does not support sr-iov testcases with guests attached.
657 This is because uio_pci_generic only supports legacy interrupts. In case
658 uio_pci_generic is selected with the vnf as QemuPciPassthrough it will be
659 modified to use igb_uio_from_src instead.
661 Note: vfio_no_iommu requires kernels equal to or greater than 4.5 and dpdk
662 16.04 or greater. Using this option will also taint the kernel.
664 Please refer to the dpdk documents at http://dpdk.org/doc/guides for more
665 information on these drivers.
667 Guest Core and Thread Binding
668 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
670 VSPERF provides options to achieve better performance by guest core binding and
671 guest vCPU thread binding as well. Core binding is to bind all the qemu threads.
672 Thread binding is to bind the house keeping threads to some CPU and vCPU thread to
673 some other CPU, this helps to reduce the noise from qemu house keeping threads.
676 .. code-block:: python
678 GUEST_CORE_BINDING = [('#EVAL(6+2*#VMINDEX)', '#EVAL(7+2*#VMINDEX)')]
680 **NOTE** By default the GUEST_THREAD_BINDING will be none, which means same as
681 the GUEST_CORE_BINDING, i.e. the vcpu threads are sharing the physical CPUs with
682 the house keeping threads. Better performance using vCPU thread binding can be
683 achieved by enabling affinity in the custom configuration file.
685 For example, if an environment requires 32,33 to be core binded and 29,30&31 for
686 guest thread binding to achieve better performance.
688 .. code-block:: python
690 VNF_AFFINITIZATION_ON = True
691 GUEST_CORE_BINDING = [('32','33')]
692 GUEST_THREAD_BINDING = [('29', '30', '31')]
697 QEMU default to a compatible subset of performance enhancing cpu features.
698 To pass all available host processor features to the guest.
700 .. code-block:: python
702 GUEST_CPU_OPTIONS = ['host,migratable=off']
704 **NOTE** To enhance the performance, cpu features tsc deadline timer for guest,
705 the guest PMU, the invariant TSC can be provided in the custom configuration file.
707 Multi-Queue Configuration
708 ^^^^^^^^^^^^^^^^^^^^^^^^^
710 VSPerf currently supports multi-queue with the following limitations:
712 1. Requires QEMU 2.5 or greater and any OVS version higher than 2.5. The
713 default upstream package versions installed by VSPerf satisfies this
716 2. Guest image must have ethtool utility installed if using l2fwd or linux
717 bridge inside guest for loopback.
719 3. If using OVS versions 2.5.0 or less enable old style multi-queue as shown
720 in the ''02_vswitch.conf'' file.
722 .. code-block:: python
724 OVS_OLD_STYLE_MQ = True
726 To enable multi-queue for dpdk modify the ''02_vswitch.conf'' file.
728 .. code-block:: python
730 VSWITCH_DPDK_MULTI_QUEUES = 2
732 **NOTE:** you should consider using the switch affinity to set a pmd cpu mask
733 that can optimize your performance. Consider the numa of the NIC in use if this
734 applies by checking /sys/class/net/<eth_name>/device/numa_node and setting an
735 appropriate mask to create PMD threads on the same numa node.
737 When multi-queue is enabled, each dpdk or dpdkvhostuser port that is created
738 on the switch will set the option for multiple queues. If old style multi queue
739 has been enabled a global option for multi queue will be used instead of the
742 To enable multi-queue on the guest modify the ''04_vnf.conf'' file.
744 .. code-block:: python
746 GUEST_NIC_QUEUES = [2]
748 Enabling multi-queue at the guest will add multiple queues to each NIC port when
749 qemu launches the guest.
751 In case of Vanilla OVS, multi-queue is enabled on the tuntap ports and nic
752 queues will be enabled inside the guest with ethtool. Simply enabling the
753 multi-queue on the guest is sufficient for Vanilla OVS multi-queue.
755 Testpmd should be configured to take advantage of multi-queue on the guest if
756 using DPDKVhostUser. This can be done by modifying the ''04_vnf.conf'' file.
758 .. code-block:: python
760 GUEST_TESTPMD_PARAMS = ['-l 0,1,2,3,4 -n 4 --socket-mem 512 -- '
761 '--burst=64 -i --txqflags=0xf00 '
762 '--nb-cores=4 --rxq=2 --txq=2 '
765 **NOTE:** The guest SMP cores must be configured to allow for testpmd to use the
766 optimal number of cores to take advantage of the multiple guest queues.
768 In case of using Vanilla OVS and qemu virtio-net you can increase performance
769 by binding vhost-net threads to cpus. This can be done by enabling the affinity
770 in the ''04_vnf.conf'' file. This can be done to non multi-queue enabled
771 configurations as well as there will be 2 vhost-net threads.
773 .. code-block:: python
775 VSWITCH_VHOST_NET_AFFINITIZATION = True
777 VSWITCH_VHOST_CPU_MAP = [4,5,8,11]
779 **NOTE:** This method of binding would require a custom script in a real
782 **NOTE:** For optimal performance guest SMPs and/or vhost-net threads should be
783 on the same numa as the NIC in use if possible/applicable. Testpmd should be
784 assigned at least (nb_cores +1) total cores with the cpu mask.
789 VSPERF provides options to support jumbo frame testing with a jumbo frame supported
790 NIC and traffic generator for the following vswitches:
796 3. TestPMD loopback with or without a guest
798 **NOTE:** There is currently no support for SR-IOV or VPP at this time with jumbo
801 All packet forwarding applications for pxp testing is supported.
803 To enable jumbo frame testing simply enable the option in the conf files and set the
804 maximum size that will be used.
806 .. code-block:: python
808 VSWITCH_JUMBO_FRAMES_ENABLED = True
809 VSWITCH_JUMBO_FRAMES_SIZE = 9000
811 To enable jumbo frame testing with OVSVanilla the NIC in test on the host must have
812 its mtu size changed manually using ifconfig or applicable tools:
814 .. code-block:: console
816 ifconfig eth1 mtu 9000 up
818 **NOTE:** To make the setting consistent across reboots you should reference the OS
819 documents as it differs from distribution to distribution.
821 To start a test for jumbo frames modify the conf file packet sizes or pass the option
822 through the VSPERF command line.
824 .. code-block:: python
826 TEST_PARAMS = {'TRAFFICGEN_PKT_SIZES':(2000,9000)}
828 .. code-block:: python
830 ./vsperf --test-params "TRAFFICGEN_PKT_SIZES=2000,9000"
832 It is recommended to increase the memory size for OvsDpdkVhostUser testing from the default
833 1024. Your size required may vary depending on the number of guests in your testing. 4096
834 appears to work well for most typical testing scenarios.
836 .. code-block:: python
838 DPDK_SOCKET_MEM = ['4096', '0']
840 **NOTE:** For Jumbo frames to work with DpdkVhostUser, mergable buffers will be enabled by
841 default. If testing with mergable buffers in QEMU is desired, disable Jumbo Frames and only
842 test non jumbo frame sizes. Test Jumbo Frames sizes separately to avoid this collision.
845 Executing Packet Forwarding tests
846 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
848 To select the applications which will forward packets,
849 the following parameters should be configured:
851 .. code-block:: python
856 or use ``--vswitch`` and ``--fwdapp`` CLI arguments:
858 .. code-block:: console
860 $ ./vsperf phy2phy_cont --conf-file user_settings.py \
864 Supported Packet Forwarding applications are:
866 .. code-block:: console
868 'testpmd' - testpmd from dpdk
871 1. Update your ''10_custom.conf'' file to use the appropriate variables
872 for selected Packet Forwarder:
874 .. code-block:: python
876 # testpmd configuration
878 # packet forwarding mode supported by testpmd; Please see DPDK documentation
879 # for comprehensive list of modes supported by your version.
880 # e.g. io|mac|mac_retry|macswap|flowgen|rxonly|txonly|csum|icmpecho|...
881 # Note: Option "mac_retry" has been changed to "mac retry" since DPDK v16.07
882 TESTPMD_FWD_MODE = 'csum'
883 # checksum calculation layer: ip|udp|tcp|sctp|outer-ip
884 TESTPMD_CSUM_LAYER = 'ip'
885 # checksum calculation place: hw (hardware) | sw (software)
886 TESTPMD_CSUM_CALC = 'sw'
887 # recognize tunnel headers: on|off
888 TESTPMD_CSUM_PARSE_TUNNEL = 'off'
892 .. code-block:: console
894 $ ./vsperf phy2phy_tput --conf-file <path_to_settings_py>
896 Executing Packet Forwarding tests with one guest
897 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
899 TestPMD with DPDK 16.11 or greater can be used to forward packets as a switch to a single guest using TestPMD vdev
900 option. To set this configuration the following parameters should be used.
902 .. code-block:: python
907 or use ``--vswitch`` and ``--fwdapp`` CLI arguments:
909 .. code-block:: console
911 $ ./vsperf pvp_tput --conf-file user_settings.py \
915 Guest forwarding application only supports TestPMD in this configuration.
917 .. code-block:: python
919 GUEST_LOOPBACK = ['testpmd']
921 For optimal performance one cpu per port +1 should be used for TestPMD. Also set additional params for packet forwarding
922 application to use the correct number of nb-cores.
924 .. code-block:: python
926 DPDK_SOCKET_MEM = ['1024', '0']
927 VSWITCHD_DPDK_ARGS = ['-l', '46,44,42,40,38', '-n', '4']
928 TESTPMD_ARGS = ['--nb-cores=4', '--txq=1', '--rxq=1']
930 For guest TestPMD 3 VCpus should be assigned with the following TestPMD params.
932 .. code-block:: python
934 GUEST_TESTPMD_PARAMS = ['-l 0,1,2 -n 4 --socket-mem 1024 -- '
935 '--burst=64 -i --txqflags=0xf00 '
936 '--disable-hw-vlan --nb-cores=2 --txq=1 --rxq=1']
938 Execution of TestPMD can be run with the following command line
940 .. code-block:: console
942 ./vsperf pvp_tput --vswitch=none --fwdapp=TestPMD --conf-file <path_to_settings_py>
944 **NOTE:** To achieve the best 0% loss numbers with rfc2544 throughput testing, other tunings should be applied to host
945 and guest such as tuned profiles and CPU tunings to prevent possible interrupts to worker threads.
947 VSPERF modes of operation
948 ^^^^^^^^^^^^^^^^^^^^^^^^^
950 VSPERF can be run in different modes. By default it will configure vSwitch,
951 traffic generator and VNF. However it can be used just for configuration
952 and execution of traffic generator. Another option is execution of all
953 components except traffic generator itself.
955 Mode of operation is driven by configuration parameter -m or --mode
957 .. code-block:: console
959 -m MODE, --mode MODE vsperf mode of operation;
961 "normal" - execute vSwitch, VNF and traffic generator
962 "trafficgen" - execute only traffic generator
963 "trafficgen-off" - execute vSwitch and VNF
964 "trafficgen-pause" - execute vSwitch and VNF but wait before traffic transmission
966 In case, that VSPERF is executed in "trafficgen" mode, then configuration
967 of traffic generator can be modified through ``TRAFFIC`` dictionary passed to the
968 ``--test-params`` option. It is not needed to specify all values of ``TRAFFIC``
969 dictionary. It is sufficient to specify only values, which should be changed.
970 Detailed description of ``TRAFFIC`` dictionary can be found at
971 :ref:`configuration-of-traffic-dictionary`.
973 Example of execution of VSPERF in "trafficgen" mode:
975 .. code-block:: console
977 $ ./vsperf -m trafficgen --trafficgen IxNet --conf-file vsperf.conf \
978 --test-params "TRAFFIC={'traffic_type':'rfc2544_continuous','bidir':'False','framerate':60}"
983 The ``--matrix`` command line argument analyses and displays the performance of
984 all the tests run. Using the metric specified by ``MATRIX_METRIC`` in the conf-file,
985 the first test is set as the baseline and all the other tests are compared to it.
986 The ``MATRIX_METRIC`` must always refer to a numeric value to enable comparision.
987 A table, with the test ID, metric value, the change of the metric in %, testname
988 and the test parameters used for each test, is printed out as well as saved into the
991 Example of 2 tests being compared using Performance Matrix:
993 .. code-block:: console
995 $ ./vsperf --conf-file user_settings.py \
996 --test-params "['TRAFFICGEN_PKT_SIZES=(64,)',"\
997 "'TRAFFICGEN_PKT_SIZES=(128,)']" \
998 phy2phy_cont phy2phy_cont --matrix
1002 .. code-block:: console
1004 +------+--------------+---------------------+----------+---------------------------------------+
1005 | ID | Name | throughput_rx_fps | Change | Parameters, CUMULATIVE_PARAMS = False |
1006 +======+==============+=====================+==========+=======================================+
1007 | 0 | phy2phy_cont | 23749000.000 | 0 | 'TRAFFICGEN_PKT_SIZES': [64] |
1008 +------+--------------+---------------------+----------+---------------------------------------+
1009 | 1 | phy2phy_cont | 16850500.000 | -29.048 | 'TRAFFICGEN_PKT_SIZES': [128] |
1010 +------+--------------+---------------------+----------+---------------------------------------+
1013 Code change verification by pylint
1014 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1016 Every developer participating in VSPERF project should run
1017 pylint before his python code is submitted for review. Project
1018 specific configuration for pylint is available at 'pylint.rc'.
1020 Example of manual pylint invocation:
1022 .. code-block:: console
1024 $ pylint --rcfile ./pylintrc ./vsperf
1029 Custom image fails to boot
1030 ~~~~~~~~~~~~~~~~~~~~~~~~~~
1032 Using custom VM images may not boot within VSPerf pxp testing because of
1033 the drive boot and shared type which could be caused by a missing scsi
1034 driver inside the image. In case of issues you can try changing the drive
1037 .. code-block:: python
1039 GUEST_BOOT_DRIVE_TYPE = ['ide']
1040 GUEST_SHARED_DRIVE_TYPE = ['ide']
1042 OVS with DPDK and QEMU
1043 ~~~~~~~~~~~~~~~~~~~~~~~
1045 If you encounter the following error: "before (last 100 chars):
1046 '-path=/dev/hugepages,share=on: unable to map backing store for
1047 hugepages: Cannot allocate memory\r\n\r\n" during qemu initialization,
1048 check the amount of hugepages on your system:
1050 .. code-block:: console
1052 $ cat /proc/meminfo | grep HugePages
1055 By default the vswitchd is launched with 1Gb of memory, to change
1056 this, modify --socket-mem parameter in conf/02_vswitch.conf to allocate
1057 an appropriate amount of memory:
1059 .. code-block:: python
1061 DPDK_SOCKET_MEM = ['1024', '0']
1062 VSWITCHD_DPDK_ARGS = ['-c', '0x4', '-n', '4']
1063 VSWITCHD_DPDK_CONFIG = {
1064 'dpdk-init' : 'true',
1065 'dpdk-lcore-mask' : '0x4',
1066 'dpdk-socket-mem' : '1024,0',
1069 Note: Option ``VSWITCHD_DPDK_ARGS`` is used for vswitchd, which supports ``--dpdk``
1070 parameter. In recent vswitchd versions, option ``VSWITCHD_DPDK_CONFIG`` will be
1071 used to configure vswitchd via ``ovs-vsctl`` calls.
1077 For more information and details refer to the rest of vSwitchPerfuser documentation.