aboutsummaryrefslogtreecommitdiffstats
path: root/docs/testing/developer/requirements/vswitchperf_ltp.rst
diff options
context:
space:
mode:
authorTrevor Cooper <trevor.cooper@intel.com>2017-03-21 13:25:49 -0700
committerTrevor Cooper <trevor.cooper@intel.com>2017-03-21 13:25:49 -0700
commit32a5263216d79ad34041dca55357278f092bb931 (patch)
treee3bf10ba8190ad93d2b114c1903cc43e2dfbf6b6 /docs/testing/developer/requirements/vswitchperf_ltp.rst
parent59668b1aa68da507335ef351619d9f50019df762 (diff)
Moved doc files to testing document structure testing/user ... testing/developer and modified doc index to match dir structure
Change-Id: I4b1a535808a48773505fa7874c61707cd349fced Signed-off-by: Trevor Cooper <trevor.cooper@intel.com>
Diffstat (limited to 'docs/testing/developer/requirements/vswitchperf_ltp.rst')
-rw-r--r--docs/testing/developer/requirements/vswitchperf_ltp.rst1348
1 files changed, 1348 insertions, 0 deletions
diff --git a/docs/testing/developer/requirements/vswitchperf_ltp.rst b/docs/testing/developer/requirements/vswitchperf_ltp.rst
new file mode 100644
index 00000000..2b74d676
--- /dev/null
+++ b/docs/testing/developer/requirements/vswitchperf_ltp.rst
@@ -0,0 +1,1348 @@
+.. This work is licensed under a Creative Commons Attribution 4.0 International License.
+.. http://creativecommons.org/licenses/by/4.0
+.. (c) OPNFV, Intel Corporation, AT&T and others.
+
+.. 3.1
+
+*****************************
+VSPERF LEVEL TEST PLAN (LTP)
+*****************************
+
+===============
+Introduction
+===============
+
+The objective of the OPNFV project titled
+**Characterize vSwitch Performance for Telco NFV Use Cases**, is to
+evaluate the performance of virtual switches to identify its suitability for a
+Telco Network Function Virtualization (NFV) environment. The intention of this
+Level Test Plan (LTP) document is to specify the scope, approach, resources,
+and schedule of the virtual switch performance benchmarking activities in
+OPNFV. The test cases will be identified in a separate document called the
+Level Test Design (LTD) document.
+
+This document is currently in draft form.
+
+.. 3.1.1
+
+
+.. _doc-id:
+
+Document identifier
+=========================
+
+The document id will be used to uniquely identify versions of the LTP. The
+format for the document id will be: OPNFV\_vswitchperf\_LTP\_REL\_STATUS, where
+by the status is one of: draft, reviewed, corrected or final. The document id
+for this version of the LTP is: OPNFV\_vswitchperf\_LTP\_Colorado\_REVIEWED.
+
+.. 3.1.2
+
+.. _scope:
+
+Scope
+==========
+
+The main purpose of this project is to specify a suite of
+performance tests in order to objectively measure the current packet
+transfer characteristics of a virtual switch in the NFVI. The intent of
+the project is to facilitate the performance testing of any virtual switch.
+Thus, a generic suite of tests shall be developed, with no hard dependencies to
+a single implementation. In addition, the test case suite shall be
+architecture independent.
+
+The test cases developed in this project shall not form part of a
+separate test framework, all of these tests may be inserted into the
+Continuous Integration Test Framework and/or the Platform Functionality
+Test Framework - if a vSwitch becomes a standard component of an OPNFV
+release.
+
+.. 3.1.3
+
+References
+===============
+
+* `RFC 1242 Benchmarking Terminology for Network Interconnection
+ Devices <http://www.ietf.org/rfc/rfc1242.txt>`__
+* `RFC 2544 Benchmarking Methodology for Network Interconnect
+ Devices <http://www.ietf.org/rfc/rfc2544.txt>`__
+* `RFC 2285 Benchmarking Terminology for LAN Switching
+ Devices <http://www.ietf.org/rfc/rfc2285.txt>`__
+* `RFC 2889 Benchmarking Methodology for LAN Switching
+ Devices <http://www.ietf.org/rfc/rfc2889.txt>`__
+* `RFC 3918 Methodology for IP Multicast
+ Benchmarking <http://www.ietf.org/rfc/rfc3918.txt>`__
+* `RFC 4737 Packet Reordering
+ Metrics <http://www.ietf.org/rfc/rfc4737.txt>`__
+* `RFC 5481 Packet Delay Variation Applicability
+ Statement <http://www.ietf.org/rfc/rfc5481.txt>`__
+* `RFC 6201 Device Reset
+ Characterization <http://tools.ietf.org/html/rfc6201>`__
+
+.. 3.1.4
+
+Level in the overall sequence
+===============================
+The level of testing conducted by vswitchperf in the overall testing sequence (among
+all the testing projects in OPNFV) is the performance benchmarking of a
+specific component (the vswitch) in the OPNFV platfrom. It's expected that this
+testing will follow on from the functional and integration testing conducted by
+other testing projects in OPNFV, namely Functest and Yardstick.
+
+.. 3.1.5
+
+Test classes and overall test conditions
+=========================================
+A benchmark is defined by the IETF as: A standardized test that serves as a
+basis for performance evaluation and comparison. It's important to note that
+benchmarks are not Functional tests. They do not provide PASS/FAIL criteria,
+and most importantly ARE NOT performed on live networks, or performed with live
+network traffic.
+
+In order to determine the packet transfer characteristics of a virtual switch,
+the benchmarking tests will be broken down into the following categories:
+
+- **Throughput Tests** to measure the maximum forwarding rate (in
+ frames per second or fps) and bit rate (in Mbps) for a constant load
+ (as defined by `RFC1242 <https://www.rfc-editor.org/rfc/rfc1242.txt>`__)
+ without traffic loss.
+- **Packet and Frame Delay Tests** to measure average, min and max
+ packet and frame delay for constant loads.
+- **Stream Performance Tests** (TCP, UDP) to measure bulk data transfer
+ performance, i.e. how fast systems can send and receive data through
+ the virtual switch.
+- **Request/Response Performance** Tests (TCP, UDP) the measure the
+ transaction rate through the virtual switch.
+- **Packet Delay Tests** to understand latency distribution for
+ different packet sizes and over an extended test run to uncover
+ outliers.
+- **Scalability Tests** to understand how the virtual switch performs
+ as the number of flows, active ports, complexity of the forwarding
+ logic's configuration... it has to deal with increases.
+- **Control Path and Datapath Coupling** Tests, to understand how
+ closely coupled the datapath and the control path are as well as the
+ effect of this coupling on the performance of the DUT.
+- **CPU and Memory Consumption Tests** to understand the virtual
+ switch’s footprint on the system, this includes:
+
+ * CPU core utilization.
+ * CPU cache utilization.
+ * Memory footprint.
+ * System bus (QPI, PCI, ..) utilization.
+ * Memory lanes utilization.
+ * CPU cycles consumed per packet.
+ * Time To Establish Flows Tests.
+
+- **Noisy Neighbour Tests**, to understand the effects of resource
+ sharing on the performance of a virtual switch.
+
+**Note:** some of the tests above can be conducted simultaneously where
+the combined results would be insightful, for example Packet/Frame Delay
+and Scalability.
+
+
+
+.. 3.2
+
+.. _details-of-LTP:
+
+===================================
+Details of the Level Test Plan
+===================================
+
+This section describes the following items:
+* Test items and their identifiers (TestItems_)
+* Test Traceability Matrix (TestMatrix_)
+* Features to be tested (FeaturesToBeTested_)
+* Features not to be tested (FeaturesNotToBeTested_)
+* Approach (Approach_)
+* Item pass/fail criteria (PassFailCriteria_)
+* Suspension criteria and resumption requirements (SuspensionResumptionReqs_)
+
+.. 3.2.1
+
+.. _TestItems:
+
+Test items and their identifiers
+==================================
+The test item/application vsperf is trying to test are virtual switches and in
+particular their performance in an nfv environment. vsperf will first try to
+measure the maximum achievable performance by a virtual switch and then it will
+focus in on usecases that are as close to real life deployment scenarios as
+possible.
+
+.. 3.2.2
+
+.. _TestMatrix:
+
+Test Traceability Matrix
+==========================
+vswitchperf leverages the "3x3" matrix (introduced in
+https://tools.ietf.org/html/draft-ietf-bmwg-virtual-net-02) to achieve test
+traceability. The matrix was expanded to 3x4 to accommodate scale metrics when
+displaying the coverage of many metrics/benchmarks). Test case covreage in the
+LTD is tracked using the following catagories:
+
+
++---------------+-------------+------------+---------------+-------------+
+| | | | | |
+| | SPEED | ACCURACY | RELIABILITY | SCALE |
+| | | | | |
++---------------+-------------+------------+---------------+-------------+
+| | | | | |
+| Activation | X | X | X | X |
+| | | | | |
++---------------+-------------+------------+---------------+-------------+
+| | | | | |
+| Operation | X | X | X | X |
+| | | | | |
++---------------+-------------+------------+---------------+-------------+
+| | | | | |
+| De-activation | | | | |
+| | | | | |
++---------------+-------------+------------+---------------+-------------+
+
+X = denotes a test catagory that has 1 or more test cases defined.
+
+.. 3.2.3
+
+.. _FeaturesToBeTested:
+
+Features to be tested
+==========================
+
+Characterizing virtual switches (i.e. Device Under Test (DUT) in this document)
+includes measuring the following performance metrics:
+
+- **Throughput** as defined by `RFC1242
+ <https://www.rfc-editor.org/rfc/rfc1242.txt>`__: The maximum rate at which
+ **none** of the offered frames are dropped by the DUT. The maximum frame
+ rate and bit rate that can be transmitted by the DUT without any error
+ should be recorded. Note there is an equivalent bit rate and a specific
+ layer at which the payloads contribute to the bits. Errors and
+ improperly formed frames or packets are dropped.
+- **Packet delay** introduced by the DUT and its cumulative effect on
+ E2E networks. Frame delay can be measured equivalently.
+- **Packet delay variation**: measured from the perspective of the
+ VNF/application. Packet delay variation is sometimes called "jitter".
+ However, we will avoid the term "jitter" as the term holds different
+ meaning to different groups of people. In this document we will
+ simply use the term packet delay variation. The preferred form for this
+ metric is the PDV form of delay variation defined in `RFC5481
+ <https://www.rfc-editor.org/rfc/rfc5481.txt>`__. The most relevant
+ measurement of PDV considers the delay variation of a single user flow,
+ as this will be relevant to the size of end-system buffers to compensate
+ for delay variation. The measurement system's ability to store the
+ delays of individual packets in the flow of interest is a key factor
+ that determines the specific measurement method. At the outset, it is
+ ideal to view the complete PDV distribution. Systems that can capture
+ and store packets and their delays have the freedom to calculate the
+ reference minimum delay and to determine various quantiles of the PDV
+ distribution accurately (in post-measurement processing routines).
+ Systems without storage must apply algorithms to calculate delay and
+ statistical measurements on the fly. For example, a system may store
+ temporary estimates of the mimimum delay and the set of (100) packets
+ with the longest delays during measurement (to calculate a high quantile,
+ and update these sets with new values periodically.
+ In some cases, a limited number of delay histogram bins will be
+ available, and the bin limits will need to be set using results from
+ repeated experiments. See section 8 of `RFC5481
+ <https://www.rfc-editor.org/rfc/rfc5481.txt>`__.
+- **Packet loss** (within a configured waiting time at the receiver): All
+ packets sent to the DUT should be accounted for.
+- **Burst behaviour**: measures the ability of the DUT to buffer packets.
+- **Packet re-ordering**: measures the ability of the device under test to
+ maintain sending order throughout transfer to the destination.
+- **Packet correctness**: packets or Frames must be well-formed, in that
+ they include all required fields, conform to length requirements, pass
+ integrity checks, etc.
+- **Availability and capacity** of the DUT i.e. when the DUT is fully “up”
+ and connected, following measurements should be captured for
+ DUT without any network packet load:
+
+ - Includes average power consumption of the CPUs (in various power states) and
+ system over specified period of time. Time period should not be less
+ than 60 seconds.
+ - Includes average per core CPU utilization over specified period of time.
+ Time period should not be less than 60 seconds.
+ - Includes the number of NIC interfaces supported.
+ - Includes headroom of VM workload processing cores (i.e. available
+ for applications).
+
+.. 3.2.4
+
+.. _FeaturesNotToBeTested:
+
+Features not to be tested
+==========================
+vsperf doesn't intend to define or perform any functional tests. The aim is to
+focus on performance.
+
+.. 3.2.5
+
+.. _Approach:
+
+Approach
+==============
+The testing approach adoped by the vswitchperf project is black box testing,
+meaning the test inputs can be generated and the outputs captured and
+completely evaluated from the outside of the System Under Test. Some metrics
+can be collected on the SUT, such as cpu or memory utilization if the
+collection has no/minimal impact on benchmark.
+This section will look at the deployment scenarios and the general methodology
+used by vswitchperf. In addition, this section will also specify the details of
+the Test Report that must be collected for each of the test cases.
+
+.. 3.2.5.1
+
+Deployment Scenarios
+--------------------------
+The following represents possible deployment test scenarios which can
+help to determine the performance of both the virtual switch and the
+datapaths to physical ports (to NICs) and to logical ports (to VNFs):
+
+.. 3.2.5.1.1
+
+.. _Phy2Phy:
+
+Physical port → vSwitch → physical port
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+.. code-block:: console
+
+ _
+ +--------------------------------------------------+ |
+ | +--------------------+ | |
+ | | | | |
+ | | v | | Host
+ | +--------------+ +--------------+ | |
+ | | phy port | vSwitch | phy port | | |
+ +---+--------------+------------+--------------+---+ _|
+ ^ :
+ | |
+ : v
+ +--------------------------------------------------+
+ | |
+ | traffic generator |
+ | |
+ +--------------------------------------------------+
+
+.. 3.2.5.1.2
+
+.. _PVP:
+
+Physical port → vSwitch → VNF → vSwitch → physical port
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+.. code-block:: console
+
+ _
+ +---------------------------------------------------+ |
+ | | |
+ | +-------------------------------------------+ | |
+ | | Application | | |
+ | +-------------------------------------------+ | |
+ | ^ : | |
+ | | | | | Guest
+ | : v | |
+ | +---------------+ +---------------+ | |
+ | | logical port 0| | logical port 1| | |
+ +---+---------------+-----------+---------------+---+ _|
+ ^ :
+ | |
+ : v _
+ +---+---------------+----------+---------------+---+ |
+ | | logical port 0| | logical port 1| | |
+ | +---------------+ +---------------+ | |
+ | ^ : | |
+ | | | | | Host
+ | : v | |
+ | +--------------+ +--------------+ | |
+ | | phy port | vSwitch | phy port | | |
+ +---+--------------+------------+--------------+---+ _|
+ ^ :
+ | |
+ : v
+ +--------------------------------------------------+
+ | |
+ | traffic generator |
+ | |
+ +--------------------------------------------------+
+
+.. 3.2.5.1.3
+
+.. _PVVP:
+
+Physical port → vSwitch → VNF → vSwitch → VNF → vSwitch → physical port
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: console
+
+ _
+ +----------------------+ +----------------------+ |
+ | Guest 1 | | Guest 2 | |
+ | +---------------+ | | +---------------+ | |
+ | | Application | | | | Application | | |
+ | +---------------+ | | +---------------+ | |
+ | ^ | | | ^ | | |
+ | | v | | | v | | Guests
+ | +---------------+ | | +---------------+ | |
+ | | logical ports | | | | logical ports | | |
+ | | 0 1 | | | | 0 1 | | |
+ +---+---------------+--+ +---+---------------+--+ _|
+ ^ : ^ :
+ | | | |
+ : v : v _
+ +---+---------------+---------+---------------+--+ |
+ | | 0 1 | | 3 4 | | |
+ | | logical ports | | logical ports | | |
+ | +---------------+ +---------------+ | |
+ | ^ | ^ | | | Host
+ | | L-----------------+ v | |
+ | +--------------+ +--------------+ | |
+ | | phy ports | vSwitch | phy ports | | |
+ +---+--------------+----------+--------------+---+ _|
+ ^ ^ : :
+ | | | |
+ : : v v
+ +--------------------------------------------------+
+ | |
+ | traffic generator |
+ | |
+ +--------------------------------------------------+
+
+.. 3.2.5.1.4
+
+Physical port → VNF → vSwitch → VNF → physical port
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: console
+
+ _
+ +----------------------+ +----------------------+ |
+ | Guest 1 | | Guest 2 | |
+ |+-------------------+ | | +-------------------+| |
+ || Application | | | | Application || |
+ |+-------------------+ | | +-------------------+| |
+ | ^ | | | ^ | | | Guests
+ | | v | | | v | |
+ |+-------------------+ | | +-------------------+| |
+ || logical ports | | | | logical ports || |
+ || 0 1 | | | | 0 1 || |
+ ++--------------------++ ++--------------------++ _|
+ ^ : ^ :
+ (PCI passthrough) | | (PCI passthrough)
+ | v : | _
+ +--------++------------+-+------------++---------+ |
+ | | || 0 | | 1 || | | |
+ | | ||logical port| |logical port|| | | |
+ | | |+------------+ +------------+| | | |
+ | | | | ^ | | | |
+ | | | L-----------------+ | | | |
+ | | | | | | | Host
+ | | | vSwitch | | | |
+ | | +-----------------------------+ | | |
+ | | | | |
+ | | v | |
+ | +--------------+ +--------------+ | |
+ | | phy port/VF | | phy port/VF | | |
+ +-+--------------+--------------+--------------+-+ _|
+ ^ :
+ | |
+ : v
+ +--------------------------------------------------+
+ | |
+ | traffic generator |
+ | |
+ +--------------------------------------------------+
+
+.. 3.2.5.1.5
+
+Physical port → vSwitch → VNF
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: console
+
+ _
+ +---------------------------------------------------+ |
+ | | |
+ | +-------------------------------------------+ | |
+ | | Application | | |
+ | +-------------------------------------------+ | |
+ | ^ | |
+ | | | | Guest
+ | : | |
+ | +---------------+ | |
+ | | logical port 0| | |
+ +---+---------------+-------------------------------+ _|
+ ^
+ |
+ : _
+ +---+---------------+------------------------------+ |
+ | | logical port 0| | |
+ | +---------------+ | |
+ | ^ | |
+ | | | | Host
+ | : | |
+ | +--------------+ | |
+ | | phy port | vSwitch | |
+ +---+--------------+------------ -------------- ---+ _|
+ ^
+ |
+ :
+ +--------------------------------------------------+
+ | |
+ | traffic generator |
+ | |
+ +--------------------------------------------------+
+
+.. 3.2.5.1.6
+
+VNF → vSwitch → physical port
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: console
+
+ _
+ +---------------------------------------------------+ |
+ | | |
+ | +-------------------------------------------+ | |
+ | | Application | | |
+ | +-------------------------------------------+ | |
+ | : | |
+ | | | | Guest
+ | v | |
+ | +---------------+ | |
+ | | logical port | | |
+ +-------------------------------+---------------+---+ _|
+ :
+ |
+ v _
+ +------------------------------+---------------+---+ |
+ | | logical port | | |
+ | +---------------+ | |
+ | : | |
+ | | | | Host
+ | v | |
+ | +--------------+ | |
+ | vSwitch | phy port | | |
+ +-------------------------------+--------------+---+ _|
+ :
+ |
+ v
+ +--------------------------------------------------+
+ | |
+ | traffic generator |
+ | |
+ +--------------------------------------------------+
+
+.. 3.2.5.1.7
+
+VNF → vSwitch → VNF → vSwitch
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: console
+
+ _
+ +-------------------------+ +-------------------------+ |
+ | Guest 1 | | Guest 2 | |
+ | +-----------------+ | | +-----------------+ | |
+ | | Application | | | | Application | | |
+ | +-----------------+ | | +-----------------+ | |
+ | : | | ^ | |
+ | | | | | | | Guest
+ | v | | : | |
+ | +---------------+ | | +---------------+ | |
+ | | logical port 0| | | | logical port 0| | |
+ +-----+---------------+---+ +---+---------------+-----+ _|
+ : ^
+ | |
+ v : _
+ +----+---------------+------------+---------------+-----+ |
+ | | port 0 | | port 1 | | |
+ | +---------------+ +---------------+ | |
+ | : ^ | |
+ | | | | | Host
+ | +--------------------+ | |
+ | | |
+ | vswitch | |
+ +-------------------------------------------------------+ _|
+
+.. 3.2.5.1.8
+
+HOST 1(Physical port → virtual switch → VNF → virtual switch → Physical port)
+→ HOST 2(Physical port → virtual switch → VNF → virtual switch → Physical port)
+
+HOST 1 (PVP) → HOST 2 (PVP)
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. code-block:: console
+
+ _
+ +----------------------+ +----------------------+ |
+ | Guest 1 | | Guest 2 | |
+ | +---------------+ | | +---------------+ | |
+ | | Application | | | | Application | | |
+ | +---------------+ | | +---------------+ | |
+ | ^ | | | ^ | | |
+ | | v | | | v | | Guests
+ | +---------------+ | | +---------------+ | |
+ | | logical ports | | | | logical ports | | |
+ | | 0 1 | | | | 0 1 | | |
+ +---+---------------+--+ +---+---------------+--+ _|
+ ^ : ^ :
+ | | | |
+ : v : v _
+ +---+---------------+--+ +---+---------------+--+ |
+ | | 0 1 | | | | 3 4 | | |
+ | | logical ports | | | | logical ports | | |
+ | +---------------+ | | +---------------+ | |
+ | ^ | | | ^ | | | Hosts
+ | | v | | | v | |
+ | +--------------+ | | +--------------+ | |
+ | | phy ports | | | | phy ports | | |
+ +---+--------------+---+ +---+--------------+---+ _|
+ ^ : : :
+ | +-----------------+ |
+ : v
+ +--------------------------------------------------+
+ | |
+ | traffic generator |
+ | |
+ +--------------------------------------------------+
+
+
+
+**Note:** For tests where the traffic generator and/or measurement
+receiver are implemented on VM and connected to the virtual switch
+through vNIC, the issues of shared resources and interactions between
+the measurement devices and the device under test must be considered.
+
+**Note:** Some RFC 2889 tests require a full-mesh sending and receiving
+pattern involving more than two ports. This possibility is illustrated in the
+Physical port → vSwitch → VNF → vSwitch → VNF → vSwitch → physical port
+diagram above (with 2 sending and 2 receiving ports, though all ports
+could be used bi-directionally).
+
+**Note:** When Deployment Scenarios are used in RFC 2889 address learning
+or cache capacity testing, an additional port from the vSwitch must be
+connected to the test device. This port is used to listen for flooded
+frames.
+
+.. 3.2.5.2
+
+General Methodology:
+--------------------------
+To establish the baseline performance of the virtual switch, tests would
+initially be run with a simple workload in the VNF (the recommended
+simple workload VNF would be `DPDK <http://www.dpdk.org/>`__'s testpmd
+application forwarding packets in a VM or vloop\_vnf a simple kernel
+module that forwards traffic between two network interfaces inside the
+virtualized environment while bypassing the networking stack).
+Subsequently, the tests would also be executed with a real Telco
+workload running in the VNF, which would exercise the virtual switch in
+the context of higher level Telco NFV use cases, and prove that its
+underlying characteristics and behaviour can be measured and validated.
+Suitable real Telco workload VNFs are yet to be identified.
+
+.. 3.2.5.2.1
+
+.. _default-test-parameters:
+
+Default Test Parameters
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The following list identifies the default parameters for suite of
+tests:
+
+- Reference application: Simple forwarding or Open Source VNF.
+- Frame size (bytes): 64, 128, 256, 512, 1024, 1280, 1518, 2K, 4k OR
+ Packet size based on use-case (e.g. RTP 64B, 256B) OR Mix of packet sizes as
+ maintained by the Functest project <https://wiki.opnfv.org/traffic_profile_management>.
+- Reordering check: Tests should confirm that packets within a flow are
+ not reordered.
+- Duplex: Unidirectional / Bidirectional. Default: Full duplex with
+ traffic transmitting in both directions, as network traffic generally
+ does not flow in a single direction. By default the data rate of
+ transmitted traffic should be the same in both directions, please
+ note that asymmetric traffic (e.g. downlink-heavy) tests will be
+ mentioned explicitly for the relevant test cases.
+- Number of Flows: Default for non scalability tests is a single flow.
+ For scalability tests the goal is to test with maximum supported
+ flows but where possible will test up to 10 Million flows. Start with
+ a single flow and scale up. By default flows should be added
+ sequentially, tests that add flows simultaneously will explicitly
+ call out their flow addition behaviour. Packets are generated across
+ the flows uniformly with no burstiness. For multi-core tests should
+ consider the number of packet flows based on vSwitch/VNF multi-thread
+ implementation and behavior.
+
+- Traffic Types: UDP, SCTP, RTP, GTP and UDP traffic.
+- Deployment scenarios are:
+- Physical → virtual switch → physical.
+- Physical → virtual switch → VNF → virtual switch → physical.
+- Physical → virtual switch → VNF → virtual switch → VNF → virtual
+ switch → physical.
+- Physical → VNF → virtual switch → VNF → physical.
+- Physical → virtual switch → VNF.
+- VNF → virtual switch → Physical.
+- VNF → virtual switch → VNF.
+
+Tests MUST have these parameters unless otherwise stated. **Test cases
+with non default parameters will be stated explicitly**.
+
+**Note**: For throughput tests unless stated otherwise, test
+configurations should ensure that traffic traverses the installed flows
+through the virtual switch, i.e. flows are installed and have an appropriate
+time out that doesn't expire before packet transmission starts.
+
+.. 3.2.5.2.2
+
+Flow Classification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Virtual switches classify packets into flows by processing and matching
+particular header fields in the packet/frame and/or the input port where
+the packets/frames arrived. The vSwitch then carries out an action on
+the group of packets that match the classification parameters. Thus a
+flow is considered to be a sequence of packets that have a shared set of
+header field values or have arrived on the same port and have the same
+action applied to them. Performance results can vary based on the
+parameters the vSwitch uses to match for a flow. The recommended flow
+classification parameters for L3 vSwitch performance tests are: the
+input port, the source IP address, the destination IP address and the
+Ethernet protocol type field. It is essential to increase the flow
+time-out time on a vSwitch before conducting any performance tests that
+do not measure the flow set-up time. Normally the first packet of a
+particular flow will install the flow in the vSwitch which adds an
+additional latency, subsequent packets of the same flow are not subject
+to this latency if the flow is already installed on the vSwitch.
+
+.. 3.2.5.2.3
+
+Test Priority
+~~~~~~~~~~~~~~~~~~~~~
+
+Tests will be assigned a priority in order to determine which tests
+should be implemented immediately and which tests implementations
+can be deferred.
+
+Priority can be of following types: - Urgent: Must be implemented
+immediately. - High: Must be implemented in the next release. - Medium:
+May be implemented after the release. - Low: May or may not be
+implemented at all.
+
+.. 3.2.5.2.4
+
+SUT Setup
+~~~~~~~~~~~~~~~~~~
+
+The SUT should be configured to its "default" state. The
+SUT's configuration or set-up must not change between tests in any way
+other than what is required to do the test. All supported protocols must
+be configured and enabled for each test set up.
+
+.. 3.2.5.2.5
+
+Port Configuration
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The DUT should be configured with n ports where
+n is a multiple of 2. Half of the ports on the DUT should be used as
+ingress ports and the other half of the ports on the DUT should be used
+as egress ports. Where a DUT has more than 2 ports, the ingress data
+streams should be set-up so that they transmit packets to the egress
+ports in sequence so that there is an even distribution of traffic
+across ports. For example, if a DUT has 4 ports 0(ingress), 1(ingress),
+2(egress) and 3(egress), the traffic stream directed at port 0 should
+output a packet to port 2 followed by a packet to port 3. The traffic
+stream directed at port 1 should also output a packet to port 2 followed
+by a packet to port 3.
+
+.. 3.2.5.2.6
+
+Frame Formats
+~~~~~~~~~~~~~~~~~~~~~
+
+**Frame formats Layer 2 (data link layer) protocols**
+
+- Ethernet II
+
+.. code-block:: console
+
+ +---------------------------+-----------+
+ | Ethernet Header | Payload | Check Sum |
+ +-----------------+---------+-----------+
+ |_________________|_________|___________|
+ 14 Bytes 46 - 1500 4 Bytes
+ Bytes
+
+
+**Layer 3 (network layer) protocols**
+
+- IPv4
+
+.. code-block:: console
+
+ +-----------------+-----------+---------+-----------+
+ | Ethernet Header | IP Header | Payload | Checksum |
+ +-----------------+-----------+---------+-----------+
+ |_________________|___________|_________|___________|
+ 14 Bytes 20 bytes 26 - 1480 4 Bytes
+ Bytes
+
+- IPv6
+
+.. code-block:: console
+
+ +-----------------+-----------+---------+-----------+
+ | Ethernet Header | IP Header | Payload | Checksum |
+ +-----------------+-----------+---------+-----------+
+ |_________________|___________|_________|___________|
+ 14 Bytes 40 bytes 26 - 1460 4 Bytes
+ Bytes
+
+**Layer 4 (transport layer) protocols**
+
+ - TCP
+ - UDP
+ - SCTP
+
+.. code-block:: console
+
+ +-----------------+-----------+-----------------+---------+-----------+
+ | Ethernet Header | IP Header | Layer 4 Header | Payload | Checksum |
+ +-----------------+-----------+-----------------+---------+-----------+
+ |_________________|___________|_________________|_________|___________|
+ 14 Bytes 40 bytes 20 Bytes 6 - 1460 4 Bytes
+ Bytes
+
+
+**Layer 5 (application layer) protocols**
+
+ - RTP
+ - GTP
+
+.. code-block:: console
+
+ +-----------------+-----------+-----------------+---------+-----------+
+ | Ethernet Header | IP Header | Layer 4 Header | Payload | Checksum |
+ +-----------------+-----------+-----------------+---------+-----------+
+ |_________________|___________|_________________|_________|___________|
+ 14 Bytes 20 bytes 20 Bytes >= 6 Bytes 4 Bytes
+
+.. 3.2.5.2.7
+
+Packet Throughput
+~~~~~~~~~~~~~~~~~~~~~~~~~
+There is a difference between an Ethernet frame,
+an IP packet, and a UDP datagram. In the seven-layer OSI model of
+computer networking, packet refers to a data unit at layer 3 (network
+layer). The correct term for a data unit at layer 2 (data link layer) is
+a frame, and at layer 4 (transport layer) is a segment or datagram.
+
+Important concepts related to 10GbE performance are frame rate and
+throughput. The MAC bit rate of 10GbE, defined in the IEEE standard 802
+.3ae, is 10 billion bits per second. Frame rate is based on the bit rate
+and frame format definitions. Throughput, defined in IETF RFC 1242, is
+the highest rate at which the system under test can forward the offered
+load, without loss.
+
+The frame rate for 10GbE is determined by a formula that divides the 10
+billion bits per second by the preamble + frame length + inter-frame
+gap.
+
+The maximum frame rate is calculated using the minimum values of the
+following parameters, as described in the IEEE 802 .3ae standard:
+
+- Preamble: 8 bytes \* 8 = 64 bits
+- Frame Length: 64 bytes (minimum) \* 8 = 512 bits
+- Inter-frame Gap: 12 bytes (minimum) \* 8 = 96 bits
+
+Therefore, Maximum Frame Rate (64B Frames)
+= MAC Transmit Bit Rate / (Preamble + Frame Length + Inter-frame Gap)
+= 10,000,000,000 / (64 + 512 + 96)
+= 10,000,000,000 / 672
+= 14,880,952.38 frame per second (fps)
+
+.. 3.2.5.3
+
+RFCs for testing virtual switch performance
+--------------------------------------------------
+
+The starting point for defining the suite of tests for benchmarking the
+performance of a virtual switch is to take existing RFCs and standards
+that were designed to test their physical counterparts and adapting them
+for testing virtual switches. The rationale behind this is to establish
+a fair comparison between the performance of virtual and physical
+switches. This section outlines the RFCs that are used by this
+specification.
+
+.. 3.2.5.3.1
+
+RFC 1242 Benchmarking Terminology for Network Interconnection
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Devices RFC 1242 defines the terminology that is used in describing
+performance benchmarking tests and their results. Definitions and
+discussions covered include: Back-to-back, bridge, bridge/router,
+constant load, data link frame size, frame loss rate, inter frame gap,
+latency, and many more.
+
+.. 3.2.5.3.2
+
+RFC 2544 Benchmarking Methodology for Network Interconnect Devices
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+RFC 2544 outlines a benchmarking methodology for network Interconnect
+Devices. The methodology results in performance metrics such as latency,
+frame loss percentage, and maximum data throughput.
+
+In this document network “throughput” (measured in millions of frames
+per second) is based on RFC 2544, unless otherwise noted. Frame size
+refers to Ethernet frames ranging from smallest frames of 64 bytes to
+largest frames of 9K bytes.
+
+Types of tests are:
+
+1. Throughput test defines the maximum number of frames per second
+ that can be transmitted without any error, or 0% loss ratio.
+ In some Throughput tests (and those tests with long duration),
+ evaluation of an additional frame loss ratio is suggested. The
+ current ratio (10^-7 %) is based on understanding the typical
+ user-to-user packet loss ratio needed for good application
+ performance and recognizing that a single transfer through a
+ vswitch must contribute a tiny fraction of user-to-user loss.
+ Further, the ratio 10^-7 % also recognizes practical limitations
+ when measuring loss ratio.
+
+2. Latency test measures the time required for a frame to travel from
+ the originating device through the network to the destination device.
+ Please note that RFC2544 Latency measurement will be superseded with
+ a measurement of average latency over all successfully transferred
+ packets or frames.
+
+3. Frame loss test measures the network’s
+ response in overload conditions - a critical indicator of the
+ network’s ability to support real-time applications in which a
+ large amount of frame loss will rapidly degrade service quality.
+
+4. Burst test assesses the buffering capability of a virtual switch. It
+ measures the maximum number of frames received at full line rate
+ before a frame is lost. In carrier Ethernet networks, this
+ measurement validates the excess information rate (EIR) as defined in
+ many SLAs.
+
+5. System recovery to characterize speed of recovery from an overload
+ condition.
+
+6. Reset to characterize speed of recovery from device or software
+ reset. This type of test has been updated by `RFC6201
+ <https://www.rfc-editor.org/rfc/rfc6201.txt>`__ as such,
+ the methodology defined by this specification will be that of RFC 6201.
+
+Although not included in the defined RFC 2544 standard, another crucial
+measurement in Ethernet networking is packet delay variation. The
+definition set out by this specification comes from
+`RFC5481 <https://www.rfc-editor.org/rfc/rfc5481.txt>`__.
+
+.. 3.2.5.3.3
+
+RFC 2285 Benchmarking Terminology for LAN Switching Devices
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+RFC 2285 defines the terminology that is used to describe the
+terminology for benchmarking a LAN switching device. It extends RFC
+1242 and defines: DUTs, SUTs, Traffic orientation and distribution,
+bursts, loads, forwarding rates, etc.
+
+.. 3.2.5.3.4
+
+RFC 2889 Benchmarking Methodology for LAN Switching
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+RFC 2889 outlines a benchmarking methodology for LAN switching, it
+extends RFC 2544. The outlined methodology gathers performance
+metrics for forwarding, congestion control, latency, address handling
+and finally filtering.
+
+.. 3.2.5.3.5
+
+RFC 3918 Methodology for IP Multicast Benchmarking
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+RFC 3918 outlines a methodology for IP Multicast benchmarking.
+
+.. 3.2.5.3.6
+
+RFC 4737 Packet Reordering Metrics
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+RFC 4737 describes metrics for identifying and counting re-ordered
+packets within a stream, and metrics to measure the extent each
+packet has been re-ordered.
+
+.. 3.2.5.3.7
+
+RFC 5481 Packet Delay Variation Applicability Statement
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+RFC 5481 defined two common, but different forms of delay variation
+metrics, and compares the metrics over a range of networking
+circumstances and tasks. The most suitable form for vSwitch
+benchmarking is the "PDV" form.
+
+.. 3.2.5.3.8
+
+RFC 6201 Device Reset Characterization
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+RFC 6201 extends the methodology for characterizing the speed of
+recovery of the DUT from device or software reset described in RFC
+2544.
+
+.. 3.2.6:
+
+.. _PassFailCriteria:
+
+Item pass/fail criteria
+=========================
+
+vswitchperf does not specify Pass/Fail criteria for the tests in terms of a
+threshold, as benchmarks do not (and should not do this). The results/metrics
+for a test are simply reported. If it had to be defined, a test is considered
+to have passed if it succesfully completed and a relavent metric was
+recorded/reported for the SUT.
+
+.. 3.2.7:
+
+.. _SuspensionResumptionReqs:
+
+Suspension criteria and resumption requirements
+================================================
+In the case of a throughput test, a test should be suspended if a virtual
+switch is failing to forward any traffic. A test should be restarted from a
+clean state if the intention is to carry out the test again.
+
+.. 3.2.8:
+
+.. _TestDelierables:
+
+Test deliverables
+==================
+Each test should produce a test report that details SUT information as well as
+the test results. There are a number of parameters related to the system, DUT
+and tests that can affect the repeatability of a test results and should be
+recorded. In order to minimise the variation in the results of a test,
+it is recommended that the test report includes the following information:
+
+- Hardware details including:
+
+ - Platform details.
+ - Processor details.
+ - Memory information (see below)
+ - Number of enabled cores.
+ - Number of cores used for the test.
+ - Number of physical NICs, as well as their details (manufacturer,
+ versions, type and the PCI slot they are plugged into).
+ - NIC interrupt configuration.
+ - BIOS version, release date and any configurations that were
+ modified.
+
+- Software details including:
+
+ - OS version (for host and VNF)
+ - Kernel version (for host and VNF)
+ - GRUB boot parameters (for host and VNF).
+ - Hypervisor details (Type and version).
+ - Selected vSwitch, version number or commit id used.
+ - vSwitch launch command line if it has been parameterised.
+ - Memory allocation to the vSwitch – which NUMA node it is using,
+ and how many memory channels.
+ - Where the vswitch is built from source: compiler details including
+ versions and the flags that were used to compile the vSwitch.
+ - DPDK or any other SW dependency version number or commit id used.
+ - Memory allocation to a VM - if it's from Hugpages/elsewhere.
+ - VM storage type: snapshot/independent persistent/independent
+ non-persistent.
+ - Number of VMs.
+ - Number of Virtual NICs (vNICs), versions, type and driver.
+ - Number of virtual CPUs and their core affinity on the host.
+ - Number vNIC interrupt configuration.
+ - Thread affinitization for the applications (including the vSwitch
+ itself) on the host.
+ - Details of Resource isolation, such as CPUs designated for
+ Host/Kernel (isolcpu) and CPUs designated for specific processes
+ (taskset).
+
+- Memory Details
+
+ - Total memory
+ - Type of memory
+ - Used memory
+ - Active memory
+ - Inactive memory
+ - Free memory
+ - Buffer memory
+ - Swap cache
+ - Total swap
+ - Used swap
+ - Free swap
+
+- Test duration.
+- Number of flows.
+- Traffic Information:
+
+ - Traffic type - UDP, TCP, IMIX / Other.
+ - Packet Sizes.
+
+- Deployment Scenario.
+
+**Note**: Tests that require additional parameters to be recorded will
+explicitly specify this.
+
+
+.. 3.3:
+
+.. _TestManagement:
+
+Test management
+=================
+This section will detail the test activities that will be conducted by vsperf
+as well as the infrastructure that will be used to complete the tests in OPNFV.
+
+.. 3.3.1:
+
+Planned activities and tasks; test progression
+=================================================
+A key consideration when conducting any sort of benchmark is trying to
+ensure the consistency and repeatability of test results between runs.
+When benchmarking the performance of a virtual switch there are many
+factors that can affect the consistency of results. This section
+describes these factors and the measures that can be taken to limit
+their effects. In addition, this section will outline some system tests
+to validate the platform and the VNF before conducting any vSwitch
+benchmarking tests.
+
+**System Isolation:**
+
+When conducting a benchmarking test on any SUT, it is essential to limit
+(and if reasonable, eliminate) any noise that may interfere with the
+accuracy of the metrics collected by the test. This noise may be
+introduced by other hardware or software (OS, other applications), and
+can result in significantly varying performance metrics being collected
+between consecutive runs of the same test. In the case of characterizing
+the performance of a virtual switch, there are a number of configuration
+parameters that can help increase the repeatability and stability of
+test results, including:
+
+- OS/GRUB configuration:
+
+ - maxcpus = n where n >= 0; limits the kernel to using 'n'
+ processors. Only use exactly what you need.
+ - isolcpus: Isolate CPUs from the general scheduler. Isolate all
+ CPUs bar one which will be used by the OS.
+ - use taskset to affinitize the forwarding application and the VNFs
+ onto isolated cores. VNFs and the vSwitch should be allocated
+ their own cores, i.e. must not share the same cores. vCPUs for the
+ VNF should be affinitized to individual cores also.
+ - Limit the amount of background applications that are running and
+ set OS to boot to runlevel 3. Make sure to kill any unnecessary
+ system processes/daemons.
+ - Only enable hardware that you need to use for your test – to
+ ensure there are no other interrupts on the system.
+ - Configure NIC interrupts to only use the cores that are not
+ allocated to any other process (VNF/vSwitch).
+
+- NUMA configuration: Any unused sockets in a multi-socket system
+ should be disabled.
+- CPU pinning: The vSwitch and the VNF should each be affinitized to
+ separate logical cores using a combination of maxcpus, isolcpus and
+ taskset.
+- BIOS configuration: BIOS should be configured for performance where
+ an explicit option exists, sleep states should be disabled, any
+ virtualization optimization technologies should be enabled, and
+ hyperthreading should also be enabled, turbo boost and overclocking
+ should be disabled.
+
+**System Validation:**
+
+System validation is broken down into two sub-categories: Platform
+validation and VNF validation. The validation test itself involves
+verifying the forwarding capability and stability for the sub-system
+under test. The rationale behind system validation is two fold. Firstly
+to give a tester confidence in the stability of the platform or VNF that
+is being tested; and secondly to provide base performance comparison
+points to understand the overhead introduced by the virtual switch.
+
+* Benchmark platform forwarding capability: This is an OPTIONAL test
+ used to verify the platform and measure the base performance (maximum
+ forwarding rate in fps and latency) that can be achieved by the
+ platform without a vSwitch or a VNF. The following diagram outlines
+ the set-up for benchmarking Platform forwarding capability:
+
+ .. code-block:: console
+
+ __
+ +--------------------------------------------------+ |
+ | +------------------------------------------+ | |
+ | | | | |
+ | | l2fw or DPDK L2FWD app | | Host
+ | | | | |
+ | +------------------------------------------+ | |
+ | | NIC | | |
+ +---+------------------------------------------+---+ __|
+ ^ :
+ | |
+ : v
+ +--------------------------------------------------+
+ | |
+ | traffic generator |
+ | |
+ +--------------------------------------------------+
+
+* Benchmark VNF forwarding capability: This test is used to verify
+ the VNF and measure the base performance (maximum forwarding rate in
+ fps and latency) that can be achieved by the VNF without a vSwitch.
+ The performance metrics collected by this test will serve as a key
+ comparison point for NIC passthrough technologies and vSwitches. VNF
+ in this context refers to the hypervisor and the VM. The following
+ diagram outlines the set-up for benchmarking VNF forwarding
+ capability:
+
+ .. code-block:: console
+
+ __
+ +--------------------------------------------------+ |
+ | +------------------------------------------+ | |
+ | | | | |
+ | | VNF | | |
+ | | | | |
+ | +------------------------------------------+ | |
+ | | Passthrough/SR-IOV | | Host
+ | +------------------------------------------+ | |
+ | | NIC | | |
+ +---+------------------------------------------+---+ __|
+ ^ :
+ | |
+ : v
+ +--------------------------------------------------+
+ | |
+ | traffic generator |
+ | |
+ +--------------------------------------------------+
+
+
+**Methodology to benchmark Platform/VNF forwarding capability**
+
+
+The recommended methodology for the platform/VNF validation and
+benchmark is: - Run `RFC2889 <https://www.rfc-editor.org/rfc/rfc2289.txt>`__
+Maximum Forwarding Rate test, this test will produce maximum
+forwarding rate and latency results that will serve as the
+expected values. These expected values can be used in
+subsequent steps or compared with in subsequent validation tests. -
+Transmit bidirectional traffic at line rate/max forwarding rate
+(whichever is higher) for at least 72 hours, measure throughput (fps)
+and latency. - Note: Traffic should be bidirectional. - Establish a
+baseline forwarding rate for what the platform can achieve. - Additional
+validation: After the test has completed for 72 hours run bidirectional
+traffic at the maximum forwarding rate once more to see if the system is
+still functional and measure throughput (fps) and latency. Compare the
+measure the new obtained values with the expected values.
+
+**NOTE 1**: How the Platform is configured for its forwarding capability
+test (BIOS settings, GRUB configuration, runlevel...) is how the
+platform should be configured for every test after this
+
+**NOTE 2**: How the VNF is configured for its forwarding capability test
+(# of vCPUs, vNICs, Memory, affinitization…) is how it should be
+configured for every test that uses a VNF after this.
+
+**Methodology to benchmark the VNF to vSwitch to VNF deployment scenario**
+
+vsperf has identified the following concerns when benchmarking the VNF to
+vSwitch to VNF deployment scenario:
+
+* The accuracy of the timing synchronization between VNFs/VMs.
+* The clock accuracy of a VNF/VM if they were to be used as traffic generators.
+* VNF traffic generator/receiver may be using resources of the system under
+ test, causing at least three forms of workload to increase as the traffic
+ load increases (generation, switching, receiving).
+
+The recommendation from vsperf is that tests for this sceanario must
+include an external HW traffic generator to act as the tester/traffic transmitter
+and receiver. The perscribed methodology to benchmark this deployment scanrio with
+an external tester involves the following three steps:
+
+#. Determine the forwarding capability and latency through the virtual interface
+connected to the VNF/VM.
+
+.. Figure:: vm2vm_virtual_interface_benchmark.png
+
+ Virtual interfaces performance benchmark
+
+#. Determine the forwarding capability and latency through the VNF/hypervisor.
+
+.. Figure:: vm2vm_hypervisor_benchmark.png
+
+ Hypervisor performance benchmark
+
+#. Determine the forwarding capability and latency for the VNF to vSwitch to VNF
+ taking the information from the previous two steps into account.
+
+.. Figure:: vm2vm_benchmark.png
+
+ VNF to vSwitch to VNF performance benchmark
+
+vsperf also identified an alternative configuration for the final step:
+
+.. Figure:: vm2vm_alternative_benchmark.png
+
+ VNF to vSwitch to VNF alternative performance benchmark
+
+.. 3.3.2:
+
+Environment/infrastructure
+============================
+Intel is providing a hosted test-bed with nine bare-metal environments
+allocated to different OPNFV projects. Currently a number of servers in
+`Intel POD 3 <https://wiki.opnfv.org/display/pharos/Intel+Pod3>`__ are
+allocated to vsperf:
+
+ * pod3-wcp-node3 and pod3-wcp-node4 which are used for CI jobs.
+ * pod3-node6 which is used as a vsperf sandbox environment.
+
+vsperf CI
+---------
+vsperf CI jobs are broken down into:
+
+ * Daily job:
+
+ * Runs everyday takes about 10 hours to complete.
+ * TESTCASES_DAILY='phy2phy_tput back2back phy2phy_tput_mod_vlan
+ phy2phy_scalability pvp_tput pvp_back2back pvvp_tput pvvp_back2back'.
+ * TESTPARAM_DAILY='--test-params TRAFFICGEN_PKT_SIZES=(64,128,512,1024,1518)'.
+
+ * Merge job:
+
+ * Runs whenever patches are merged to master.
+ * Runs a basic Sanity test.
+
+ * Verify job:
+
+ * Runs every time a patch is pushed to gerrit.
+ * Builds documentation.
+
+Scripts:
+--------
+There are 2 scripts that are part of VSPERFs CI:
+
+ * build-vsperf.sh: Lives in the VSPERF repository in the ci/ directory and is
+ used to run vsperf with the appropriate cli parameters.
+ * vswitchperf.yml: YAML description of our jenkins job. lives in the RELENG
+ repository.
+
+More info on vsperf CI can be found here:
+https://wiki.opnfv.org/display/vsperf/VSPERF+CI
+
+.. 3.3.3:
+
+Responsibilities and authority
+===============================
+The group responsible for managing, designing, preparing and executing the
+tests listed in the LTD are the vsperf committers and contributors. The vsperf
+committers and contributors should work with the relavent OPNFV projects to
+ensure that the infrastructure is in place for testing vswitches, and that the
+results are published to common end point (a results database).
+