RFC 9004




Internet Engineering Task Force (IETF)                         A. Morton
Request for Comments: 9004                                     AT&T Labs
Updates: 2544                                                   May 2021
Category: Informational
ISSN: 2070-1721


        Updates for the Back-to-Back Frame Benchmark in RFC 2544

Abstract



   Fundamental benchmarking methodologies for network interconnect
   devices of interest to the IETF are defined in RFC 2544.  This memo
   updates the procedures of the test to measure the Back-to-Back Frames
   benchmark of RFC 2544, based on further experience.

   This memo updates Section 26.4 of RFC 2544.

Status of This Memo



   This document is not an Internet Standards Track specification; it is
   published for informational purposes.

   This document is a product of the Internet Engineering Task Force
   (IETF).  It represents the consensus of the IETF community.  It has
   received public review and has been approved for publication by the
   Internet Engineering Steering Group (IESG).  Not all documents
   approved by the IESG are candidates for any level of Internet
   Standard; see Section 2 of RFC 7841.

   Information about the current status of this document, any errata,
   and how to provide feedback on it may be obtained at
   https://www.rfc-editor.org/info/rfc9004.

Copyright Notice



   Copyright (c) 2021 IETF Trust and the persons identified as the
   document authors.  All rights reserved.

   This document is subject to BCP 78 and the IETF Trust's Legal
   Provisions Relating to IETF Documents
   (https://trustee.ietf.org/license-info) in effect on the date of
   publication of this document.  Please review these documents
   carefully, as they describe your rights and restrictions with respect
   to this document.  Code Components extracted from this document must
   include Simplified BSD License text as described in Section 4.e of
   the Trust Legal Provisions and are provided without warranty as
   described in the Simplified BSD License.

Table of Contents



   1.  Introduction
   2.  Requirements Language
   3.  Scope and Goals
   4.  Motivation
   5.  Prerequisites
   6.  Back-to-Back Frames
     6.1.  Preparing the List of Frame Sizes
     6.2.  Test for a Single Frame Size
     6.3.  Test Repetition and Benchmark
     6.4.  Benchmark Calculations
   7.  Reporting
   8.  Security Considerations
   9.  IANA Considerations
   10. References
     10.1.  Normative References
     10.2.  Informative References
   Acknowledgments

   Author's Address



1.  Introduction



   The IETF's fundamental benchmarking methodologies are defined in
   [RFC2544], supported by the terms and definitions in [RFC1242].
   [RFC2544] actually obsoletes an earlier specification, [RFC1944].
   Over time, the benchmarking community has updated [RFC2544] several
   times, including the Device Reset benchmark [RFC6201] and the
   important Applicability Statement [RFC6815] concerning use outside
   the Isolated Test Environment (ITE) required for accurate
   benchmarking.  Other specifications implicitly update [RFC2544], such
   as the IPv6 benchmarking methodologies in [RFC5180].

   Recent testing experience with the Back-to-Back Frame test and
   benchmark in Section 26.4 of [RFC2544] indicates that an update is
   warranted [OPNFV-2017] [VSPERF-b2b].  In particular, analysis of the
   results indicates that buffer size matters when compensating for
   interruptions of software-packet processing, and this finding
   increases the importance of the Back-to-Back Frame characterization
   described here.  This memo provides additional rationale and the
   updated method.

   [RFC2544] provides its own requirements language consistent with
   [RFC2119], since [RFC1944] (which it obsoletes) predates [RFC2119].
   All three memos share common authorship.  Today, [RFC8174] clarifies
   the usage of requirements language, so the requirements language
   presented in this memo are expressed in accordance with [RFC8174].
   They are intended for those performing/reporting laboratory tests to
   improve clarity and repeatability, and for those designing devices
   that facilitate these tests.

2.  Requirements Language



   The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
   "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
   "OPTIONAL" in this document are to be interpreted as described in
   BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all
   capitals, as shown here.

3.  Scope and Goals



   The scope of this memo is to define an updated method to
   unambiguously perform tests, measure the benchmark(s), and report the
   results for Back-to-Back Frames (as described in Section 26.4 of
   [RFC2544]).

   The goal is to provide more efficient test procedures where possible
   and expand reporting with additional interpretation of the results.
   The tests described in this memo address the cases in which the
   maximum frame rate of a single ingress port cannot be transferred to
   an egress port without loss (for some frame sizes of interest).

   Benchmarks as described in [RFC2544] rely on test conditions with
   constant frame sizes, with the goal of understanding what network-
   device capability has been tested.  Tests with the smallest size
   stress the header-processing capacity, and tests with the largest
   size stress the overall bit-processing capacity.  Tests with sizes in
   between may determine the transition between these two capacities.
   However, conditions simultaneously sending a mixture of Internet
   (IMIX) frame sizes, such as those described in [RFC6985], MUST NOT be
   used in Back-to-Back Frame testing.

   Section 3 of [RFC8239] describes buffer-size testing for physical
   networking devices in a data center.  Those methods measure buffer
   latency directly with traffic on multiple ingress ports that overload
   an egress port on the Device Under Test (DUT) and are not subject to
   the revised calculations presented in this memo.  Likewise, the
   methods of [RFC8239] SHOULD be used for test cases where the egress-
   port buffer is the known point of overload.

4.  Motivation



   Section 3.1 of [RFC1242] describes the rationale for the Back-to-Back
   Frames benchmark.  To summarize, there are several reasons that
   devices on a network produce bursts of frames at the minimum allowed
   spacing; and it is, therefore, worthwhile to understand the DUT limit
   on the length of such bursts in practice.  The same document also
   states:

   |  Tests of this parameter are intended to determine the extent of
   |  data buffering in the device.

   Since this test was defined, there have been occasional discussions
   of the stability and repeatability of the results, both over time and
   across labs.  Fortunately, the Open Platform for Network Function
   Virtualization (OPNFV) project on Virtual Switch Performance (VSPERF)
   Continuous Integration (CI) [VSPERF-CI] testing routinely repeats
   Back-to-Back Frame tests to verify that test functionality has been
   maintained through development of the test-control programs.  These
   tests were used as a basis to evaluate stability and repeatability,
   even across lab setups when the test platform was migrated to new DUT
   hardware at the end of 2016.

   When the VSPERF CI results were examined [VSPERF-b2b], several
   aspects of the results were considered notable:

   1.  Back-to-Back Frame benchmark was very consistent for some fixed
       frame sizes, and somewhat variable for other frame sizes.

   2.  The number of Back-to-Back Frames with zero loss reported for
       large frame sizes was unexpectedly long (translating to 30
       seconds of buffer time), and no explanation or measurement limit
       condition was indicated.  It was important that the buffering
       time calculations were part of the referenced testing and
       analysis [VSPERF-b2b], because the calculated buffer time of 30
       seconds for some frame sizes was clearly wrong or highly suspect.
       On the other hand, a result expressed only as a large number of
       Back-to-Back Frames does not permit such an easy comparison with
       reality.

   3.  Calculation of the extent of buffer time in the DUT helped to
       explain the results observed with all frame sizes.  For example,
       tests with some frame sizes cannot exceed the frame-header-
       processing rate of the DUT, thus, no buffering occurs.
       Therefore, the results depended on the test equipment and not the
       DUT.

   4.  It was found that a better estimate of the DUT buffer time could
       be calculated using measurements of both the longest burst in
       frames without loss and results from the Throughput tests
       conducted according to Section 26.1 of [RFC2544].  It is apparent
       that the DUT's frame-processing rate empties the buffer during a
       trial and tends to increase the "implied" buffer-size estimate
       (measured according to Section 26.4 of [RFC2544] because many
       frames have departed the buffer when the burst of frames ends).
       A calculation using the Throughput measurement can reveal a
       "corrected" buffer-size estimate.

   Further, if the Throughput tests of Section 26.1 of [RFC2544] are
   conducted as a prerequisite, the number of frame sizes required for
   Back-to-Back Frame benchmarking can be reduced to one or more of the
   small frame sizes, or the results for large frame sizes can be noted
   as invalid in the results if tested anyway.  These are the larger
   frame sizes for which the Back-to-Back Frame rate cannot exceed the
   frame-header-processing rate of the DUT and little or no buffering
   occurs.

   The material below provides the details of the calculation to
   estimate the actual buffer storage available in the DUT, using
   results from the Throughput tests for each frame size and the Max
   Theoretical Frame Rate for the DUT links (which constrain the minimum
   frame spacing).

   In reality, there are many buffers and packet-header-processing steps
   in a typical DUT.  The simplified model used in these calculations
   for the DUT includes a packet-header-processing function with limited
   rate of operation, as shown in Figure 1.

                        |------------ DUT --------|
   Generator -> Ingress -> Buffer -> HeaderProc -> Egress -> Receiver

                 Figure 1: Simplified Model for DUT Testing

   So, in the Back-to-Back Frame testing:

   1.  The ingress burst arrives at Max Theoretical Frame Rate, and
       initially the frames are buffered.

   2.  The packet-header-processing function (HeaderProc) operates at
       the "Measured Throughput" (Section 26.1 of [RFC2544]), removing
       frames from the buffer (this is the best approximation we have,
       another acceptable approximation is the received frame rate
       during Back-to-back Frame testing, if Measured Throughput is not
       available).

   3.  Frames that have been processed are clearly not in the buffer, so
       the Corrected DUT Buffer Time equation (Section 6.4) estimates
       and removes the frames that the DUT forwarded on egress during
       the burst.  We define buffer time as the number of frames
       occupying the buffer divided by the Max Theoretical Frame Rate
       (on ingress) for the frame size under test.

   4.  A helpful concept is the buffer-filling rate, which is the
       difference between the Max Theoretical Frame Rate (ingress) and
       the Measured Throughput (HeaderProc on egress).  If the actual
       buffer size in frames is known, the time to fill the buffer
       during a measurement can be calculated using the filling rate, as
       a check on measurements.  However, the buffer in the model
       represents many buffers of different sizes in the DUT data path.

   Knowledge of approximate buffer storage size (in time or bytes) may
   be useful in estimating whether frame losses will occur if DUT
   forwarding is temporarily suspended in a production deployment due to
   an unexpected interruption of frame processing (an interruption of
   duration greater than the estimated buffer would certainly cause lost
   frames).  In Section 6, the calculations for the correct buffer time
   use the combination of offered load at Max Theoretical Frame Rate and
   header-processing speed at 100% of Measured Throughput.  Other
   combinations are possible, such as changing the percent of Measured
   Throughput to account for other processes reducing the header
   processing rate.

   The presentation of OPNFV VSPERF evaluation and development of
   enhanced search algorithms [VSPERF-BSLV] was given and discussed at
   IETF 102.  The enhancements are intended to compensate for transient
   processor interrupts that may cause loss at near-Throughput levels of
   offered load.  Subsequent analysis of the results indicates that
   buffers within the DUT can compensate for some interrupts, and this
   finding increases the importance of the Back-to-Back Frame
   characterization described here.

5.  Prerequisites



   The test setup MUST be consistent with Figure 1 of [RFC2544], or
   Figure 2 of that document when the tester's sender and receiver are
   different devices.  Other mandatory testing aspects described in
   [RFC2544] MUST be included, unless explicitly modified in the next
   section.

   The ingress and egress link speeds and link-layer protocols MUST be
   specified and used to compute the Max Theoretical Frame Rate when
   respecting the minimum interframe gap.

   The test results for the Throughput benchmark conducted according to
   Section 26.1 of [RFC2544] for all frame sizes RECOMMENDED by
   [RFC2544] MUST be available to reduce the tested-frame-size list or
   to note invalid results for individual frame sizes (because the burst
   length may be essentially infinite for large frame sizes).

   Note that:

   *  the Throughput and the Back-to-Back Frame measurement-
      configuration traffic characteristics (unidirectional or
      bidirectional, and number of flows generated) MUST match.

   *  the Throughput measurement MUST be taken under zero-loss
      conditions, according to Section 26.1 of [RFC2544].

   The Back-to-Back Benchmark described in Section 3.1 of [RFC1242] MUST
   be measured directly by the tester, where buffer size is inferred
   from Back-to-Back Frame bursts and associated packet-loss
   measurements.  Therefore, sources of frame loss that are unrelated to
   consistent evaluation of buffer size SHOULD be identified and removed
   or mitigated.  Example sources include:

   *  On-path active components that are external to the DUT

   *  Operating-system environment interrupting DUT operation

   *  Shared-resource contention between the DUT and other off-path
      component(s) impacting DUT's behavior, sometimes called the "noisy
      neighbor" problem with virtualized network functions.

   Mitigations applicable to some of the sources above are discussed in
   Section 6.2, with the other measurement requirements described below
   in Section 6.

6.  Back-to-Back Frames



   Objective: To characterize the ability of a DUT to process Back-to-
   Back Frames as defined in [RFC1242].

   The procedure follows.

6.1.  Preparing the List of Frame Sizes



   From the list of RECOMMENDED frame sizes (Section 9 of [RFC2544]),
   select the subset of frame sizes whose Measured Throughput (during
   prerequisite testing) was less than the Max Theoretical Frame Rate of
   the DUT/test setup.  These are the only frame sizes where it is
   possible to produce a burst of frames that cause the DUT buffers to
   fill and eventually overflow, producing one or more discarded frames.

6.2.  Test for a Single Frame Size



   Each trial in the test requires the tester to send a burst of frames
   (after idle time) with the minimum interframe gap and to count the
   corresponding frames forwarded by the DUT.

   The duration of the trial includes three REQUIRED components:

   1.  The time to send the burst of frames (at the back-to-back rate),
       determined by the search algorithm.

   2.  The time to receive the transferred burst of frames (at the
       [RFC2544] Throughput rate), possibly truncated by buffer
       overflow, and certainly including the latency of the DUT.

   3.  At least 2 seconds not overlapping the time to receive the burst
       (Component 2, above), to ensure that DUT buffers have depleted.
       Longer times MUST be used when conditions warrant, such as when
       buffer times >2 seconds are measured or when burst sending times
       are >2 seconds, but care is needed, since this time component
       directly increases trial duration, and many trials and tests
       comprise a complete benchmarking study.

   The upper search limit for the time to send each burst MUST be
   configurable to values as high as 30 seconds (buffer time results
   reported at or near the configured upper limit are likely invalid,
   and the test MUST be repeated with a higher search limit).

   If all frames have been received, the tester increases the length of
   the burst according to the search algorithm and performs another
   trial.

   If the received frame count is less than the number of frames in the
   burst, then the limit of DUT processing and buffering may have been
   exceeded, and the burst length for the next trial is determined by
   the search algorithm (the burst length is typically reduced, but see
   below).

   Classic search algorithms have been adapted for use in benchmarking,
   where the search requires discovery of a pair of outcomes, one with
   no loss and another with loss, at load conditions within the
   acceptable tolerance or accuracy.  Conditions encountered when
   benchmarking the infrastructure for network function virtualization
   require algorithm enhancement.  Fortunately, the adaptation of Binary
   Search, and an enhanced Binary Search with Loss Verification, have
   been specified in Clause 12.3 of [TST009].  These algorithms can
   easily be used for Back-to-Back Frame benchmarking by replacing the
   offered load level with burst length in frames.  [TST009], Annex B
   describes the theory behind the enhanced Binary Search with Loss
   Verification algorithm.

   There are also promising works in progress that may prove useful in
   Back-to-Back Frame benchmarking.  [BMWG-MLRSEARCH] and
   [BMWG-PLRSEARCH] are two such examples.

   Either the [TST009] Binary Search or Binary Search with Loss
   Verification algorithms MUST be used, and input parameters to the
   algorithm(s) MUST be reported.

   The tester usually imposes a (configurable) minimum step size for
   burst length, and the step size MUST be reported with the results (as
   this influences the accuracy and variation of test results).

   The original Section 26.4 of [RFC2544] definition is stated below:

   |  The back-to-back value is the number of frames in the longest
   |  burst that the DUT will handle without the loss of any frames.

6.3.  Test Repetition and Benchmark



   On this topic, Section 26.4 of [RFC2544] requires:

   |  The trial length MUST be at least 2 seconds and SHOULD be repeated
   |  at least 50 times with the average of the recorded values being
   |  reported.

   Therefore, the Back-to-Back Frame benchmark is the average of burst-
   length values over repeated tests to determine the longest burst of
   frames that the DUT can successfully process and buffer without frame
   loss.  Each of the repeated tests completes an independent search
   process.

   In this update, the test MUST be repeated N times (the number of
   repetitions is now a variable that must be reported) for each frame
   size in the subset list, and each Back-to-Back Frame value MUST be
   made available for further processing (below).

6.4.  Benchmark Calculations



   For each frame size, calculate the following summary statistics for
   longest Back-to-Back Frame values over the N tests:

   *  Average (Benchmark)

   *  Minimum

   *  Maximum

   *  Standard Deviation

   Further, calculate the Implied DUT Buffer Time and the Corrected DUT
   Buffer Time in seconds, as follows:

   Implied DUT buffer time =

      Average num of Back-to-back Frames / Max Theoretical Frame Rate

   The formula above is simply expressing the burst of frames in units
   of time.

   The next step is to apply a correction factor that accounts for the
   DUT's frame forwarding operation during the test (assuming the simple
   model of the DUT composed of a buffer and a forwarding function,
   described in Section 4).

   Corrected DUT Buffer Time =
                     /                                         \
      Implied DUT    |Implied DUT       Measured Throughput    |
   =  Buffer Time -  |Buffer Time * -------------------------- |
                     |              Max Theoretical Frame Rate |
                     \                                         /

   where:

   1.  The "Measured Throughput" is the [RFC2544] Throughput Benchmark
       for the frame size tested, as augmented by methods including the
       Binary Search with Loss Verification algorithm in [TST009] where
       applicable and MUST be expressed in frames per second in this
       equation.

   2.  The "Max Theoretical Frame Rate" is a calculated value for the
       interface speed and link-layer technology used, and it MUST be
       expressed in frames per second in this equation.

   The term on the far right in the formula for Corrected DUT Buffer
   Time accounts for all the frames in the burst that were transmitted
   by the DUT *while the burst of frames was sent in*.  So, these frames
   are not in the buffer, and the buffer size is more accurately
   estimated by excluding them.  If Measured Throughput is not
   available, an acceptable approximation is the received frame rate
   (see Forwarding Rate in [RFC2889] measured during Back-to-back Frame
   testing).

7.  Reporting



   The Back-to-Back Frame results SHOULD be reported in the format of a
   table with a row for each of the tested frame sizes.  There SHOULD be
   columns for the frame size and the resultant average frame count for
   each type of data stream tested.

   The number of tests averaged for the benchmark, N, MUST be reported.

   The minimum, maximum, and standard deviation across all complete
   tests SHOULD also be reported (they are referred to as
   "Min,Max,StdDev" in Table 1).

   The Corrected DUT Buffer Time SHOULD also be reported.

   If the tester operates using a limited maximum burst length in
   frames, then this maximum length SHOULD be reported.

    +=============+================+================+================+
    | Frame Size, | Ave B2B        | Min,Max,StdDev | Corrected Buff |
    | octets      | Length, frames |                | Time, Sec      |
    +=============+================+================+================+
    | 64          | 26000          | 25500,27000,20 | 0.00004        |
    +-------------+----------------+----------------+----------------+

                   Table 1: Back-to-Back Frame Results

   Static and configuration parameters (reported with Table 1):

   *  Number of test repetitions, N

   *  Minimum Step Size (during searches), in frames.


   If the tester has a specific (actual) frame rate of interest (less
   than the Throughput rate), it is useful to estimate the buffer time
   at that actual frame rate:

   Actual Buffer Time =
                                      Max Theoretical Frame Rate
        = Corrected DUT Buffer Time * --------------------------
                                          Actual Frame Rate

   and report this value, properly labeled.

8.  Security Considerations



   Benchmarking activities as described in this memo are limited to
   technology characterization using controlled stimuli in a laboratory
   environment, with dedicated address space and the other constraints
   of [RFC2544].

   The benchmarking network topology will be an independent test setup
   and MUST NOT be connected to devices that may forward the test
   traffic into a production network or misroute traffic to the test
   management network.  See [RFC6815].

   Further, benchmarking is performed on an "opaque-box" (a.k.a.
   "black-box") basis, relying solely on measurements observable
   external to the Device or System Under Test (SUT).

   The DUT developers are commonly independent from the personnel and
   institutions conducting benchmarking studies.  DUT developers might
   have incentives to alter the performance of the DUT if the test
   conditions can be detected.  Special capabilities SHOULD NOT exist in
   the DUT/SUT specifically for benchmarking purposes.  Procedures
   described in this document are not designed to detect such activity.
   Additional testing outside of the scope of this document would be
   needed and has been used successfully in the past to discover such
   malpractices.

   Any implications for network security arising from the DUT/SUT SHOULD
   be identical in the lab and in production networks.

9.  IANA Considerations



   This document has no IANA actions.

10.  References



10.1.  Normative References



   [RFC1242]  Bradner, S., "Benchmarking Terminology for Network
              Interconnection Devices", RFC 1242, DOI 10.17487/RFC1242,
              July 1991, <https://www.rfc-editor.org/info/rfc1242>.

   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119,
              DOI 10.17487/RFC2119, March 1997,
              <https://www.rfc-editor.org/info/rfc2119>.

   [RFC2544]  Bradner, S. and J. McQuaid, "Benchmarking Methodology for
              Network Interconnect Devices", RFC 2544,
              DOI 10.17487/RFC2544, March 1999,
              <https://www.rfc-editor.org/info/rfc2544>.

   [RFC6985]  Morton, A., "IMIX Genome: Specification of Variable Packet
              Sizes for Additional Testing", RFC 6985,
              DOI 10.17487/RFC6985, July 2013,
              <https://www.rfc-editor.org/info/rfc6985>.

   [RFC8174]  Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
              2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
              May 2017, <https://www.rfc-editor.org/info/rfc8174>.

   [RFC8239]  Avramov, L. and J. Rapp, "Data Center Benchmarking
              Methodology", RFC 8239, DOI 10.17487/RFC8239, August 2017,
              <https://www.rfc-editor.org/info/rfc8239>.

   [TST009]   ETSI, "Network Functions Virtualisation (NFV) Release 3;
              Testing; Specification of Networking Benchmarks and
              Measurement Methods for NFVI", Rapporteur: A. Morton, ETSI
              GS NFV-TST 009 v3.4.1, December 2020,
              <https://www.etsi.org/deliver/etsi_gs/NFV-
              TST/001_099/009/03.04.01_60/gs_NFV-TST009v030401p.pdf>.

10.2.  Informative References



   [BMWG-MLRSEARCH]
              Konstantynowicz, M., Ed. and V. Polák, Ed., "Multiple Loss
              Ratio Search for Packet Throughput (MLRsearch)", Work in
              Progress, Internet-Draft, draft-ietf-bmwg-mlrsearch-00, 9
              February 2021, <https://tools.ietf.org/html/draft-ietf-
              bmwg-mlrsearch-00>.

   [BMWG-PLRSEARCH]
              Konstantynowicz, M., Ed. and V. Polák, Ed., "Probabilistic
              Loss Ratio Search for Packet Throughput (PLRsearch)", Work
              in Progress, Internet-Draft, draft-vpolak-bmwg-plrsearch-
              03, 6 March 2020, <https://tools.ietf.org/html/draft-
              vpolak-bmwg-plrsearch-03>.

   [OPNFV-2017]
              Cooper, T., Rao, S., and A. Morton, "Dataplane
              Performance, Capacity, and Benchmarking in OPNFV", 15 June
              2017,
              <https://wiki.anuket.io/download/attachments/4404001/
              VSPERF-Dataplane-Perf-Cap-Bench.pdf?version=1&modification
              Date=1621191833500&api=v2>.

   [RFC1944]  Bradner, S. and J. McQuaid, "Benchmarking Methodology for
              Network Interconnect Devices", RFC 1944,
              DOI 10.17487/RFC1944, May 1996,
              <https://www.rfc-editor.org/info/rfc1944>.

   [RFC2889]  Mandeville, R. and J. Perser, "Benchmarking Methodology
              for LAN Switching Devices", RFC 2889,
              DOI 10.17487/RFC2889, August 2000,
              <https://www.rfc-editor.org/info/rfc2889>.

   [RFC5180]  Popoviciu, C., Hamza, A., Van de Velde, G., and D.
              Dugatkin, "IPv6 Benchmarking Methodology for Network
              Interconnect Devices", RFC 5180, DOI 10.17487/RFC5180, May
              2008, <https://www.rfc-editor.org/info/rfc5180>.

   [RFC6201]  Asati, R., Pignataro, C., Calabria, F., and C. Olvera,
              "Device Reset Characterization", RFC 6201,
              DOI 10.17487/RFC6201, March 2011,
              <https://www.rfc-editor.org/info/rfc6201>.

   [RFC6815]  Bradner, S., Dubray, K., McQuaid, J., and A. Morton,
              "Applicability Statement for RFC 2544: Use on Production
              Networks Considered Harmful", RFC 6815,
              DOI 10.17487/RFC6815, November 2012,
              <https://www.rfc-editor.org/info/rfc6815>.

   [VSPERF-b2b]
              Morton, A., "Back2Back Testing Time Series (from CI)", May
              2021, <https://wiki.anuket.io/display/HOME/
              Traffic+Generator+Testing#TrafficGeneratorTesting-
              AppendixB:Back2BackTestingTimeSeries(fromCI)>.

   [VSPERF-BSLV]
              Rao, S. and A. Morton, "Evolution of Repeatability in
              Benchmarking: Fraser Plugfest (Summary for IETF BMWG)",
              July 2018,
              <https://datatracker.ietf.org/meeting/102/materials/
              slides-102-bmwg-evolution-of-repeatability-in-
              benchmarking-fraser-plugfest-summary-for-ietf-bmwg-00>.

   [VSPERF-CI]
              Tahhan, M., "OPNFV VSPERF CI", September 2019,
              <https://wiki.anuket.io/display/HOME/VSPERF+CI>.

Acknowledgments

   Thanks to Trevor Cooper, Sridhar Rao, and Martin Klozik of the VSPERF
   project for many contributions to the early testing [VSPERF-b2b].
   Yoshiaki Itou has also investigated the topic and made useful
   suggestions.  Maciek Konstantyowicz and Vratko Polák also provided
   many comments and suggestions based on extensive integration testing
   and resulting search-algorithm proposals -- the most up-to-date
   feedback possible.  Tim Carlin also provided comments and support for
   the document.  Warren Kumari's review improved readability in several
   key passages.  David Black, Martin Duke, and Scott Bradner's comments
   improved the clarity and configuration advice on trial duration.
   Mališa Vučinić suggested additional text on DUT design cautions in
   the Security Considerations section.

Author's Address



   Al Morton
   AT&T Labs
   200 Laurel Avenue South
   Middletown, NJ 07748
   United States of America

   Phone: +1 732 420 1571
   Email: acmorton@att.com