RFC 8337






Internet Engineering Task Force (IETF)                         M. Mathis
Request for Comments: 8337                                   Google, Inc
Category: Experimental                                         A. Morton
ISSN: 2070-1721                                                AT&T Labs
                                                              March 2018


            Model-Based Metrics for Bulk Transport Capacity

Abstract



   This document introduces a new class of Model-Based Metrics designed
   to assess if a complete Internet path can be expected to meet a
   predefined Target Transport Performance by applying a suite of IP
   diagnostic tests to successive subpaths.  The subpath-at-a-time tests
   can be robustly applied to critical infrastructure, such as network
   interconnections or even individual devices, to accurately detect if
   any part of the infrastructure will prevent paths traversing it from
   meeting the Target Transport Performance.

   Model-Based Metrics rely on mathematical models to specify a Targeted
   IP Diagnostic Suite, a set of IP diagnostic tests designed to assess
   whether common transport protocols can be expected to meet a
   predetermined Target Transport Performance over an Internet path.

   For Bulk Transport Capacity, the IP diagnostics are built using test
   streams and statistical criteria for evaluating the packet transfer
   that mimic TCP over the complete path.  The temporal structure of the
   test stream (e.g., bursts) mimics TCP or other transport protocols
   carrying bulk data over a long path.  However, they are constructed
   to be independent of the details of the subpath under test, end
   systems, or applications.  Likewise, the success criteria evaluates
   the packet transfer statistics of the subpath against criteria
   determined by protocol performance models applied to the Target
   Transport Performance of the complete path.  The success criteria
   also does not depend on the details of the subpath, end systems, or
   applications.














Mathis & Morton               Experimental                      [Page 1]

RFC 8337                   Model-Based Metrics                March 2018


Status of This Memo



   This document is not an Internet Standards Track specification; it is
   published for examination, experimental implementation, and
   evaluation.

   This document defines an Experimental Protocol for the Internet
   community.  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/rfc8337.

Copyright Notice



   Copyright (c) 2018 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.



















Mathis & Morton               Experimental                      [Page 2]

RFC 8337                   Model-Based Metrics                March 2018


Table of Contents



   1. Introduction ....................................................4
   2. Overview ........................................................5
   3. Terminology .....................................................8
      3.1. General Terminology ........................................8
      3.2. Terminology about Paths ...................................10
      3.3. Properties ................................................11
      3.4. Basic Parameters ..........................................12
      3.5. Ancillary Parameters ......................................13
      3.6. Temporal Patterns for Test Streams ........................14
      3.7. Tests .....................................................15
   4. Background .....................................................16
      4.1. TCP Properties ............................................18
      4.2. Diagnostic Approach .......................................20
      4.3. New Requirements Relative to RFC 2330 .....................21
   5. Common Models and Parameters ...................................22
      5.1. Target End-to-End Parameters ..............................22
      5.2. Common Model Calculations .................................22
      5.3. Parameter Derating ........................................23
      5.4. Test Preconditions ........................................24
   6. Generating Test Streams ........................................24
      6.1. Mimicking Slowstart .......................................25
      6.2. Constant Window Pseudo CBR ................................27
      6.3. Scanned Window Pseudo CBR .................................28
      6.4. Concurrent or Channelized Testing .........................28
   7. Interpreting the Results .......................................29
      7.1. Test Outcomes .............................................29
      7.2. Statistical Criteria for Estimating run_length ............31
      7.3. Reordering Tolerance ......................................33
   8. IP Diagnostic Tests ............................................34
      8.1. Basic Data Rate and Packet Transfer Tests .................34
           8.1.1. Delivery Statistics at Paced Full Data Rate ........35
           8.1.2. Delivery Statistics at Full Data Windowed Rate .....35
           8.1.3. Background Packet Transfer Statistics Tests ........35
      8.2. Standing Queue Tests ......................................36
           8.2.1. Congestion Avoidance ...............................37
           8.2.2. Bufferbloat ........................................37
           8.2.3. Non-excessive Loss .................................38
           8.2.4. Duplex Self-Interference ...........................38
      8.3. Slowstart Tests ...........................................39
           8.3.1. Full Window Slowstart Test .........................39
           8.3.2. Slowstart AQM Test .................................39
      8.4. Sender Rate Burst Tests ...................................40
      8.5. Combined and Implicit Tests ...............................41
           8.5.1. Sustained Full-Rate Bursts Test ....................41
           8.5.2. Passive Measurements ...............................42




Mathis & Morton               Experimental                      [Page 3]

RFC 8337                   Model-Based Metrics                March 2018


   9. Example ........................................................43
      9.1. Observations about Applicability ..........................44
   10. Validation ....................................................45
   11. Security Considerations .......................................46
   12. IANA Considerations ...........................................47
   13. Informative References ........................................47
   Appendix A.  Model Derivations ....................................52
     A.1.  Queueless Reno ............................................52
   Appendix B.  The Effects of ACK Scheduling ........................53
   Acknowledgments ...................................................55
   Authors' Addresses ................................................55

1.  Introduction



   Model-Based Metrics (MBM) rely on peer-reviewed mathematical models
   to specify a Targeted IP Diagnostic Suite (TIDS), a set of IP
   diagnostic tests designed to assess whether common transport
   protocols can be expected to meet a predetermined Target Transport
   Performance over an Internet path.  This document describes the
   modeling framework to derive the test parameters for assessing an
   Internet path's ability to support a predetermined Bulk Transport
   Capacity.

   Each test in TIDS measures some aspect of IP packet transfer needed
   to meet the Target Transport Performance.  For Bulk Transport
   Capacity, the TIDS includes IP diagnostic tests to verify that there
   is sufficient IP capacity (data rate), sufficient queue space at
   bottlenecks to absorb and deliver typical transport bursts, low
   enough background packet loss ratio to not interfere with congestion
   control, and other properties described below.  Unlike typical IP
   Performance Metrics (IPPM) that yield measures of network properties,
   Model-Based Metrics nominally yield pass/fail evaluations of the
   ability of standard transport protocols to meet the specific
   performance objective over some network path.

   In most cases, the IP diagnostic tests can be implemented by
   combining existing IPPM metrics with additional controls for
   generating test streams having a specified temporal structure (bursts
   or standing queues caused by constant bit rate streams, etc.) and
   statistical criteria for evaluating packet transfer.  The temporal
   structure of the test streams mimics transport protocol behavior over
   the complete path; the statistical criteria models the transport
   protocol's response to less-than-ideal IP packet transfer.  In
   control theory terms, the tests are "open loop".  Note that running a
   test requires the coordinated activity of sending and receiving
   measurement points.





Mathis & Morton               Experimental                      [Page 4]

RFC 8337                   Model-Based Metrics                March 2018


   This document addresses Bulk Transport Capacity.  It describes an
   alternative to the approach presented in "A Framework for Defining
   Empirical Bulk Transfer Capacity Metrics" [RFC3148].  Other Model-
   Based Metrics may cover other applications and transports, such as
   Voice over IP (VoIP) over UDP, RTP, and new transport protocols.

   This document assumes a traditional Reno TCP-style, self-clocked,
   window-controlled transport protocol that uses packet loss and
   Explicit Congestion Notification (ECN) Congestion Experienced (CE)
   marks for congestion feedback.  There are currently some experimental
   protocols and congestion control algorithms that are rate based or
   otherwise fall outside of these assumptions.  In the future, these
   new protocols and algorithms may call for revised models.

   The MBM approach, i.e., mapping Target Transport Performance to a
   Targeted IP Diagnostic Suite (TIDS) of IP tests, solves some
   intrinsic problems with using TCP or other throughput-maximizing
   protocols for measurement.  In particular, all throughput-maximizing
   protocols (especially TCP congestion control) cause some level of
   congestion in order to detect when they have reached the available
   capacity limitation of the network.  This self-inflicted congestion
   obscures the network properties of interest and introduces non-linear
   dynamic equilibrium behaviors that make any resulting measurements
   useless as metrics because they have no predictive value for
   conditions or paths different from that of the measurement itself.
   In order to prevent these effects, it is necessary to avoid the
   effects of TCP congestion control in the measurement method.  These
   issues are discussed at length in Section 4.  Readers who are
   unfamiliar with basic properties of TCP and TCP-like congestion
   control may find it easier to start at Section 4 or 4.1.

   A Targeted IP Diagnostic Suite does not have such difficulties.  IP
   diagnostics can be constructed such that they make strong statistical
   statements about path properties that are independent of measurement
   details, such as vantage and choice of measurement points.

2.  Overview



   This document describes a modeling framework for deriving a Targeted
   IP Diagnostic Suite from a predetermined Target Transport
   Performance.  It is not a complete specification and relies on other
   standards documents to define important details such as packet type-P
   selection, sampling techniques, vantage selection, etc.  Fully
   Specified Targeted IP Diagnostic Suites (FSTIDSs) define all of these
   details.  A Targeted IP Diagnostic Suite (TIDS) refers to the subset
   of such a specification that is in scope for this document.  This
   terminology is further defined in Section 3.




Mathis & Morton               Experimental                      [Page 5]

RFC 8337                   Model-Based Metrics                March 2018


   Section 4 describes some key aspects of TCP behavior and what they
   imply about the requirements for IP packet transfer.  Most of the IP
   diagnostic tests needed to confirm that the path meets these
   properties can be built on existing IPPM metrics, with the addition
   of statistical criteria for evaluating packet transfer and, in a few
   cases, new mechanisms to implement the required temporal structure.
   (One group of tests, the standing queue tests described in
   Section 8.2, don't correspond to existing IPPM metrics, but suitable
   new IPPM metrics can be patterned after the existing definitions.)

   Figure 1 shows the MBM modeling and measurement framework.  The
   Target Transport Performance at the top of the figure is determined
   by the needs of the user or application, which are outside the scope
   of this document.  For Bulk Transport Capacity, the main performance
   parameter of interest is the Target Data Rate.  However, since TCP's
   ability to compensate for less-than-ideal network conditions is
   fundamentally affected by the Round-Trip Time (RTT) and the Maximum
   Transmission Unit (MTU) of the complete path, these parameters must
   also be specified in advance based on knowledge about the intended
   application setting.  They may reflect a specific application over a
   real path through the Internet or an idealized application and
   hypothetical path representing a typical user community.  Section 5
   describes the common parameters and models derived from the Target
   Transport Performance.



























Mathis & Morton               Experimental                      [Page 6]

RFC 8337                   Model-Based Metrics                March 2018


                      Target Transport Performance
            (Target Data Rate, Target RTT, and Target MTU)
                                   |
                           ________V_________
                           |  mathematical  |
                           |     models     |
                           |                |
                           ------------------
          Traffic parameters |            | Statistical criteria
                             |            |
                      _______V____________V____Targeted IP____
                     |       |   * * *    | Diagnostic Suite  |
                _____|_______V____________V________________   |
              __|____________V____________V______________  |  |
              |           IP diagnostic tests            | |  |
              |              |            |              | |  |
              | _____________V__        __V____________  | |  |
              | |   traffic    |        |   Delivery  |  | |  |
              | |   pattern    |        |  Evaluation |  | |  |
              | |  generation  |        |             |  | |  |
              | -------v--------        ------^--------  | |  |
              |   |    v    test stream via   ^      |   | |--
              |   |  -->======================>--    |   | |
              |   |       subpath under test         |   |-
              ----V----------------------------------V--- |
                  | |  |                             | |  |
                  V V  V                             V V  V
              fail/inconclusive            pass/fail/inconclusive
          (traffic generation status)           (test result)

                   Figure 1: Overall Modeling Framework

   Mathematical TCP models are used to determine traffic parameters and
   subsequently to design traffic patterns that mimic TCP (which has
   burst characteristics at multiple time scales) or other transport
   protocols delivering bulk data and operating at the Target Data Rate,
   MTU, and RTT over a full range of conditions.  Using the techniques
   described in Section 6, the traffic patterns are generated based on
   the three Target parameters of the complete path (Target Data Rate,
   Target RTT, and Target MTU), independent of the properties of
   individual subpaths.  As much as possible, the test streams are
   generated deterministically (precomputed) to minimize the extent to
   which test methodology, measurement points, measurement vantage, or
   path partitioning affect the details of the measurement traffic.

   Section 7 describes packet transfer statistics and methods to test
   against the statistical criteria provided by the mathematical models.
   Since the statistical criteria typically apply to the complete path



Mathis & Morton               Experimental                      [Page 7]

RFC 8337                   Model-Based Metrics                March 2018


   (a composition of subpaths) [RFC6049], in situ testing requires that
   the end-to-end statistical criteria be apportioned as separate
   criteria for each subpath.  Subpaths that are expected to be
   bottlenecks would then be permitted to contribute a larger fraction
   of the end-to-end packet loss budget.  In compensation, subpaths that
   are not expected to exhibit bottlenecks must be constrained to
   contribute less packet loss.  Thus, the statistical criteria for each
   subpath in each test of a TIDS is an apportioned share of the end-to-
   end statistical criteria for the complete path that was determined by
   the mathematical model.

   Section 8 describes the suite of individual tests needed to verify
   all of the required IP delivery properties.  A subpath passes if and
   only if all of the individual IP diagnostic tests pass.  Any subpath
   that fails any test indicates that some users are likely to fail to
   attain their Target Transport Performance under some conditions.  In
   addition to passing or failing, a test can be deemed inconclusive for
   a number of reasons, including the following: the precomputed traffic
   pattern was not accurately generated, the measurement results were
   not statistically significant, the test failed to meet some required
   test preconditions, etc.  If all tests pass but some are
   inconclusive, then the entire suite is deemed to be inconclusive.

   In Section 9, we present an example TIDS that might be representative
   of High Definition (HD) video and illustrate how Model-Based Metrics
   can be used to address difficult measurement situations, such as
   confirming that inter-carrier exchanges have sufficient performance
   and capacity to deliver HD video between ISPs.

   Since there is some uncertainty in the modeling process, Section 10
   describes a validation procedure to diagnose and minimize false
   positive and false negative results.

3.  Terminology



   Terms containing underscores (rather than spaces) appear in equations
   and typically have algorithmic definitions.

3.1.  General Terminology



   Target:  A general term for any parameter specified by or derived
      from the user's application or transport performance requirements.

   Target Transport Performance:  Application or transport performance
      target values for the complete path.  For Bulk Transport Capacity
      defined in this document, the Target Transport Performance
      includes the Target Data Rate, Target RTT, and Target MTU as
      described below.



Mathis & Morton               Experimental                      [Page 8]

RFC 8337                   Model-Based Metrics                March 2018


   Target Data Rate:  The specified application data rate required for
      an application's proper operation.  Conventional Bulk Transport
      Capacity (BTC) metrics are focused on the Target Data Rate;
      however, these metrics have little or no predictive value because
      they do not consider the effects of the other two parameters of
      the Target Transport Performance -- the RTT and MTU of the
      complete paths.

   Target RTT (Round-Trip Time):  The specified baseline (minimum) RTT
      of the longest complete path over which the user expects to be
      able to meet the target performance.  TCP and other transport
      protocol's ability to compensate for path problems is generally
      proportional to the number of round trips per second.  The Target
      RTT determines both key parameters of the traffic patterns (e.g.,
      burst sizes) and the thresholds on acceptable IP packet transfer
      statistics.  The Target RTT must be specified considering
      appropriate packets sizes: MTU-sized packets on the forward path
      and ACK-sized packets (typically, header_overhead) on the return
      path.  Note that Target RTT is specified and not measured; MBM
      measurements derived for a given target_RTT will be applicable to
      any path with a smaller RTT.

   Target MTU (Maximum Transmission Unit):  The specified maximum MTU
      supported by the complete path over which the application expects
      to meet the target performance.  In this document, we assume a
      1500-byte MTU unless otherwise specified.  If a subpath has a
      smaller MTU, then it becomes the Target MTU for the complete path,
      and all model calculations and subpath tests must use the same
      smaller MTU.

   Targeted IP Diagnostic Suite (TIDS):  A set of IP diagnostic tests
      designed to determine if an otherwise ideal complete path
      containing the subpath under test can sustain flows at a specific
      target_data_rate using packets with a size of target_MTU when the
      RTT of the complete path is target_RTT.

   Fully Specified Targeted IP Diagnostic Suite (FSTIDS):  A TIDS
      together with additional specifications such as measurement packet
      type ("type-p" [RFC2330]) that are out of scope for this document
      and need to be drawn from other standards documents.

   Bulk Transport Capacity (BTC):  Bulk Transport Capacity metrics
      evaluate an Internet path's ability to carry bulk data, such as
      large files, streaming (non-real-time) video, and, under some
      conditions, web images and other content.  Prior efforts to define
      BTC metrics have been based on [RFC3148], which predates our
      understanding of TCP and the requirements described in Section 4.
      In general, "Bulk Transport" indicates that performance is



Mathis & Morton               Experimental                      [Page 9]

RFC 8337                   Model-Based Metrics                March 2018


      determined by the interplay between the network, cross traffic,
      and congestion control in the transport protocol.  It excludes
      situations where performance is dominated by the RTT alone (e.g.,
      transactions) or bottlenecks elsewhere, such as in the application
      itself.

   IP diagnostic tests:  Measurements or diagnostics to determine if
      packet transfer statistics meet some precomputed target.

   traffic patterns:  The temporal patterns or burstiness of traffic
      generated by applications over transport protocols such as TCP.
      There are several mechanisms that cause bursts at various
      timescales as described in Section 4.1.  Our goal here is to mimic
      the range of common patterns (burst sizes, rates, etc.), without
      tying our applicability to specific applications, implementations,
      or technologies, which are sure to become stale.

   Explicit Congestion Notification (ECN):  See [RFC3168].

   packet transfer statistics:  Raw, detailed, or summary statistics
      about packet transfer properties of the IP layer including packet
      losses, ECN Congestion Experienced (CE) marks, reordering, or any
      other properties that may be germane to transport performance.

   packet loss ratio:  As defined in [RFC7680].

   apportioned:  To divide and allocate, for example, budgeting packet
      loss across multiple subpaths such that the losses will accumulate
      to less than a specified end-to-end loss ratio.  Apportioning
      metrics is essentially the inverse of the process described in
      [RFC5835].

   open loop:  A control theory term used to describe a class of
      techniques where systems that naturally exhibit circular
      dependencies can be analyzed by suppressing some of the
      dependencies, such that the resulting dependency graph is acyclic.

3.2.  Terminology about Paths



   See [RFC2330] and [RFC7398] for existing terms and definitions.

   data sender:  Host sending data and receiving ACKs.

   data receiver:  Host receiving data and sending ACKs.

   complete path:  The end-to-end path from the data sender to the data
      receiver.




Mathis & Morton               Experimental                     [Page 10]

RFC 8337                   Model-Based Metrics                March 2018


   subpath:  A portion of the complete path.  Note that there is no
      requirement that subpaths be non-overlapping.  A subpath can be as
      small as a single device, link, or interface.

   measurement point:  Measurement points as described in [RFC7398].

   test path:  A path between two measurement points that includes a
      subpath of the complete path under test.  If the measurement
      points are off path, the test path may include "test leads"
      between the measurement points and the subpath.

   dominant bottleneck:  The bottleneck that generally determines most
      packet transfer statistics for the entire path.  It typically
      determines a flow's self-clock timing, packet loss, and ECN CE
      marking rate, with other potential bottlenecks having less effect
      on the packet transfer statistics.  See Section 4.1 on TCP
      properties.

   front path:  The subpath from the data sender to the dominant
      bottleneck.

   back path:  The subpath from the dominant bottleneck to the receiver.

   return path:  The path taken by the ACKs from the data receiver to
      the data sender.

   cross traffic:  Other, potentially interfering, traffic competing for
      network resources (such as bandwidth and/or queue capacity).

3.3.  Properties



   The following properties are determined by the complete path and
   application.  These are described in more detail in Section 5.1.

   Application Data Rate:  General term for the data rate as seen by the
      application above the transport layer in bytes per second.  This
      is the payload data rate and explicitly excludes transport-level
      and lower-level headers (TCP/IP or other protocols),
      retransmissions, and other overhead that is not part of the total
      quantity of data delivered to the application.

   IP rate:  The actual number of IP-layer bytes delivered through a
      subpath, per unit time, including TCP and IP headers, retransmits,
      and other TCP/IP overhead.  This is the same as IP-type-P Link
      Usage in [RFC5136].






Mathis & Morton               Experimental                     [Page 11]

RFC 8337                   Model-Based Metrics                March 2018


   IP capacity:  The maximum number of IP-layer bytes that can be
      transmitted through a subpath, per unit time, including TCP and IP
      headers, retransmits, and other TCP/IP overhead.  This is the same
      as IP-type-P Link Capacity in [RFC5136].

   bottleneck IP capacity:  The IP capacity of the dominant bottleneck
      in the forward path.  All throughput-maximizing protocols estimate
      this capacity by observing the IP rate delivered through the
      bottleneck.  Most protocols derive their self-clocks from the
      timing of this data.  See Section 4.1 and Appendix B for more
      details.

   implied bottleneck IP capacity:  The bottleneck IP capacity implied
      by the ACKs returning from the receiver.  It is determined by
      looking at how much application data the ACK stream at the sender
      reports as delivered to the data receiver per unit time at various
      timescales.  If the return path is thinning, batching, or
      otherwise altering the ACK timing, the implied bottleneck IP
      capacity over short timescales might be substantially larger than
      the bottleneck IP capacity averaged over a full RTT.  Since TCP
      derives its clock from the data delivered through the bottleneck,
      the front path must have sufficient buffering to absorb any data
      bursts at the dimensions (size and IP rate) implied by the ACK
      stream, which are potentially doubled during slowstart.  If the
      return path is not altering the ACK stream, then the implied
      bottleneck IP capacity will be the same as the bottleneck IP
      capacity.  See Section 4.1 and Appendix B for more details.

   sender interface rate:  The IP rate that corresponds to the IP
      capacity of the data sender's interface.  Due to sender efficiency
      algorithms, including technologies such as TCP segmentation
      offload (TSO), nearly all modern servers deliver data in bursts at
      full interface link rate.  Today, 1 or 10 Gb/s are typical.

   header_overhead:  The IP and TCP header sizes, which are the portion
      of each MTU not available for carrying application payload.
      Without loss of generality, this is assumed to be the size for
      returning acknowledgments (ACKs).  For TCP, the Maximum Segment
      Size (MSS) is the Target MTU minus the header_overhead.

3.4.  Basic Parameters



   Basic parameters common to models and subpath tests are defined here.
   Formulas for target_window_size and target_run_length appear in
   Section 5.2.  Note that these are mixed between application transport
   performance (excludes headers) and IP performance (includes TCP
   headers and retransmissions as part of the IP payload).




Mathis & Morton               Experimental                     [Page 12]

RFC 8337                   Model-Based Metrics                March 2018


   Network power:  The observed data rate divided by the observed RTT.
      Network power indicates how effectively a transport protocol is
      filling a network.

   Window [size]:  The total quantity of data carried by packets
      in-flight plus the data represented by ACKs circulating in the
      network is referred to as the window.  See Section 4.1.  Sometimes
      used with other qualifiers (congestion window (cwnd) or receiver
      window) to indicate which mechanism is controlling the window.

   pipe size:  A general term for the number of packets needed in flight
      (the window size) to exactly fill a network path or subpath.  It
      corresponds to the window size, which maximizes network power.  It
      is often used with additional qualifiers to specify which path,
      under what conditions, etc.

   target_window_size:  The average number of packets in flight (the
      window size) needed to meet the Target Data Rate for the specified
      Target RTT and Target MTU.  It implies the scale of the bursts
      that the network might experience.

   run length:  A general term for the observed, measured, or specified
      number of packets that are (expected to be) delivered between
      losses or ECN CE marks.  Nominally, it is one over the sum of the
      loss and ECN CE marking probabilities, if they are independently
      and identically distributed.

   target_run_length:  The target_run_length is an estimate of the
      minimum number of non-congestion marked packets needed between
      losses or ECN CE marks necessary to attain the target_data_rate
      over a path with the specified target_RTT and target_MTU, as
      computed by a mathematical model of TCP congestion control.  A
      reference calculation is shown in Section 5.2 and alternatives in
      Appendix A.

   reference target_run_length:  target_run_length computed precisely by
      the method in Section 5.2.  This is likely to be slightly more
      conservative than required by modern TCP implementations.

3.5.  Ancillary Parameters



   The following ancillary parameters are used for some tests:

   derating:  Under some conditions, the standard models are too
      conservative.  The modeling framework permits some latitude in
      relaxing or "derating" some test parameters, as described in
      Section 5.3, in exchange for a more stringent TIDS validation




Mathis & Morton               Experimental                     [Page 13]

RFC 8337                   Model-Based Metrics                March 2018


      procedures, described in Section 10.  Models can be derated by
      including a multiplicative derating factor to make tests less
      stringent.

   subpath_IP_capacity:  The IP capacity of a specific subpath.

   test path:  A subpath of a complete path under test.

   test_path_RTT:  The RTT observed between two measurement points using
      packet sizes that are consistent with the transport protocol.
      This is generally MTU-sized packets of the forward path and
      packets with a size of header_overhead on the return path.

   test_path_pipe:  The pipe size of a test path.  Nominally, it is the
      test_path_RTT times the test path IP_capacity.

   test_window:  The smallest window sufficient to meet or exceed the
      target_rate when operating with a pure self-clock over a test
      path.  The test_window is typically calculated as follows (but see
      the discussion in Appendix B about the effects of channel
      scheduling on RTT):

      ceiling(target_data_rate * test_path_RTT / (target_MTU -
      header_overhead))

      On some test paths, the test_window may need to be adjusted
      slightly to compensate for the RTT being inflated by the devices
      that schedule packets.

3.6.  Temporal Patterns for Test Streams



   The terminology below is used to define temporal patterns for test
   streams.  These patterns are designed to mimic TCP behavior, as
   described in Section 4.1.

   packet headway:  Time interval between packets, specified from the
      start of one to the start of the next.  For example, if packets
      are sent with a 1 ms headway, there will be exactly 1000 packets
      per second.

   burst headway:  Time interval between bursts, specified from the
      start of the first packet of one burst to the start of the first
      packet of the next burst.  For example, if 4 packet bursts are
      sent with a 1 ms burst headway, there will be exactly 4000 packets
      per second.

   paced single packets:  Individual packets sent at the specified rate
      or packet headway.



Mathis & Morton               Experimental                     [Page 14]

RFC 8337                   Model-Based Metrics                March 2018


   paced bursts:  Bursts on a timer.  Specify any 3 of the following:
      average data rate, packet size, burst size (number of packets),
      and burst headway (burst start to start).  By default, the bursts
      are assumed to occur at full sender interface rate, such that the
      packet headway within each burst is the minimum supported by the
      sender's interface.  Under some conditions, it is useful to
      explicitly specify the packet headway within each burst.

   slowstart rate:  Paced bursts of four packets each at an average data
      rate equal to twice the implied bottleneck IP capacity (but not
      more than the sender interface rate).  This mimics TCP slowstart.
      This is a two-level burst pattern described in more detail in
      Section 6.1.  If the implied bottleneck IP capacity is more than
      half of the sender interface rate, the slowstart rate becomes the
      sender interface rate.

   slowstart burst:  A specified number of packets in a two-level burst
      pattern that resembles slowstart.  This mimics one round of TCP
      slowstart.

   repeated slowstart bursts:  Slowstart bursts repeated once per
      target_RTT.  For TCP, each burst would be twice as large as the
      prior burst, and the sequence would end at the first ECN CE mark
      or lost packet.  For measurement, all slowstart bursts would be
      the same size (nominally, target_window_size but other sizes might
      be specified), and the ECN CE marks and lost packets are counted.

3.7.  Tests



   The tests described in this document can be grouped according to
   their applicability.

   Capacity tests:  Capacity tests determine if a network subpath has
      sufficient capacity to deliver the Target Transport Performance.
      As long as the test stream is within the proper envelope for the
      Target Transport Performance, the average packet losses or ECN CE
      marks must be below the statistical criteria computed by the
      model.  As such, capacity tests reflect parameters that can
      transition from passing to failing as a consequence of cross
      traffic, additional presented load, or the actions of other
      network users.  By definition, capacity tests also consume
      significant network resources (data capacity and/or queue buffer
      space), and the test schedules must be balanced by their cost.

   Monitoring tests:  Monitoring tests are designed to capture the most
      important aspects of a capacity test without presenting excessive
      ongoing load themselves.  As such, they may miss some details of




Mathis & Morton               Experimental                     [Page 15]

RFC 8337                   Model-Based Metrics                March 2018


      the network's performance but can serve as a useful reduced-cost
      proxy for a capacity test, for example, to support continuous
      production network monitoring.

   Engineering tests:  Engineering tests evaluate how network algorithms
      (such as Active Queue Management (AQM) and channel allocation)
      interact with TCP-style self-clocked protocols and adaptive
      congestion control based on packet loss and ECN CE marks.  These
      tests are likely to have complicated interactions with cross
      traffic and, under some conditions, can be inversely sensitive to
      load.  For example, a test to verify that an AQM algorithm causes
      ECN CE marks or packet drops early enough to limit queue occupancy
      may experience a false pass result in the presence of cross
      traffic.  It is important that engineering tests be performed
      under a wide range of conditions, including both in situ and bench
      testing, and over a wide variety of load conditions.  Ongoing
      monitoring is less likely to be useful for engineering tests,
      although sparse in situ testing might be appropriate.

4.  Background



   When "Framework for IP Performance Metrics" [RFC2330] was published
   in 1998, sound Bulk Transport Capacity (BTC) measurement was known to
   be well beyond our capabilities.  Even when "A Framework for Defining
   Empirical Bulk Transfer Capacity Metrics" [RFC3148] was published, we
   knew that we didn't really understand the problem.  Now, in
   hindsight, we understand why assessing BTC is such a difficult
   problem:

   o  TCP is a control system with circular dependencies -- everything
      affects performance, including components that are explicitly not
      part of the test (for example, the host processing power is not
      in-scope of path performance tests).

   o  Congestion control is a dynamic equilibrium process, similar to
      processes observed in chemistry and other fields.  The network and
      transport protocols find an operating point that balances opposing
      forces: the transport protocol pushing harder (raising the data
      rate and/or window) while the network pushes back (raising packet
      loss ratio, RTT, and/or ECN CE marks).  By design, TCP congestion
      control keeps raising the data rate until the network gives some
      indication that its capacity has been exceeded by dropping packets
      or adding ECN CE marks.  If a TCP sender accurately fills a path
      to its IP capacity (e.g., the bottleneck is 100% utilized), then
      packet losses and ECN CE marks are mostly determined by the TCP
      sender and how aggressively it seeks additional capacity; they are
      not determined by the network itself, because the network must
      send exactly the signals that TCP needs to set its rate.



Mathis & Morton               Experimental                     [Page 16]

RFC 8337                   Model-Based Metrics                March 2018


   o  TCP's ability to compensate for network impairments (such as loss,
      delay, and delay variation, outside of those caused by TCP itself)
      is directly proportional to the number of send-ACK round-trip
      exchanges per second (i.e., inversely proportional to the RTT).
      As a consequence, an impaired subpath may pass a short RTT local
      test even though it fails when the subpath is extended by an
      effectively perfect network to some larger RTT.

   o  TCP has an extreme form of the Observer Effect (colloquially known
      as the "Heisenberg Effect").  Measurement and cross traffic
      interact in unknown and ill-defined ways.  The situation is
      actually worse than the traditional physics problem where you can
      at least estimate bounds on the relative momentum of the
      measurement and measured particles.  In general, for network
      measurement, you cannot determine even the order of magnitude of
      the effect.  It is possible to construct measurement scenarios
      where the measurement traffic starves real user traffic, yielding
      an overly inflated measurement.  The inverse is also possible: the
      user traffic can fill the network, such that the measurement
      traffic detects only minimal available capacity.  In general, you
      cannot determine which scenario might be in effect, so you cannot
      gauge the relative magnitude of the uncertainty introduced by
      interactions with other network traffic.

   o  As a consequence of the properties listed above, it is difficult,
      if not impossible, for two independent implementations (hardware
      or software) of TCP congestion control to produce equivalent
      performance results [RFC6576] under the same network conditions.

   These properties are a consequence of the dynamic equilibrium
   behavior intrinsic to how all throughput-maximizing protocols
   interact with the Internet.  These protocols rely on control systems
   based on estimated network metrics to regulate the quantity of data
   to send into the network.  The packet-sending characteristics in turn
   alter the network properties estimated by the control system metrics,
   such that there are circular dependencies between every transmission
   characteristic and every estimated metric.  Since some of these
   dependencies are nonlinear, the entire system is nonlinear, and any
   change anywhere causes a difficult-to-predict response in network
   metrics.  As a consequence, Bulk Transport Capacity metrics have not
   fulfilled the analytic framework envisioned in [RFC2330].

   Model-Based Metrics overcome these problems by making the measurement
   system open loop: the packet transfer statistics (akin to the network
   estimators) do not affect the traffic or traffic patterns (bursts),
   which are computed on the basis of the Target Transport Performance.
   A path or subpath meeting the Target Transfer Performance




Mathis & Morton               Experimental                     [Page 17]

RFC 8337                   Model-Based Metrics                March 2018


   requirements would exhibit packet transfer statistics and estimated
   metrics that would not cause the control system to slow the traffic
   below the Target Data Rate.

4.1.  TCP Properties



   TCP and other self-clocked protocols (e.g., the Stream Control
   Transmission Protocol (SCTP)) carry the vast majority of all Internet
   data.  Their dominant bulk data transport behavior is to have an
   approximately fixed quantity of data and acknowledgments (ACKs)
   circulating in the network.  The data receiver reports arriving data
   by returning ACKs to the data sender, and the data sender typically
   responds by sending approximately the same quantity of data back into
   the network.  The total quantity of data plus the data represented by
   ACKs circulating in the network is referred to as the "window".  The
   mandatory congestion control algorithms incrementally adjust the
   window by sending slightly more or less data in response to each ACK.
   The fundamentally important property of this system is that it is
   self-clocked: the data transmissions are a reflection of the ACKs
   that were delivered by the network, and the ACKs are a reflection of
   the data arriving from the network.

   A number of protocol features cause bursts of data, even in idealized
   networks that can be modeled as simple queuing systems.

   During slowstart, the IP rate is doubled on each RTT by sending twice
   as much data as was delivered to the receiver during the prior RTT.
   Each returning ACK causes the sender to transmit twice the data the
   ACK reported arriving at the receiver.  For slowstart to be able to
   fill the pipe, the network must be able to tolerate slowstart bursts
   up to the full pipe size inflated by the anticipated window reduction
   on the first loss or ECN CE mark.  For example, with classic Reno
   congestion control, an optimal slowstart has to end with a burst that
   is twice the bottleneck rate for one RTT in duration.  This burst
   causes a queue that is equal to the pipe size (i.e., the window is
   twice the pipe size), so when the window is halved in response to the
   first packet loss, the new window will be the pipe size.

   Note that if the bottleneck IP rate is less than half of the capacity
   of the front path (which is almost always the case), the slowstart
   bursts will not by themselves cause significant queues anywhere else
   along the front path; they primarily exercise the queue at the
   dominant bottleneck.

   Several common efficiency algorithms also cause bursts.  The self-
   clock is typically applied to groups of packets: the receiver's
   delayed ACK algorithm generally sends only one ACK per two data
   segments.  Furthermore, modern senders use TCP segmentation offload



Mathis & Morton               Experimental                     [Page 18]

RFC 8337                   Model-Based Metrics                March 2018


   (TSO) to reduce CPU overhead.  The sender's software stack builds
   super-sized TCP segments that the TSO hardware splits into MTU-sized
   segments on the wire.  The net effect of TSO, delayed ACK, and other
   efficiency algorithms is to send bursts of segments at full sender
   interface rate.

   Note that these efficiency algorithms are almost always in effect,
   including during slowstart, such that slowstart typically has a two-
   level burst structure.  Section 6.1 describes slowstart in more
   detail.

   Additional sources of bursts include TCP's initial window [RFC6928],
   application pauses, channel allocation mechanisms, and network
   devices that schedule ACKs.  Appendix B describes these last two
   items.  If the application pauses (e.g., stops reading or writing
   data) for some fraction of an RTT, many TCP implementations catch up
   to their earlier window size by sending a burst of data at the full
   sender interface rate.  To fill a network with a realistic
   application, the network has to be able to tolerate sender interface
   rate bursts large enough to restore the prior window following
   application pauses.

   Although the sender interface rate bursts are typically smaller than
   the last burst of a slowstart, they are at a higher IP rate so they
   potentially exercise queues at arbitrary points along the front path
   from the data sender up to and including the queue at the dominant
   bottleneck.  It is known that these bursts can hurt network
   performance, especially in conjunction with other queue pressure;
   however, we are not aware of any models for estimating the impact or
   prescribing limits on the size or frequency of sender rate bursts.

   In conclusion, to verify that a path can meet a Target Transport
   Performance, it is necessary to independently confirm that the path
   can tolerate bursts at the scales that can be caused by the above
   mechanisms.  Three cases are believed to be sufficient:

   o  Two-level slowstart bursts sufficient to get connections started
      properly.

   o  Ubiquitous sender interface rate bursts caused by efficiency
      algorithms.  We assume four packet bursts to be the most common
      case, since it matches the effects of delayed ACK during
      slowstart.  These bursts should be assumed not to significantly
      affect packet transfer statistics.







Mathis & Morton               Experimental                     [Page 19]

RFC 8337                   Model-Based Metrics                March 2018


   o  Infrequent sender interface rate bursts that are the maximum of
      the full target_window_size and the initial window size (10
      segments in [RFC6928]).  The target_run_length may be derated for
      these large fast bursts.

   If a subpath can meet the required packet loss ratio for bursts at
   all of these scales, then it has sufficient buffering at all
   potential bottlenecks to tolerate any of the bursts that are likely
   introduced by TCP or other transport protocols.

4.2.  Diagnostic Approach



   A complete path is expected to be able to attain a specified Bulk
   Transport Capacity if the path's RTT is equal to or smaller than the
   Target RTT, the path's MTU is equal to or larger than the Target MTU,
   and all of the following conditions are met:

   1.  The IP capacity is above the Target Data Rate by a sufficient
       margin to cover all TCP/IP overheads.  This can be confirmed by
       the tests described in Section 8.1 or any number of IP capacity
       tests adapted to implement MBM.

   2.  The observed packet transfer statistics are better than required
       by a suitable TCP performance model (e.g., fewer packet losses or
       ECN CE marks).  See Section 8.1 or any number of low- or fixed-
       rate packet loss tests outside of MBM.

   3.  There is sufficient buffering at the dominant bottleneck to
       absorb a slowstart burst large enough to get the flow out of
       slowstart at a suitable window size.  See Section 8.3.

   4.  There is sufficient buffering in the front path to absorb and
       smooth sender interface rate bursts at all scales that are likely
       to be generated by the application, any channel arbitration in
       the ACK path, or any other mechanisms.  See Section 8.4.

   5.  When there is a slowly rising standing queue at the bottleneck,
       then the onset of packet loss has to be at an appropriate point
       (in time or in queue depth) and has to be progressive, for
       example, by use of Active Queue Management [RFC7567].  See
       Section 8.2.

   6.  When there is a standing queue at a bottleneck for a shared media
       subpath (e.g., a half-duplex link), there must be a suitable
       bound on the interaction between ACKs and data, for example, due
       to the channel arbitration mechanism.  See Section 8.2.4.





Mathis & Morton               Experimental                     [Page 20]

RFC 8337                   Model-Based Metrics                March 2018


   Note that conditions 1 through 4 require capacity tests for
   validation and thus may need to be monitored on an ongoing basis.
   Conditions 5 and 6 require engineering tests, which are best
   performed in controlled environments (e.g., bench tests).  They won't
   generally fail due to load but may fail in the field (e.g., due to
   configuration errors, etc.) and thus should be spot checked.

   A tool that can perform many of the tests is available from
   [MBMSource].

4.3.  New Requirements Relative to RFC 2330



   Model-Based Metrics are designed to fulfill some additional
   requirements that were not recognized at the time RFC 2330 [RFC2330]
   was published.  These missing requirements may have significantly
   contributed to policy difficulties in the IP measurement space.  Some
   additional requirements are:

   o  IP metrics must be actionable by the ISP -- they have to be
      interpreted in terms of behaviors or properties at the IP or lower
      layers that an ISP can test, repair, and verify.

   o  Metrics should be spatially composable, such that measures of
      concatenated paths should be predictable from subpaths.

   o  Metrics must be vantage point invariant over a significant range
      of measurement point choices, including off-path measurement
      points.  The only requirements for Measurement Point (MP)
      selection should be that the RTT between the MPs is below some
      reasonable bound and that the effects of the "test leads"
      connecting MPs to the subpath under test can be calibrated out of
      the measurements.  The latter might be accomplished if the test
      leads are effectively ideal or their properties can be deducted
      from the measurements between the MPs.  While many tests require
      that the test leads have at least as much IP capacity as the
      subpath under test, some do not, for example, the Background
      Packet Transfer Statistics Tests described in Section 8.1.3.

   o  Metric measurements should be repeatable by multiple parties with
      no specialized access to MPs or diagnostic infrastructure.  It
      should be possible for different parties to make the same
      measurement and observe the same results.  In particular, it is
      important that both a consumer (or the consumer's delegate) and
      ISP be able to perform the same measurement and get the same
      result.  Note that vantage independence is key to meeting this
      requirement.





Mathis & Morton               Experimental                     [Page 21]

RFC 8337                   Model-Based Metrics                March 2018


5.  Common Models and Parameters



5.1.  Target End-to-End Parameters



   The target end-to-end parameters are the Target Data Rate, Target
   RTT, and Target MTU as defined in Section 3.  These parameters are
   determined by the needs of the application or the ultimate end user
   and the complete Internet path over which the application is expected
   to operate.  The target parameters are in units that make sense to
   layers above the TCP layer: payload bytes delivered to the
   application.  They exclude overheads associated with TCP and IP
   headers, retransmits and other protocols (e.g., DNS).  Note that
   IP-based network services include TCP headers and retransmissions as
   part of delivered payload; this difference (header_overhead) is
   recognized in calculations below.

   Other end-to-end parameters defined in Section 3 include the
   effective bottleneck data rate, the sender interface data rate, and
   the TCP and IP header sizes.

   The target_data_rate must be smaller than all subpath IP capacities
   by enough headroom to carry the transport protocol overhead,
   explicitly including retransmissions and an allowance for
   fluctuations in TCP's actual data rate.  Specifying a
   target_data_rate with insufficient headroom is likely to result in
   brittle measurements that have little predictive value.

   Note that the target parameters can be specified for a hypothetical
   path (for example, to construct TIDS designed for bench testing in
   the absence of a real application) or for a live in situ test of
   production infrastructure.

   The number of concurrent connections is explicitly not a parameter in
   this model.  If a subpath requires multiple connections in order to
   meet the specified performance, that must be stated explicitly, and
   the procedure described in Section 6.4 applies.

5.2.  Common Model Calculations



   The Target Transport Performance is used to derive the
   target_window_size and the reference target_run_length.

   The target_window_size is the average window size in packets needed
   to meet the target_rate, for the specified target_RTT and target_MTU.
   To calculate target_window_size:

   target_window_size = ceiling(target_rate * target_RTT / (target_MTU -
   header_overhead))



Mathis & Morton               Experimental                     [Page 22]

RFC 8337                   Model-Based Metrics                March 2018


   The target_run_length is an estimate of the minimum required number
   of unmarked packets that must be delivered between losses or ECN CE
   marks, as computed by a mathematical model of TCP congestion control.
   The derivation here is parallel to the derivation in [MSMO97] and, by
   design, is quite conservative.

   The reference target_run_length is derived as follows.  Assume the
   subpath_IP_capacity is infinitesimally larger than the
   target_data_rate plus the required header_overhead.  Then,
   target_window_size also predicts the onset of queuing.  A larger
   window will cause a standing queue at the bottleneck.

   Assume the transport protocol is using standard Reno-style Additive
   Increase Multiplicative Decrease (AIMD) congestion control [RFC5681]
   (but not Appropriate Byte Counting [RFC3465]) and the receiver is
   using standard delayed ACKs.  Reno increases the window by one packet
   every pipe size worth of ACKs.  With delayed ACKs, this takes two
   RTTs per increase.  To exactly fill the pipe, the spacing of losses
   must be no closer than when the peak of the AIMD sawtooth reached
   exactly twice the target_window_size.  Otherwise, the multiplicative
   window reduction triggered by the loss would cause the network to be
   underfilled.  Per [MSMO97] the number of packets between losses must
   be the area under the AIMD sawtooth.  They must be no more frequent
   than every 1 in ((3/2)*target_window_size)*(2*target_window_size)
   packets, which simplifies to:

   target_run_length = 3*(target_window_size^2)

   Note that this calculation is very conservative and is based on a
   number of assumptions that may not apply.  Appendix A discusses these
   assumptions and provides some alternative models.  If a different
   model is used, an FSTIDS must document the actual method for
   computing target_run_length and the ratio between alternate
   target_run_length and the reference target_run_length calculated
   above, along with a discussion of the rationale for the underlying
   assumptions.

   Most of the individual parameters for the tests in Section 8 are
   derived from target_window_size and target_run_length.

5.3.  Parameter Derating



   Since some aspects of the models are very conservative, the MBM
   framework permits some latitude in derating test parameters.  Rather
   than trying to formalize more complicated models, we permit some test
   parameters to be relaxed as long as they meet some additional
   procedural constraints:




Mathis & Morton               Experimental                     [Page 23]

RFC 8337                   Model-Based Metrics                March 2018


   o  The FSTIDS must document and justify the actual method used to
      compute the derated metric parameters.

   o  The validation procedures described in Section 10 must be used to
      demonstrate the feasibility of meeting the Target Transport
      Performance with infrastructure that just barely passes the
      derated tests.

   o  The validation process for an FSTIDS itself must be documented in
      such a way that other researchers can duplicate the validation
      experiments.

   Except as noted, all tests below assume no derating.  Tests for which
   there is not currently a well-established model for the required
   parameters explicitly include derating as a way to indicate
   flexibility in the parameters.

5.4.  Test Preconditions



   Many tests have preconditions that are required to assure their
   validity.  Examples include the presence or non-presence of cross
   traffic on specific subpaths; negotiating ECN; and a test stream
   preamble of appropriate length to achieve stable access to network
   resources in the presence of reactive network elements (as defined in
   Section 1.1 of [RFC7312]).  If preconditions are not properly
   satisfied for some reason, the tests should be considered to be
   inconclusive.  In general, it is useful to preserve diagnostic
   information as to why the preconditions were not met and any test
   data that was collected even if it is not useful for the intended
   test.  Such diagnostic information and partial test data may be
   useful for improving the test or test procedures themselves.

   It is important to preserve the record that a test was scheduled;
   otherwise, precondition enforcement mechanisms can introduce sampling
   bias.  For example, canceling tests due to cross traffic on
   subscriber access links might introduce sampling bias in tests of the
   rest of the network by reducing the number of tests during peak
   network load.

   Test preconditions and failure actions must be specified in an
   FSTIDS.

6.  Generating Test Streams



   Many important properties of Model-Based Metrics, such as vantage
   independence, are a consequence of using test streams that have
   temporal structures that mimic TCP or other transport protocols
   running over a complete path.  As described in Section 4.1, self-



Mathis & Morton               Experimental                     [Page 24]

RFC 8337                   Model-Based Metrics                March 2018


   clocked protocols naturally have burst structures related to the RTT
   and pipe size of the complete path.  These bursts naturally get
   larger (contain more packets) as either the Target RTT or Target Data
   Rate get larger or the Target MTU gets smaller.  An implication of
   these relationships is that test streams generated by running self-
   clocked protocols over short subpaths may not adequately exercise the
   queuing at any bottleneck to determine if the subpath can support the
   full Target Transport Performance over the complete path.

   Failing to authentically mimic TCP's temporal structure is part of
   the reason why simple performance tools such as iPerf, netperf, nc,
   etc., have the reputation for yielding false pass results over short
   test paths, even when a subpath has a flaw.

   The definitions in Section 3 are sufficient for most test streams.
   We describe the slowstart and standing queue test streams in more
   detail.

   In conventional measurement practice, stochastic processes are used
   to eliminate many unintended correlations and sample biases.
   However, MBM tests are designed to explicitly mimic temporal
   correlations caused by network or protocol elements themselves.  Some
   portions of these systems, such as traffic arrival (e.g., test
   scheduling), are naturally stochastic.  Other behaviors, such as
   back-to-back packet transmissions, are dominated by implementation-
   specific deterministic effects.  Although these behaviors always
   contain non-deterministic elements and might be modeled
   stochastically, these details typically do not contribute
   significantly to the overall system behavior.  Furthermore, it is
   known that real protocols are subject to failures caused by network
   property estimators suffering from bias due to correlation in their
   own traffic.  For example, TCP's RTT estimator used to determine the
   Retransmit Timeout (RTO), can be fooled by periodic cross traffic or
   start-stop applications.  For these reasons, many details of the test
   streams are specified deterministically.

   It may prove useful to introduce fine-grained noise sources into the
   models used for generating test streams in an update of Model-Based
   Metrics, but the complexity is not warranted at the time this
   document was written.

6.1.  Mimicking Slowstart



   TCP slowstart has a two-level burst structure as shown in Figure 2.
   The fine time structure is caused by efficiency algorithms that
   deliberately batch work (CPU, channel allocation, etc.) to better
   amortize certain network and host overheads.  ACKs passing through
   the return path typically cause the sender to transmit small bursts



Mathis & Morton               Experimental                     [Page 25]

RFC 8337                   Model-Based Metrics                March 2018


   of data at the full sender interface rate.  For example, TCP
   Segmentation Offload (TSO) and Delayed Acknowledgment both contribute
   to this effect.  During slowstart, these bursts are at the same
   headway as the returning ACKs but are typically twice as large (e.g.,
   have twice as much data) as the ACK reported was delivered to the
   receiver.  Due to variations in delayed ACK and algorithms such as
   Appropriate Byte Counting [RFC3465], different pairs of senders and
   receivers produce slightly different burst patterns.  Without loss of
   generality, we assume each ACK causes four packet sender interface
   rate bursts at an average headway equal to the ACK headway; this
   corresponds to sending at an average rate equal to twice the
   effective bottleneck IP rate.  Each slowstart burst consists of a
   series of four packet sender interface rate bursts such that the
   total number of packets is the current window size (as of the last
   packet in the burst).

   The coarse time structure is due to each RTT being a reflection of
   the prior RTT.  For real transport protocols, each slowstart burst is
   twice as large (twice the window) as the previous burst but is spread
   out in time by the network bottleneck, such that each successive RTT
   exhibits the same effective bottleneck IP rate.  The slowstart phase
   ends on the first lost packet or ECN mark, which is intended to
   happen after successive slowstart bursts merge in time: the next
   burst starts before the bottleneck queue is fully drained and the
   prior burst is complete.

   For the diagnostic tests described below, we preserve the fine time
   structure but manipulate the coarse structure of the slowstart bursts
   (burst size and headway) to measure the ability of the dominant
   bottleneck to absorb and smooth slowstart bursts.

   Note that a stream of repeated slowstart bursts has three different
   average rates, depending on the averaging time interval.  At the
   finest timescale (a few packet times at the sender interface), the
   peak of the average IP rate is the same as the sender interface rate;
   at a medium timescale (a few ACK times at the dominant bottleneck),
   the peak of the average IP rate is twice the implied bottleneck IP
   capacity; and at timescales longer than the target_RTT and when the
   burst size is equal to the target_window_size, the average rate is
   equal to the target_data_rate.  This pattern corresponds to repeating
   the last RTT of TCP slowstart when delayed ACK and sender-side byte
   counting are present but without the limits specified in Appropriate
   Byte Counting [RFC3465].








Mathis & Morton               Experimental                     [Page 26]

RFC 8337                   Model-Based Metrics                March 2018


   time ==>    ( - equals one packet)

   Fine time structure of the packet stream:

   ----  ----  ----  ----  ----

   |<>| sender interface rate bursts (typically 3 or 4 packets)
   |<===>| burst headway (from the ACK headway)

   \____repeating sender______/
          rate bursts

   Coarse (RTT-level) time structure of the packet stream:

   ----  ----  ----  ----  ----                     ----  ---- ...

   |<========================>| slowstart burst size (from the window)
   |<==============================================>| slowstart headway
                                                       (from the RTT)
   \__________________________/                     \_________ ...
       one slowstart burst                     Repeated slowstart bursts

               Figure 2: Multiple Levels of Slowstart Bursts

6.2.  Constant Window Pseudo CBR



   Pseudo constant bit rate (CBR) is implemented by running a standard
   self-clocked protocol such as TCP with a fixed window size.  If that
   window size is test_window, the data rate will be slightly above the
   target_rate.

   Since the test_window is constrained to be an integer number of
   packets, for small RTTs or low data rates, there may not be
   sufficiently precise control over the data rate.  Rounding the
   test_window up (as defined above) is likely to result in data rates
   that are higher than the target rate, but reducing the window by one
   packet may result in data rates that are too small.  Also, cross
   traffic potentially raises the RTT, implicitly reducing the rate.
   Cross traffic that raises the RTT nearly always makes the test more
   strenuous (i.e., more demanding for the network path).

   Note that Constant Window Pseudo CBR (and Scanned Window Pseudo CBR
   in the next section) both rely on a self-clock that is at least
   partially derived from the properties of the subnet under test.  This
   introduces the possibility that the subnet under test exhibits
   behaviors such as extreme RTT fluctuations that prevent these
   algorithms from accurately controlling data rates.




Mathis & Morton               Experimental                     [Page 27]

RFC 8337                   Model-Based Metrics                March 2018


   An FSTIDS specifying a Constant Window Pseudo CBR test must
   explicitly indicate under what conditions errors in the data rate
   cause tests to be inconclusive.  Conventional paced measurement
   traffic may be more appropriate for these environments.

6.3.  Scanned Window Pseudo CBR



   Scanned Window Pseudo CBR is similar to the Constant Window Pseudo
   CBR described above, except the window is scanned across a range of
   sizes designed to include two key events: the onset of queuing and
   the onset of packet loss or ECN CE marks.  The window is scanned by
   incrementing it by one packet every 2*target_window_size delivered
   packets.  This mimics the additive increase phase of standard Reno
   TCP congestion avoidance when delayed ACKs are in effect.  Normally,
   the window increases are separated by intervals slightly longer than
   twice the target_RTT.

   There are two ways to implement this test: 1) applying a window clamp
   to standard congestion control in a standard protocol such as TCP and
   2) stiffening a non-standard transport protocol.  When standard
   congestion control is in effect, any losses or ECN CE marks cause the
   transport to revert to a window smaller than the clamp, such that the
   scanning clamp loses control of the window size.  The NPAD (Network
   Path and Application Diagnostics) pathdiag tool is an example of this
   class of algorithms [Pathdiag].

   Alternatively, a non-standard congestion control algorithm can
   respond to losses by transmitting extra data, such that it maintains
   the specified window size independent of losses or ECN CE marks.
   Such a stiffened transport explicitly violates mandatory Internet
   congestion control [RFC5681] and is not suitable for in situ testing.
   It is only appropriate for engineering testing under laboratory
   conditions.  The Windowed Ping tool implements such a test [WPING].
   This tool has been updated (see [mpingSource]).

   The test procedures in Section 8.2 describe how to the partition the
   scans into regions and how to interpret the results.

6.4.  Concurrent or Channelized Testing



   The procedures described in this document are only directly
   applicable to single-stream measurement, e.g., one TCP connection or
   measurement stream.  In an ideal world, we would disallow all
   performance claims based on multiple concurrent streams, but this is
   not practical due to at least two issues.  First, many very high-rate
   link technologies are channelized and at last partially pin the flow-
   to-channel mapping to minimize packet reordering within flows.




Mathis & Morton               Experimental                     [Page 28]

RFC 8337                   Model-Based Metrics                March 2018


   Second, TCP itself has scaling limits.  Although the former problem
   might be overcome through different design decisions, the latter
   problem is more deeply rooted.

   All congestion control algorithms that are philosophically aligned
   with [RFC5681] (e.g., claim some level of TCP compatibility,
   friendliness, or fairness) have scaling limits; that is, as a long
   fat network (LFN) with a fixed RTT and MTU gets faster, these
   congestion control algorithms get less accurate and, as a
   consequence, have difficulty filling the network [CCscaling].  These
   properties are a consequence of the original Reno AIMD congestion
   control design and the requirement in [RFC5681] that all transport
   protocols have similar responses to congestion.

   There are a number of reasons to want to specify performance in terms
   of multiple concurrent flows; however, this approach is not
   recommended for data rates below several megabits per second, which
   can be attained with run lengths under 10000 packets on many paths.
   Since the required run length is proportional to the square of the
   data rate, at higher rates, the run lengths can be unreasonably
   large, and multiple flows might be the only feasible approach.

   If multiple flows are deemed necessary to meet aggregate performance
   targets, then this must be stated both in the design of the TIDS and
   in any claims about network performance.  The IP diagnostic tests
   must be performed concurrently with the specified number of
   connections.  For the tests that use bursty test streams, the bursts
   should be synchronized across streams unless there is a priori
   knowledge that the applications have some explicit mechanism to
   stagger their own bursts.  In the absence of an explicit mechanism to
   stagger bursts, many network and application artifacts will sometimes
   implicitly synchronize bursts.  A test that does not control burst
   synchronization may be prone to false pass results for some
   applications.

7.  Interpreting the Results



7.1.  Test Outcomes



   To perform an exhaustive test of a complete network path, each test
   of the TIDS is applied to each subpath of the complete path.  If any
   subpath fails any test, then a standard transport protocol running
   over the complete path can also be expected to fail to attain the
   Target Transport Performance under some conditions.

   In addition to passing or failing, a test can be deemed to be
   inconclusive for a number of reasons.  Proper instrumentation and
   treatment of inconclusive outcomes is critical to the accuracy and



Mathis & Morton               Experimental                     [Page 29]

RFC 8337                   Model-Based Metrics                March 2018


   robustness of Model-Based Metrics.  Tests can be inconclusive if the
   precomputed traffic pattern or data rates were not accurately
   generated; the measurement results were not statistically
   significant; the required preconditions for the test were not met; or
   other causes.  See Section 5.4.

   For example, consider a test that implements Constant Window Pseudo
   CBR (Section 6.2) by adding rate controls and detailed IP packet
   transfer instrumentation to TCP (e.g., using the extended performance
   statistics for TCP as described in [RFC4898]).  TCP includes built-in
   control systems that might interfere with the sending data rate.  If
   such a test meets the required packet transfer statistics (e.g., run
   length) while failing to attain the specified data rate, it must be
   treated as an inconclusive result, because we cannot a priori
   determine if the reduced data rate was caused by a TCP problem or a
   network problem or if the reduced data rate had a material effect on
   the observed packet transfer statistics.

   Note that for capacity tests, if the observed packet transfer
   statistics meet the statistical criteria for failing (based on
   acceptance of hypothesis H1 in Section 7.2), the test can be
   considered to have failed because it doesn't really matter that the
   test didn't attain the required data rate.

   The important new properties of MBM, such as vantage independence,
   are a direct consequence of opening the control loops in the
   protocols, such that the test stream does not depend on network
   conditions or IP packets received.  Any mechanism that introduces
   feedback between the path's measurements and the test stream
   generation is at risk of introducing nonlinearities that spoil these
   properties.  Any exceptional event that indicates that such feedback
   has happened should cause the test to be considered inconclusive.

   Inconclusive tests may be caused by situations in which a test
   outcome is ambiguous because of network limitations or an unknown
   limitation on the IP diagnostic test itself, which may have been
   caused by some uncontrolled feedback from the network.

   Note that procedures that attempt to search the target parameter
   space to find the limits on a parameter such as target_data_rate are
   at risk of breaking the location-independent properties of Model-
   Based Metrics if any part of the boundary between passing,
   inconclusive, or failing results is sensitive to RTT (which is
   normally the case).  For example, the maximum data rate for a
   marginal link (e.g., exhibiting excess errors) is likely to be
   sensitive to the test_path_RTT.  The maximum observed data rate over
   the test path has very little value for predicting the maximum rate
   over a different path.



Mathis & Morton               Experimental                     [Page 30]

RFC 8337                   Model-Based Metrics                March 2018


   One of the goals for evolving TIDS designs will be to keep sharpening
   the distinctions between inconclusive, passing, and failing tests.
   The criteria for inconclusive, passing, and failing tests must be
   explicitly stated for every test in the TIDS or FSTIDS.

   One of the goals for evolving the testing process, procedures, tools,
   and measurement point selection should be to minimize the number of
   inconclusive tests.

   It may be useful to keep raw packet transfer statistics and ancillary
   metrics [RFC3148] for deeper study of the behavior of the network
   path and to measure the tools themselves.  Raw packet transfer
   statistics can help to drive tool evolution.  Under some conditions,
   it might be possible to re-evaluate the raw data for satisfying
   alternate Target Transport Performance.  However, it is important to
   guard against sampling bias and other implicit feedback that can
   cause false results and exhibit measurement point vantage
   sensitivity.  Simply applying different delivery criteria based on a
   different Target Transport Performance is insufficient if the test
   traffic patterns (bursts, etc.) do not match the alternate Target
   Transport Performance.

7.2.  Statistical Criteria for Estimating run_length



   When evaluating the observed run_length, we need to determine
   appropriate packet stream sizes and acceptable error levels for
   efficient measurement.  In practice, can we compare the empirically
   estimated packet loss and ECN CE marking ratios with the targets as
   the sample size grows?  How large a sample is needed to say that the
   measurements of packet transfer indicate a particular run length is
   present?

   The generalized measurement can be described as recursive testing:
   send packets (individually or in patterns) and observe the packet
   transfer performance (packet loss ratio, other metric, or any marking
   we define).

   As each packet is sent and measured, we have an ongoing estimate of
   the performance in terms of the ratio of packet loss or ECN CE marks
   to total packets (i.e., an empirical probability).  We continue to
   send until conditions support a conclusion or a maximum sending limit
   has been reached.

   We have a target_mark_probability, one mark per target_run_length,
   where a "mark" is defined as a lost packet, a packet with ECN CE
   mark, or other signal.  This constitutes the null hypothesis:





Mathis & Morton               Experimental                     [Page 31]

RFC 8337                   Model-Based Metrics                March 2018


   H0:  no more than one mark in target_run_length =
      3*(target_window_size)^2 packets

   We can stop sending packets if ongoing measurements support accepting
   H0 with the specified Type I error = alpha (= 0.05, for example).

   We also have an alternative hypothesis to evaluate: is performance
   significantly lower than the target_mark_probability?  Based on
   analysis of typical values and practical limits on measurement
   duration, we choose four times the H0 probability:

   H1:  one or more marks in (target_run_length/4) packets

   and we can stop sending packets if measurements support rejecting H0
   with the specified Type II error = beta (= 0.05, for example), thus
   preferring the alternate hypothesis H1.

   H0 and H1 constitute the success and failure outcomes described
   elsewhere in this document; while the ongoing measurements do not
   support either hypothesis, the current status of measurements is
   inconclusive.

   The problem above is formulated to match the Sequential Probability
   Ratio Test (SPRT) [Wald45] [Montgomery90].  Note that as originally
   framed, the events under consideration were all manufacturing
   defects.  In networking, ECN CE marks and lost packets are not
   defects but signals, indicating that the transport protocol should
   slow down.

   The Sequential Probability Ratio Test also starts with a pair of
   hypotheses specified as above:

   H0:  p0 = one defect in target_run_length

   H1:  p1 = one defect in target_run_length/4

   As packets are sent and measurements collected, the tester evaluates
   the cumulative defect count against two boundaries representing H0
   Acceptance or Rejection (and acceptance of H1):

   Acceptance line:  Xa = -h1 + s*n

   Rejection line:  Xr = h2 + s*n

   where n increases linearly for each packet sent and






Mathis & Morton               Experimental                     [Page 32]

RFC 8337                   Model-Based Metrics                March 2018


   h1 =  { log((1-alpha)/beta) }/k

   h2 =  { log((1-beta)/alpha) }/k

   k  =  log{ (p1(1-p0)) / (p0(1-p1)) }

   s  =  [ log{ (1-p0)/(1-p1) } ]/k

   for p0 and p1 as defined in the null and alternative hypotheses
   statements above, and alpha and beta as the Type I and Type II
   errors.

   The SPRT specifies simple stopping rules:

   o  Xa < defect_count(n) < Xr: continue testing

   o  defect_count(n) <= Xa: Accept H0

   o  defect_count(n) >= Xr: Accept H1

   The calculations above are implemented in the R-tool for Statistical
   Analysis [Rtool], in the add-on package for Cross-Validation via
   Sequential Testing (CVST) [CVST].

   Using the equations above, we can calculate the minimum number of
   packets (n) needed to accept H0 when x defects are observed.  For
   example, when x = 0:

   Xa = 0  = -h1 + s*n

   and  n = h1 / s

   Note that the derivations in [Wald45] and [Montgomery90] differ.
   Montgomery's simplified derivation of SPRT may assume a Bernoulli
   processes, where the packet loss probabilities are independent and
   identically distributed, making the SPRT more accessible.  Wald's
   seminal paper showed that this assumption is not necessary.  It helps
   to remember that the goal of SPRT is not to estimate the value of the
   packet loss rate but only whether or not the packet loss ratio is
   likely (1) low enough (when we accept the H0 null hypothesis),
   yielding success or (2) too high (when we accept the H1 alternate
   hypothesis), yielding failure.

7.3.  Reordering Tolerance



   All tests must be instrumented for packet-level reordering [RFC4737].
   However, there is no consensus for how much reordering should be
   acceptable.  Over the last two decades, the general trend has been to



Mathis & Morton               Experimental                     [Page 33]

RFC 8337                   Model-Based Metrics                March 2018


   make protocols and applications more tolerant to reordering (for
   example, see [RFC5827]), in response to the gradual increase in
   reordering in the network.  This increase has been due to the
   deployment of technologies such as multithreaded routing lookups and
   Equal-Cost Multipath (ECMP) routing.  These techniques increase
   parallelism in the network and are critical to enabling overall
   Internet growth to exceed Moore's Law.

   With transport retransmission strategies, there are fundamental
   trade-offs among reordering tolerance, how quickly losses can be
   repaired, and overhead from spurious retransmissions.  In advance of
   new retransmission strategies, we propose the following strawman:
   transport protocols should be able to adapt to reordering as long as
   the reordering extent is not more than the maximum of one quarter
   window or 1 ms, whichever is larger.  (These values come from
   experience prototyping Early Retransmit [RFC5827] and related
   algorithms.  They agree with the values being proposed for "RACK: a
   time-based fast loss detection algorithm" [RACK].)  Within this limit
   on reorder extent, there should be no bound on reordering density.

   By implication, recording that is less than these bounds should not
   be treated as a network impairment.  However, [RFC4737] still
   applies: reordering should be instrumented, and the maximum
   reordering that can be properly characterized by the test (because of
   the bound on history buffers) should be recorded with the measurement
   results.

   Reordering tolerance and diagnostic limitations, such as the size of
   the history buffer used to diagnose packets that are way out of
   order, must be specified in an FSTIDS.

8.  IP Diagnostic Tests



   The IP diagnostic tests below are organized according to the
   technique used to generate the test stream as described in Section 6.
   All of the results are evaluated in accordance with Section 7,
   possibly with additional test-specific criteria.

   We also introduce some combined tests that are more efficient when
   networks are expected to pass but conflate diagnostic signatures when
   they fail.

8.1.  Basic Data Rate and Packet Transfer Tests



   We propose several versions of the basic data rate and packet
   transfer statistics test that differ in how the data rate is
   controlled.  The data can be paced on a timer or window controlled
   (and self-clocked).  The first two tests implicitly confirm that



Mathis & Morton               Experimental                     [Page 34]

RFC 8337                   Model-Based Metrics                March 2018


   sub_path has sufficient raw capacity to carry the target_data_rate.
   They are recommended for relatively infrequent testing, such as an
   installation or periodic auditing process.  The third test,
   Background Packet Transfer Statistics, is a low-rate test designed
   for ongoing monitoring for changes in subpath quality.

8.1.1.  Delivery Statistics at Paced Full Data Rate



   This test confirms that the observed run length is at least the
   target_run_length while relying on timer to send data at the
   target_rate using the procedure described in Section 6.1 with a burst
   size of 1 (single packets) or 2 (packet pairs).

   The test is considered to be inconclusive if the packet transmission
   cannot be accurately controlled for any reason.

   RFC 6673 [RFC6673] is appropriate for measuring packet transfer
   statistics at full data rate.

8.1.2.  Delivery Statistics at Full Data Windowed Rate



   This test confirms that the observed run length is at least the
   target_run_length while sending at an average rate approximately
   equal to the target_data_rate, by controlling (or clamping) the
   window size of a conventional transport protocol to test_window.

   Since losses and ECN CE marks cause transport protocols to reduce
   their data rates, this test is expected to be less precise about
   controlling its data rate.  It should not be considered inconclusive
   as long as at least some of the round trips reached the full
   target_data_rate without incurring losses or ECN CE marks.  To pass
   this test, the network must deliver target_window_size packets in
   target_RTT time without any losses or ECN CE marks at least once per
   two target_window_size round trips, in addition to meeting the run
   length statistical test.

8.1.3.  Background Packet Transfer Statistics Tests



   The Background Packet Transfer Statistics Test is a low-rate version
   of the target rate test above, designed for ongoing lightweight
   monitoring for changes in the observed subpath run length without
   disrupting users.  It should be used in conjunction with one of the
   above full-rate tests because it does not confirm that the subpath
   can support raw data rate.

   RFC 6673 [RFC6673] is appropriate for measuring background packet
   transfer statistics.




Mathis & Morton               Experimental                     [Page 35]

RFC 8337                   Model-Based Metrics                March 2018


8.2.  Standing Queue Tests



   These engineering tests confirm that the bottleneck is well behaved
   across the onset of packet loss, which typically follows after the
   onset of queuing.  Well behaved generally means lossless for
   transient queues, but once the queue has been sustained for a
   sufficient period of time (or reaches a sufficient queue depth),
   there should be a small number of losses or ECN CE marks to signal to
   the transport protocol that it should reduce its window or data rate.
   Losses that are too early can prevent the transport from averaging at
   the target_data_rate.  Losses that are too late indicate that the
   queue might not have an appropriate AQM [RFC7567] and, as a
   consequence, be subject to bufferbloat [wikiBloat].  Queues without
   AQM have the potential to inflict excess delays on all flows sharing
   the bottleneck.  Excess losses (more than half of the window) at the
   onset of loss make loss recovery problematic for the transport
   protocol.  Non-linear, erratic, or excessive RTT increases suggest
   poor interactions between the channel acquisition algorithms and the
   transport self-clock.  All of the tests in this section use the same
   basic scanning algorithm, described here, but score the link or
   subpath on the basis of how well it avoids each of these problems.

   Some network technologies rely on virtual queues or other techniques
   to meter traffic without adding any queuing delay, in which case the
   data rate will vary with the window size all the way up to the onset
   of load-induced packet loss or ECN CE marks.  For these technologies,
   the discussion of queuing in Section 6.3 does not apply, but it is
   still necessary to confirm that the onset of losses or ECN CE marks
   be at an appropriate point and progressive.  If the network
   bottleneck does not introduce significant queuing delay, modify the
   procedure described in Section 6.3 to start the scan at a window
   equal to or slightly smaller than the test_window.

   Use the procedure in Section 6.3 to sweep the window across the onset
   of queuing and the onset of loss.  The tests below all assume that
   the scan emulates standard additive increase and delayed ACK by
   incrementing the window by one packet for every 2*target_window_size
   packets delivered.  A scan can typically be divided into three
   regions: below the onset of queuing, a standing queue, and at or
   beyond the onset of loss.

   Below the onset of queuing, the RTT is typically fairly constant, and
   the data rate varies in proportion to the window size.  Once the data
   rate reaches the subpath IP rate, the data rate becomes fairly
   constant, and the RTT increases in proportion to the increase in
   window size.  The precise transition across the start of queuing can
   be identified by the maximum network power, defined to be the ratio




Mathis & Morton               Experimental                     [Page 36]

RFC 8337                   Model-Based Metrics                March 2018


   data rate over the RTT.  The network power can be computed at each
   window size, and the window with the maximum is taken as the start of
   the queuing region.

   If there is random background loss (e.g., bit errors), precise
   determination of the onset of queue-induced packet loss may require
   multiple scans.  At window sizes large enough to cause loss in
   queues, all transport protocols are expected to experience periodic
   losses determined by the interaction between the congestion control
   and AQM algorithms.  For standard congestion control algorithms, the
   periodic losses are likely to be relatively widely spaced, and the
   details are typically dominated by the behavior of the transport
   protocol itself.  For the case of stiffened transport protocols (with
   non-standard, aggressive congestion control algorithms), the details
   of periodic losses will be dominated by how the window increase
   function responds to loss.

8.2.1.  Congestion Avoidance



   A subpath passes the congestion avoidance standing queue test if more
   than target_run_length packets are delivered between the onset of
   queuing (as determined by the window with the maximum network power
   as described above) and the first loss or ECN CE mark.  If this test
   is implemented using a standard congestion control algorithm with a
   clamp, it can be performed in situ in the production internet as a
   capacity test.  For an example of such a test, see [Pathdiag].

   For technologies that do not have conventional queues, use the
   test_window in place of the onset of queuing.  That is, a subpath
   passes the congestion avoidance standing queue test if more than
   target_run_length packets are delivered between the start of the scan
   at test_window and the first loss or ECN CE mark.

8.2.2.  Bufferbloat



   This test confirms that there is some mechanism to limit buffer
   occupancy (e.g., that prevents bufferbloat).  Note that this is not
   strictly a requirement for single-stream bulk transport capacity;
   however, if there is no mechanism to limit buffer queue occupancy,
   then a single stream with sufficient data to deliver is likely to
   cause the problems described in [RFC7567] and [wikiBloat].  This may
   cause only minor symptoms for the dominant flow but has the potential
   to make the subpath unusable for other flows and applications.

   The test will pass if the onset of loss occurs before a standing
   queue has introduced delay greater than twice the target_RTT or
   another well-defined and specified limit.  Note that there is not yet
   a model for how much standing queue is acceptable.  The factor of two



Mathis & Morton               Experimental                     [Page 37]

RFC 8337                   Model-Based Metrics                March 2018


   chosen here reflects a rule of thumb.  In conjunction with the
   previous test, this test implies that the first loss should occur at
   a queuing delay that is between one and two times the target_RTT.

   Specified RTT limits that are larger than twice the target_RTT must
   be fully justified in the FSTIDS.

8.2.3.  Non-excessive Loss



   This test confirms that the onset of loss is not excessive.  The test
   will pass if losses are equal to or less than the increase in the
   cross traffic plus the test stream window increase since the previous
   RTT.  This could be restated as non-decreasing total throughput of
   the subpath at the onset of loss.  (Note that when there is a
   transient drop in subpath throughput and there is not already a
   standing queue, a subpath that passes other queue tests in this
   document will have sufficient queue space to hold one full RTT worth
   of data).

   Note that token bucket policers will not pass this test, which is as
   intended.  TCP often stumbles badly if more than a small fraction of
   the packets are dropped in one RTT.  Many TCP implementations will
   require a timeout and slowstart to recover their self-clock.  Even if
   they can recover from the massive losses, the sudden change in
   available capacity at the bottleneck wastes serving and front-path
   capacity until TCP can adapt to the new rate [Policing].

8.2.4.  Duplex Self-Interference



   This engineering test confirms a bound on the interactions between
   the forward data path and the ACK return path when they share a half-
   duplex link.

   Some historical half-duplex technologies had the property that each
   direction held the channel until it completely drained its queue.
   When a self-clocked transport protocol, such as TCP, has data and
   ACKs passing in opposite directions through such a link, the behavior
   often reverts to stop-and-wait.  Each additional packet added to the
   window raises the observed RTT by two packet times, once as the
   additional packet passes through the data path and once for the
   additional delay incurred by the ACK waiting on the return path.

   The Duplex Self-Interference Test fails if the RTT rises by more than
   a fixed bound above the expected queuing time computed from the
   excess window divided by the subpath IP capacity.  This bound must be
   smaller than target_RTT/2 to avoid reverting to stop-and-wait
   behavior (e.g., data packets and ACKs both have to be released at
   least twice per RTT).



Mathis & Morton               Experimental                     [Page 38]

RFC 8337                   Model-Based Metrics                March 2018


8.3.  Slowstart Tests



   These tests mimic slowstart: data is sent at twice the effective
   bottleneck rate to exercise the queue at the dominant bottleneck.

8.3.1.  Full Window Slowstart Test



   This capacity test confirms that slowstart is not likely to exit
   prematurely.  To perform this test, send slowstart bursts that are
   target_window_size total packets and accumulate packet transfer
   statistics as described in Section 7.2 to score the outcome.  The
   test will pass if it is statistically significant that the observed
   number of good packets delivered between losses or ECN CE marks is
   larger than the target_run_length.  The test will fail if it is
   statistically significant that the observed interval between losses
   or ECN CE marks is smaller than the target_run_length.

   The test is deemed inconclusive if the elapsed time to send the data
   burst is not less than half of the time to receive the ACKs.  (That
   is, it is acceptable to send data too fast, but sending it slower
   than twice the actual bottleneck rate as indicated by the ACKs is
   deemed inconclusive).  The headway for the slowstart bursts should be
   the target_RTT.

   Note that these are the same parameters that are used for the
   Sustained Full-Rate Bursts Test, except the burst rate is at
   slowstart rate rather than sender interface rate.

8.3.2.  Slowstart AQM Test



   To perform this test, do a continuous slowstart (send data
   continuously at twice the implied IP bottleneck capacity) until the
   first loss; stop and allow the network to drain and repeat; gather
   statistics on how many packets were delivered before the loss, the
   pattern of losses, maximum observed RTT, and window size; and justify
   the results.  There is not currently sufficient theory to justify
   requiring any particular result; however, design decisions that
   affect the outcome of this tests also affect how the network balances
   between long and short flows (the "mice vs. elephants" problem).  The
   queue sojourn time for the first packet delivered after the first
   loss should be at least one half of the target_RTT.

   This engineering test should be performed on a quiescent network or
   testbed, since cross traffic has the potential to change the results
   in ill-defined ways.






Mathis & Morton               Experimental                     [Page 39]

RFC 8337                   Model-Based Metrics                March 2018


8.4.  Sender Rate Burst Tests



   These tests determine how well the network can deliver bursts sent at
   the sender's interface rate.  Note that this test most heavily
   exercises the front path and is likely to include infrastructure that
   may be out of scope for an access ISP, even though the bursts might
   be caused by ACK compression, thinning, or channel arbitration in the
   access ISP.  See Appendix B.

   Also, there are a several details about sender interface rate bursts
   that are not fully defined here.  These details, such as the assumed
   sender interface rate, should be explicitly stated in an FSTIDS.

   Current standards permit TCP to send full window bursts following an
   application pause.  (Congestion Window Validation [RFC2861] and
   updates to support Rate-Limited Traffic [RFC7661] are not required).
   Since full window bursts are consistent with standard behavior, it is
   desirable that the network be able to deliver such bursts; otherwise,
   application pauses will cause unwarranted losses.  Note that the AIMD
   sawtooth requires a peak window that is twice target_window_size, so
   the worst-case burst may be 2*target_window_size.

   It is also understood in the application and serving community that
   interface rate bursts have a cost to the network that has to be
   balanced against other costs in the servers themselves.  For example,
   TCP Segmentation Offload (TSO) reduces server CPU in exchange for
   larger network bursts, which increase the stress on network buffer
   memory.  Some newer TCP implementations can pace traffic at scale
   [TSO_pacing] [TSO_fq_pacing].  It remains to be determined if and how
   quickly these changes will be deployed.

   There is not yet theory to unify these costs or to provide a
   framework for trying to optimize global efficiency.  We do not yet
   have a model for how many server rate bursts should be tolerated by
   the network.  Some bursts must be tolerated by the network, but it is
   probably unreasonable to expect the network to be able to efficiently
   deliver all data as a series of bursts.

   For this reason, this is the only test for which we encourage
   derating.  A TIDS could include a table containing pairs of derating
   parameters: burst sizes and how much each burst size is permitted to
   reduce the run length, relative to the target_run_length.









Mathis & Morton               Experimental                     [Page 40]

RFC 8337                   Model-Based Metrics                March 2018


8.5.  Combined and Implicit Tests



   Combined tests efficiently confirm multiple network properties in a
   single test, possibly as a side effect of normal content delivery.
   They require less measurement traffic than other testing strategies
   at the cost of conflating diagnostic signatures when they fail.
   These are by far the most efficient for monitoring networks that are
   nominally expected to pass all tests.

8.5.1.  Sustained Full-Rate Bursts Test



   The Sustained Full-Rate Bursts Test implements a combined worst-case
   version of all of the capacity tests above.  To perform this test,
   send target_window_size bursts of packets at server interface rate
   with target_RTT burst headway (burst start to next burst start), and
   verify that the observed packet transfer statistics meets the
   target_run_length.

   Key observations:

   o  The subpath under test is expected to go idle for some fraction of
      the time, determined by the difference between the time to drain
      the queue at the subpath_IP_capacity and the target_RTT.  If the
      queue does not drain completely, it may be an indication that the
      subpath has insufficient IP capacity or that there is some other
      problem with the test (e.g., it is inconclusive).

   o  The burst sensitivity can be derated by sending smaller bursts
      more frequently (e.g., by sending target_window_size*derate packet
      bursts every target_RTT*derate, where "derate" is less than one).

   o  When not derated, this test is the most strenuous capacity test.

   o  A subpath that passes this test is likely to be able to sustain
      higher rates (close to subpath_IP_capacity) for paths with RTTs
      significantly smaller than the target_RTT.

   o  This test can be implemented with instrumented TCP [RFC4898],
      using a specialized measurement application at one end (e.g.,
      [MBMSource]) and a minimal service at the other end (e.g.,
      [RFC863] and [RFC864]).

   o  This test is efficient to implement, since it does not require
      per-packet timers, and can make use of TSO in modern network
      interfaces.






Mathis & Morton               Experimental                     [Page 41]

RFC 8337                   Model-Based Metrics                March 2018


   o  If a subpath is known to pass the standing queue engineering tests
      (particularly that it has a progressive onset of loss at an
      appropriate queue depth), then the Sustained Full-Rate Bursts Test
      is sufficient to assure that the subpath under test will not
      impair Bulk Transport Capacity at the target performance under all
      conditions.  See Section 8.2 for a discussion of the standing
      queue tests.

   Note that this test is clearly independent of the subpath RTT or
   other details of the measurement infrastructure, as long as the
   measurement infrastructure can accurately and reliably deliver the
   required bursts to the subpath under test.

8.5.2.  Passive Measurements



   Any non-throughput-maximizing application, such as fixed-rate
   streaming media, can be used to implement passive or hybrid (defined
   in [RFC7799]) versions of Model-Based Metrics with some additional
   instrumentation and possibly a traffic shaper or other controls in
   the servers.  The essential requirement is that the data transmission
   be constrained such that even with arbitrary application pauses and
   bursts, the data rate and burst sizes stay within the envelope
   defined by the individual tests described above.

   If the application's serving data rate can be constrained to be less
   than or equal to the target_data_rate and the serving_RTT (the RTT
   between the sender and client) is less than the target_RTT, this
   constraint is most easily implemented by clamping the transport
   window size to serving_window_clamp (which is set to the test_window
   and computed for the actual serving path).

   Under the above constraints, the serving_window_clamp will limit both
   the serving data rate and burst sizes to be no larger than the
   parameters specified by the procedures in Section 8.1.2, 8.4, or
   8.5.1.  Since the serving RTT is smaller than the target_RTT, the
   worst-case bursts that might be generated under these conditions will
   be smaller than called for by Section 8.4, and the sender rate burst
   sizes are implicitly derated by the serving_window_clamp divided by
   the target_window_size at the very least.  (Depending on the
   application behavior, the data might be significantly smoother than
   specified by any of the burst tests.)

   In an alternative implementation, the data rate and bursts might be
   explicitly controlled by a programmable traffic shaper or by pacing
   at the sender.  This would provide better control over transmissions
   but is more complicated to implement, although the required
   technology is available [TSO_pacing] [TSO_fq_pacing].




Mathis & Morton               Experimental                     [Page 42]

RFC 8337                   Model-Based Metrics                March 2018


   Note that these techniques can be applied to any content delivery
   that can be operated at a constrained data rate to inhibit TCP
   equilibrium behavior.

   Furthermore, note that Dynamic Adaptive Streaming over HTTP (DASH) is
   generally in conflict with passive Model-Based Metrics measurement,
   because it is a rate-maximizing protocol.  It can still meet the
   requirement here if the rate can be capped, for example, by knowing a
   priori the maximum rate needed to deliver a particular piece of
   content.

9.  Example



   In this section, we illustrate a TIDS designed to confirm that an
   access ISP can reliably deliver HD video from multiple content
   providers to all of its customers.  With modern codecs, minimal HD
   video (720p) generally fits in 2.5 Mb/s.  Due to the ISP's
   geographical size, network topology, and modem characteristics, the
   ISP determines that most content is within a 50 ms RTT of its users.
   (This example RTT is sufficient to cover the propagation delay to
   continental Europe or to either coast of the United States with low-
   delay modems; it is sufficient to cover somewhat smaller geographical
   regions if the modems require additional delay to implement advanced
   compression and error recovery.)

                +----------------------+-------+---------+
                | End-to-End Parameter | value | units   |
                +----------------------+-------+---------+
                | target_rate          | 2.5   | Mb/s    |
                | target_RTT           | 50    | ms      |
                | target_MTU           | 1500  | bytes   |
                | header_overhead      | 64    | bytes   |
                |                      |       |         |
                | target_window_size   | 11    | packets |
                | target_run_length    | 363   | packets |
                +----------------------+-------+---------+

                    Table 1: 2.5 Mb/s over a 50 ms Path

   Table 1 shows the default TCP model with no derating and, as such, is
   quite conservative.  The simplest TIDS would be to use the Sustained
   Full-Rate Bursts Test, described in Section 8.5.1.  Such a test would
   send 11 packet bursts every 50 ms and confirm that there was no more
   than 1 packet loss per 33 bursts (363 total packets in 1.650
   seconds).






Mathis & Morton               Experimental                     [Page 43]

RFC 8337                   Model-Based Metrics                March 2018


   Since this number represents the entire end-to-end loss budget,
   independent subpath tests could be implemented by apportioning the
   packet loss ratio across subpaths.  For example, 50% of the losses
   might be allocated to the access or last mile link to the user, 40%
   to the network interconnections with other ISPs, and 1% to each
   internal hop (assuming no more than 10 internal hops).  Then, all of
   the subpaths can be tested independently, and the spatial composition
   of passing subpaths would be expected to be within the end-to-end
   loss budget.

9.1.  Observations about Applicability



   Guidance on deploying and using MBM belong in a future document.
   However, the example above illustrates some of the issues that may
   need to be considered.

   Note that another ISP, with different geographical coverage,
   topology, or modem technology may need to assume a different
   target_RTT and, as a consequence, a different target_window_size and
   target_run_length, even for the same target_data rate.  One of the
   implications of this is that infrastructure shared by multiple ISPs,
   such as Internet Exchange Points (IXPs) and other interconnects may
   need to be evaluated on the basis of the most stringent
   target_window_size and target_run_length of any participating ISP.
   One way to do this might be to choose target parameters for
   evaluating such shared infrastructure on the basis of a hypothetical
   reference path that does not necessarily match any actual paths.

   Testing interconnects has generally been problematic: conventional
   performance tests run between measurement points adjacent to either
   side of the interconnect are not generally useful.  Unconstrained TCP
   tests, such as iPerf [iPerf], are usually overly aggressive due to
   the small RTT (often less than 1 ms).  With a short RTT, these tools
   are likely to report inflated data rates because on a short RTT,
   these tools can tolerate very high packet loss ratios and can push
   other cross traffic off of the network.  As a consequence, these
   measurements are useless for predicting actual user performance over
   longer paths and may themselves be quite disruptive.  Model-Based
   Metrics solves this problem.  The interconnect can be evaluated with
   the same TIDS as other subpaths.  Continuing our example, if the
   interconnect is apportioned 40% of the losses, 11 packet bursts sent
   every 50 ms should have fewer than one loss per 82 bursts (902
   packets).








Mathis & Morton               Experimental                     [Page 44]

RFC 8337                   Model-Based Metrics                March 2018


10.  Validation



   Since some aspects of the models are likely to be too conservative,
   Section 5.2 permits alternate protocol models, and Section 5.3
   permits test parameter derating.  If either of these techniques is
   used, we require demonstrations that such a TIDS can robustly detect
   subpaths that will prevent authentic applications using state-of-the-
   art protocol implementations from meeting the specified Target
   Transport Performance.  This correctness criteria is potentially
   difficult to prove, because it implicitly requires validating a TIDS
   against all possible paths and subpaths.  The procedures described
   here are still experimental.

   We suggest two approaches, both of which should be applied.  First,
   publish a fully open description of the TIDS, including what
   assumptions were used and how it was derived, such that the research
   community can evaluate the design decisions, test them, and comment
   on their applicability.  Second, demonstrate that applications do
   meet the Target Transport Performance when running over a network
   testbed that has the tightest possible constraints that still allow
   the tests in the TIDS to pass.

   This procedure resembles an epsilon-delta proof in calculus.
   Construct a test network such that all of the individual tests of the
   TIDS pass by only small (infinitesimal) margins, and demonstrate that
   a variety of authentic applications running over real TCP
   implementations (or other protocols as appropriate) meets the Target
   Transport Performance over such a network.  The workloads should
   include multiple types of streaming media and transaction-oriented
   short flows (e.g., synthetic web traffic).

   For example, for the HD streaming video TIDS described in Section 9,
   the IP capacity should be exactly the header_overhead above 2.5 Mb/s,
   the per packet random background loss ratio should be 1/363 (for a
   run length of 363 packets), the bottleneck queue should be 11
   packets, and the front path should have just enough buffering to
   withstand 11 packet interface rate bursts.  We want every one of the
   TIDS tests to fail if we slightly increase the relevant test
   parameter, so, for example, sending a 12-packet burst should cause
   excess (possibly deterministic) packet drops at the dominant queue at
   the bottleneck.  This network has the tightest possible constraints
   that can be expected to pass the TIDS, yet it should be possible for
   a real application using a stock TCP implementation in the vendor's
   default configuration to attain 2.5 Mb/s over a 50 ms path.

   The most difficult part of setting up such a testbed is arranging for
   it to have the tightest possible constraints that still allow it to
   pass the individual tests.  Two approaches are suggested:



Mathis & Morton               Experimental                     [Page 45]

RFC 8337                   Model-Based Metrics                March 2018


   o  constraining (configuring) the network devices not to use all
      available resources (e.g., by limiting available buffer space or
      data rate)

   o  pre-loading subpaths with cross traffic

   Note that it is important that a single tightly constrained
   environment just barely passes all tests; otherwise, there is a
   chance that TCP can exploit extra latitude in some parameters (such
   as data rate) to partially compensate for constraints in other
   parameters (e.g., queue space).  This effect is potentially
   bidirectional: extra latitude in the queue space tests has the
   potential to enable TCP to compensate for insufficient data-rate
   headroom.

   To the extent that a TIDS is used to inform public dialog, it should
   be fully documented publicly, including the details of the tests,
   what assumptions were used, and how it was derived.  All of the
   details of the validation experiment should also be published with
   sufficient detail for the experiments to be replicated by other
   researchers.  All components should be either open source or fully
   described proprietary implementations that are available to the
   research community.

11.  Security Considerations



   Measurement is often used to inform business and policy decisions
   and, as a consequence, is potentially subject to manipulation.
   Model-Based Metrics are expected to be a huge step forward because
   equivalent measurements can be performed from multiple vantage
   points, such that performance claims can be independently validated
   by multiple parties.

   Much of the acrimony in the Net Neutrality debate is due to the
   historical lack of any effective vantage-independent tools to
   characterize network performance.  Traditional methods for measuring
   Bulk Transport Capacity are sensitive to RTT and as a consequence
   often yield very different results when run local to an ISP or
   interconnect and when run over a customer's complete path.  Neither
   the ISP nor customer can repeat the other's measurements, leading to
   high levels of distrust and acrimony.  Model-Based Metrics are
   expected to greatly improve this situation.

   Note that in situ measurements sometimes require sending synthetic
   measurement traffic between arbitrary locations in the network and,
   as such, are potentially attractive platforms for launching DDoS





Mathis & Morton               Experimental                     [Page 46]

RFC 8337                   Model-Based Metrics                March 2018


   attacks.  All active measurement tools and protocols must be designed
   to minimize the opportunities for these misuses.  See the discussion
   in Section 7 of [RFC7594].

   Some of the tests described in this document are not intended for
   frequent network monitoring since they have the potential to cause
   high network loads and might adversely affect other traffic.

   This document only describes a framework for designing a Fully
   Specified Targeted IP Diagnostic Suite.  Each FSTIDS must include its
   own security section.

12.  IANA Considerations



   This document has no IANA actions.

13.  Informative References



   [RFC863]   Postel, J., "Discard Protocol", STD 21, RFC 863,
              DOI 10.17487/RFC0863, May 1983,
              <https://www.rfc-editor.org/info/rfc863>.

   [RFC864]   Postel, J., "Character Generator Protocol", STD 22,
              RFC 864, DOI 10.17487/RFC0864, May 1983,
              <https://www.rfc-editor.org/info/rfc864>.

   [RFC2330]  Paxson, V., Almes, G., Mahdavi, J., and M. Mathis,
              "Framework for IP Performance Metrics", RFC 2330,
              DOI 10.17487/RFC2330, May 1998,
              <https://www.rfc-editor.org/info/rfc2330>.

   [RFC2861]  Handley, M., Padhye, J., and S. Floyd, "TCP Congestion
              Window Validation", RFC 2861, DOI 10.17487/RFC2861, June
              2000, <https://www.rfc-editor.org/info/rfc2861>.

   [RFC3148]  Mathis, M. and M. Allman, "A Framework for Defining
              Empirical Bulk Transfer Capacity Metrics", RFC 3148,
              DOI 10.17487/RFC3148, July 2001,
              <https://www.rfc-editor.org/info/rfc3148>.

   [RFC3168]  Ramakrishnan, K., Floyd, S., and D. Black, "The Addition
              of Explicit Congestion Notification (ECN) to IP",
              RFC 3168, DOI 10.17487/RFC3168, September 2001,
              <https://www.rfc-editor.org/info/rfc3168>.

   [RFC3465]  Allman, M., "TCP Congestion Control with Appropriate Byte
              Counting (ABC)", RFC 3465, DOI 10.17487/RFC3465, February
              2003, <https://www.rfc-editor.org/info/rfc3465>.



Mathis & Morton               Experimental                     [Page 47]

RFC 8337                   Model-Based Metrics                March 2018


   [RFC4737]  Morton, A., Ciavattone, L., Ramachandran, G., Shalunov,
              S., and J. Perser, "Packet Reordering Metrics", RFC 4737,
              DOI 10.17487/RFC4737, November 2006,
              <https://www.rfc-editor.org/info/rfc4737>.

   [RFC4898]  Mathis, M., Heffner, J., and R. Raghunarayan, "TCP
              Extended Statistics MIB", RFC 4898, DOI 10.17487/RFC4898,
              May 2007, <https://www.rfc-editor.org/info/rfc4898>.

   [RFC5136]  Chimento, P. and J. Ishac, "Defining Network Capacity",
              RFC 5136, DOI 10.17487/RFC5136, February 2008,
              <https://www.rfc-editor.org/info/rfc5136>.

   [RFC5681]  Allman, M., Paxson, V., and E. Blanton, "TCP Congestion
              Control", RFC 5681, DOI 10.17487/RFC5681, September 2009,
              <https://www.rfc-editor.org/info/rfc5681>.

   [RFC5827]  Allman, M., Avrachenkov, K., Ayesta, U., Blanton, J., and
              P. Hurtig, "Early Retransmit for TCP and Stream Control
              Transmission Protocol (SCTP)", RFC 5827,
              DOI 10.17487/RFC5827, May 2010,
              <https://www.rfc-editor.org/info/rfc5827>.

   [RFC5835]  Morton, A., Ed. and S. Van den Berghe, Ed., "Framework for
              Metric Composition", RFC 5835, DOI 10.17487/RFC5835, April
              2010, <https://www.rfc-editor.org/info/rfc5835>.

   [RFC6049]  Morton, A. and E. Stephan, "Spatial Composition of
              Metrics", RFC 6049, DOI 10.17487/RFC6049, January 2011,
              <https://www.rfc-editor.org/info/rfc6049>.

   [RFC6576]  Geib, R., Ed., Morton, A., Fardid, R., and A. Steinmitz,
              "IP Performance Metrics (IPPM) Standard Advancement
              Testing", BCP 176, RFC 6576, DOI 10.17487/RFC6576, March
              2012, <https://www.rfc-editor.org/info/rfc6576>.

   [RFC6673]  Morton, A., "Round-Trip Packet Loss Metrics", RFC 6673,
              DOI 10.17487/RFC6673, August 2012,
              <https://www.rfc-editor.org/info/rfc6673>.

   [RFC6928]  Chu, J., Dukkipati, N., Cheng, Y., and M. Mathis,
              "Increasing TCP's Initial Window", RFC 6928,
              DOI 10.17487/RFC6928, April 2013,
              <https://www.rfc-editor.org/info/rfc6928>.







Mathis & Morton               Experimental                     [Page 48]

RFC 8337                   Model-Based Metrics                March 2018


   [RFC7312]  Fabini, J. and A. Morton, "Advanced Stream and Sampling
              Framework for IP Performance Metrics (IPPM)", RFC 7312,
              DOI 10.17487/RFC7312, August 2014,
              <https://www.rfc-editor.org/info/rfc7312>.

   [RFC7398]  Bagnulo, M., Burbridge, T., Crawford, S., Eardley, P., and
              A. Morton, "A Reference Path and Measurement Points for
              Large-Scale Measurement of Broadband Performance",
              RFC 7398, DOI 10.17487/RFC7398, February 2015,
              <https://www.rfc-editor.org/info/rfc7398>.

   [RFC7567]  Baker, F., Ed. and G. Fairhurst, Ed., "IETF
              Recommendations Regarding Active Queue Management",
              BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015,
              <https://www.rfc-editor.org/info/rfc7567>.

   [RFC7594]  Eardley, P., Morton, A., Bagnulo, M., Burbridge, T.,
              Aitken, P., and A. Akhter, "A Framework for Large-Scale
              Measurement of Broadband Performance (LMAP)", RFC 7594,
              DOI 10.17487/RFC7594, September 2015,
              <https://www.rfc-editor.org/info/rfc7594>.

   [RFC7661]  Fairhurst, G., Sathiaseelan, A., and R. Secchi, "Updating
              TCP to Support Rate-Limited Traffic", RFC 7661,
              DOI 10.17487/RFC7661, October 2015,
              <https://www.rfc-editor.org/info/rfc7661>.

   [RFC7680]  Almes, G., Kalidindi, S., Zekauskas, M., and A. Morton,
              Ed., "A One-Way Loss Metric for IP Performance Metrics
              (IPPM)", STD 82, RFC 7680, DOI 10.17487/RFC7680, January
              2016, <https://www.rfc-editor.org/info/rfc7680>.

   [RFC7799]  Morton, A., "Active and Passive Metrics and Methods (with
              Hybrid Types In-Between)", RFC 7799, DOI 10.17487/RFC7799,
              May 2016, <https://www.rfc-editor.org/info/rfc7799>.

   [AFD]      Pan, R., Breslau, L., Prabhakar, B., and S. Shenker,
              "Approximate fairness through differential dropping", ACM
              SIGCOMM Computer Communication Review, Volume 33, Issue 2,
              DOI 10.1145/956981.956985, April 2003.

   [CCscaling]
              Paganini, F., Doyle, J., and S. Low, "Scalable laws for
              stable network congestion control", Proceedings of IEEE
              Conference on Decision and Control,,
              DOI 10.1109/CDC.2001.980095, December 2001.





Mathis & Morton               Experimental                     [Page 49]

RFC 8337                   Model-Based Metrics                March 2018


   [CVST]     Krueger, T. and M. Braun, "R package: Fast Cross-
              Validation via Sequential Testing", version 0.1, 11 2012.

   [iPerf]    Wikipedia, "iPerf", November 2017,
              <https://en.wikipedia.org/w/
              index.php?title=Iperf&oldid=810583885>.

   [MBMSource]
              "mbm", July 2016, <https://github.com/m-lab/MBM>.

   [Montgomery90]
              Montgomery, D., "Introduction to Statistical Quality
              Control", 2nd Edition, ISBN 0-471-51988-X, 1990.

   [mpingSource]
              "mping", July 2016, <https://github.com/m-lab/mping>.

   [MSMO97]   Mathis, M., Semke, J., Mahdavi, J., and T. Ott, "The
              Macroscopic Behavior of the TCP Congestion Avoidance
              Algorithm", Computer Communications Review, Volume 27,
              Issue 3, DOI 10.1145/263932.264023, July 1997.

   [Pathdiag] Mathis, M., Heffner, J., O'Neil, P., and P. Siemsen,
              "Pathdiag: Automated TCP Diagnosis", Passive and Active
              Network Measurement, Lecture Notes in Computer Science,
              Volume 4979, DOI 10.1007/978-3-540-79232-1_16, 2008.

   [Policing] Flach, T., Papageorge, P., Terzis, A., Pedrosa, L., Cheng,
              Y., Karim, T., Katz-Bassett, E., and R. Govindan, "An
              Internet-Wide Analysis of Traffic Policing", Proceedings
              of ACM SIGCOMM, DOI 10.1145/2934872.2934873, August 2016.

   [RACK]     Cheng, Y., Cardwell, N., Dukkipati, N., and P. Jha, "RACK:
              a time-based fast loss detection algorithm for TCP", Work
              in Progress, draft-ietf-tcpm-rack-03, March 2018.

   [Rtool]    R Development Core Team, "R: A language and environment
              for statistical computing", R Foundation for Statistical
              Computing, Vienna, Austria, ISBN 3-900051-07-0, 2011,
              <http://www.R-project.org/>.

   [TSO_fq_pacing]
              Dumazet, E. and Y. Chen, "TSO, fair queuing, pacing:
              three's a charm", Proceedings of IETF 88, TCPM WG,
              November 2013,
              <https://www.ietf.org/proceedings/88/slides/
              slides-88-tcpm-9.pdf>.




Mathis & Morton               Experimental                     [Page 50]

RFC 8337                   Model-Based Metrics                March 2018


   [TSO_pacing]
              Corbet, J., "TSO sizing and the FQ scheduler", August
              2013, <https://lwn.net/Articles/564978/>.

   [Wald45]   Wald, A., "Sequential Tests of Statistical Hypotheses",
              The Annals of Mathematical Statistics, Volume 16, Number
              2, pp. 117-186, June 1945,
              <http://www.jstor.org/stable/2235829>.

   [wikiBloat]
              Wikipedia, "Bufferbloat", January 2018,
              <https://en.wikipedia.org/w/
              index.php?title=Bufferbloat&oldid=819293377>.

   [WPING]    Mathis, M., "Windowed Ping: An IP Level Performance
              Diagnostic", Computer Networks and ISDN Systems, Volume
              27, Issue 3, DOI 10.1016/0169-7552(94)90119-8, June 1994.


































Mathis & Morton               Experimental                     [Page 51]

RFC 8337                   Model-Based Metrics                March 2018


Appendix A.  Model Derivations



   The reference target_run_length described in Section 5.2 is based on
   very conservative assumptions: that all excess data in flight (i.e.,
   the window size) above the target_window_size contributes to a
   standing queue that raises the RTT and that classic Reno congestion
   control with delayed ACKs is in effect.  In this section we provide
   two alternative calculations using different assumptions.

   It may seem out of place to allow such latitude in a measurement
   method, but this section provides offsetting requirements.

   The estimates provided by these models make the most sense if network
   performance is viewed logarithmically.  In the operational Internet,
   data rates span more than eight orders of magnitude, RTT spans more
   than three orders of magnitude, and packet loss ratio spans at least
   eight orders of magnitude if not more.  When viewed logarithmically
   (as in decibels), these correspond to 80 dB of dynamic range.  On an
   80 dB scale, a 3 dB error is less than 4% of the scale, even though
   it represents a factor of 2 in untransformed parameter.

   This document gives a lot of latitude for calculating
   target_run_length; however, people designing a TIDS should consider
   the effect of their choices on the ongoing tussle about the relevance
   of "TCP friendliness" as an appropriate model for Internet capacity
   allocation.  Choosing a target_run_length that is substantially
   smaller than the reference target_run_length specified in Section 5.2
   strengthens the argument that it may be appropriate to abandon "TCP
   friendliness" as the Internet fairness model.  This gives developers
   incentive and permission to develop even more aggressive applications
   and protocols, for example, by increasing the number of connections
   that they open concurrently.

A.1.  Queueless Reno



   In Section 5.2, models were derived based on the assumption that the
   subpath IP rate matches the target rate plus overhead, such that the
   excess window needed for the AIMD sawtooth causes a fluctuating queue
   at the bottleneck.

   An alternate situation would be a bottleneck where there is no
   significant queue and losses are caused by some mechanism that does
   not involve extra delay, for example, by the use of a virtual queue
   as done in Approximate Fair Dropping [AFD].  A flow controlled by
   such a bottleneck would have a constant RTT and a data rate that
   fluctuates in a sawtooth due to AIMD congestion control.  Assume the





Mathis & Morton               Experimental                     [Page 52]

RFC 8337                   Model-Based Metrics                March 2018


   losses are being controlled to make the average data rate meet some
   goal that is equal to or greater than the target_rate.  The necessary
   run length to meet the target_rate can be computed as follows:

   For some value of Wmin, the window will sweep from Wmin packets to
   2*Wmin packets in 2*Wmin RTT (due to delayed ACK).  Unlike the
   queuing case where Wmin = target_window_size, we want the average of
   Wmin and 2*Wmin to be the target_window_size, so the average data
   rate is the target rate.  Thus, we want Wmin =
   (2/3)*target_window_size.

   Between losses, each sawtooth delivers (1/2)(Wmin+2*Wmin)(2Wmin)
   packets in 2*Wmin RTTs.

   Substituting these together, we get:

   target_run_length = (4/3)(target_window_size^2)

   Note that this is 44% of the reference_run_length computed earlier.
   This makes sense because under the assumptions in Section 5.2, the
   AMID sawtooth caused a queue at the bottleneck, which raised the
   effective RTT by 50%.

Appendix B.  The Effects of ACK Scheduling



   For many network technologies, simple queuing models don't apply: the
   network schedules, thins, or otherwise alters the timing of ACKs and
   data, generally to raise the efficiency of the channel allocation
   algorithms when confronted with relatively widely spaced small ACKs.
   These efficiency strategies are ubiquitous for half-duplex, wireless,
   and broadcast media.

   Altering the ACK stream by holding or thinning ACKs typically has two
   consequences: it raises the implied bottleneck IP capacity, making
   the fine-grained slowstart bursts either faster or larger, and it
   raises the effective RTT by the average time that the ACKs and data
   are delayed.  The first effect can be partially mitigated by
   re-clocking ACKs once they are beyond the bottleneck on the return
   path to the sender; however, this further raises the effective RTT.

   The most extreme example of this sort of behavior would be a half-
   duplex channel that is not released as long as the endpoint currently
   holding the channel has more traffic (data or ACKs) to send.  Such
   environments cause self-clocked protocols under full load to revert
   to extremely inefficient stop-and-wait behavior.  The channel
   constrains the protocol to send an entire window of data as a single





Mathis & Morton               Experimental                     [Page 53]

RFC 8337                   Model-Based Metrics                March 2018


   contiguous burst on the forward path, followed by the entire window
   of ACKs on the return path.  (A channel with this behavior would fail
   the Duplex Self-Interference Test described in Section 8.2.4).

   If a particular return path contains a subpath or device that alters
   the timing of the ACK stream, then the entire front path from the
   sender up to the bottleneck must be tested at the burst parameters
   implied by the ACK scheduling algorithm.  The most important
   parameter is the implied bottleneck IP capacity, which is the average
   rate at which the ACKs advance snd.una.  Note that thinning the ACK
   stream (relying on the cumulative nature of seg.ack to permit
   discarding some ACKs) causes most TCP implementations to send
   interface rate bursts to offset the longer times between ACKs in
   order to maintain the average data rate.

   Note that due to ubiquitous self-clocking in Internet protocols,
   ill-conceived channel allocation mechanisms are likely to increases
   the queuing stress on the front path because they cause larger full
   sender rate data bursts.

   Holding data or ACKs for channel allocation or other reasons (such as
   forward error correction) always raises the effective RTT relative to
   the minimum delay for the path.  Therefore, it may be necessary to
   replace target_RTT in the calculation in Section 5.2 by an
   effective_RTT, which includes the target_RTT plus a term to account
   for the extra delays introduced by these mechanisms.

























Mathis & Morton               Experimental                     [Page 54]

RFC 8337                   Model-Based Metrics                March 2018


Acknowledgments

   Ganga Maguluri suggested the statistical test for measuring loss
   probability in the target run length.  Alex Gilgur and Merry Mou
   helped with the statistics.

   Meredith Whittaker improved the clarity of the communications.

   Ruediger Geib provided feedback that greatly improved the document.

   This work was inspired by Measurement Lab: open tools running on an
   open platform, using open tools to collect open data.  See
   <http://www.measurementlab.net/>.

Authors' Addresses



   Matt Mathis
   Google, Inc
   1600 Amphitheatre Parkway
   Mountain View, CA  94043
   United States of America

   Email: mattmathis@google.com


   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


















Mathis & Morton               Experimental                     [Page 55]