RFC 2330






Network Working Group                                          V. Paxson
Request for Comments: 2330                Lawrence Berkeley National Lab
Category: Informational                                         G. Almes
                                             Advanced Network & Services
                                                              J. Mahdavi
                                                               M. Mathis
                                         Pittsburgh Supercomputer Center
                                                                May 1998


                  Framework for IP Performance Metrics


1. Status of this Memo



   This memo provides information for the Internet community.  It does
   not specify an Internet standard of any kind.  Distribution of this
   memo is unlimited.


2. Copyright Notice



   Copyright (C) The Internet Society (1998).  All Rights Reserved.

Table of Contents



   1.  STATUS OF THIS MEMO.............................................1
   2.  COPYRIGHT NOTICE................................................1
   3.  INTRODUCTION....................................................2
   4.  CRITERIA FOR IP PERFORMANCE METRICS.............................3
   5.  TERMINOLOGY FOR PATHS AND CLOUDS................................4
   6.  FUNDAMENTAL CONCEPTS............................................5
     6.1  Metrics......................................................5
     6.2  Measurement Methodology......................................6
     6.3  Measurements, Uncertainties, and Errors......................7
   7.  METRICS AND THE ANALYTICAL FRAMEWORK............................8
   8.  EMPIRICALLY SPECIFIED METRICS..................................11
   9.  TWO FORMS OF COMPOSITION.......................................12
     9.1  Spatial Composition of Metrics..............................12
     9.2  Temporal Composition of Formal Models and Empirical Metrics.13
   10.  ISSUES RELATED TO TIME........................................14
     10.1  Clock Issues...............................................14
     10.2  The Notion of "Wire Time"..................................17
   11. SINGLETONS, SAMPLES, AND STATISTICS............................19
     11.1  Methods of Collecting Samples..............................20
       11.1.1 Poisson Sampling........................................21
       11.1.2 Geometric Sampling......................................22
       11.1.3 Generating Poisson Sampling Intervals...................22



Paxson, et. al.              Informational                      [Page 1]

RFC 2330          Framework for IP Performance Metrics          May 1998


     11.2  Self-Consistency...........................................24
     11.3  Defining Statistical Distributions.........................25
     11.4  Testing For Goodness-of-Fit................................27
   12. AVOIDING STOCHASTIC METRICS....................................28
   13. PACKETS OF TYPE P..............................................29
   14. INTERNET ADDRESSES VS. HOSTS...................................30
   15. STANDARD-FORMED PACKETS........................................30
   16. ACKNOWLEDGEMENTS...............................................31
   17. SECURITY CONSIDERATIONS........................................31
   18. APPENDIX.......................................................32
   19. REFERENCES.....................................................38
   20. AUTHORS' ADDRESSES.............................................39
   21. FULL COPYRIGHT STATEMENT.......................................40


3. Introduction



   The purpose of this memo is to define a general framework for
   particular metrics to be developed by the IETF's IP Performance
   Metrics effort, begun by the Benchmarking Methodology Working Group
   (BMWG) of the Operational Requirements Area, and being continued by
   the IP Performance Metrics Working Group (IPPM) of the Transport
   Area.

   We begin by laying out several criteria for the metrics that we
   adopt.  These criteria are designed to promote an IPPM effort that
   will maximize an accurate common understanding by Internet users and
   Internet providers of the performance and reliability both of end-
   to-end paths through the Internet and of specific 'IP clouds' that
   comprise portions of those paths.

   We next define some Internet vocabulary that will allow us to speak
   clearly about Internet components such as routers, paths, and clouds.

   We then define the fundamental concepts of 'metric' and 'measurement
   methodology', which allow us to speak clearly about measurement
   issues.  Given these concepts, we proceed to discuss the important
   issue of measurement uncertainties and errors, and develop a key,
   somewhat subtle notion of how they relate to the analytical framework
   shared by many aspects of the Internet engineering discipline.  We
   then introduce the notion of empirically defined metrics, and finish
   this part of the document with a general discussion of how metrics
   can be 'composed'.

   The remainder of the document deals with a variety of issues related
   to defining sound metrics and methodologies:  how to deal with
   imperfect clocks; the notion of 'wire time' as distinct from 'host
   time'; how to aggregate sets of singleton metrics into samples and



Paxson, et. al.              Informational                      [Page 2]

RFC 2330          Framework for IP Performance Metrics          May 1998


   derive sound statistics from those samples; why it is recommended to
   avoid thinking about Internet properties in probabilistic terms (such
   as the probability that a packet is dropped), since these terms often
   include implicit assumptions about how the network behaves; the
   utility of defining metrics in terms of packets of a generic type;
   the benefits of preferring IP addresses to DNS host names; and the
   notion of 'standard-formed' packets.  An appendix discusses the
   Anderson-Darling test for gauging whether a set of values matches a
   given statistical distribution, and gives C code for an
   implementation of the test.

   In some sections of the memo, we will surround some commentary text
   with the brackets {Comment: ... }.  We stress that this commentary is
   only commentary, and is not itself part of the framework document or
   a proposal of particular metrics.  In some cases this commentary will
   discuss some of the properties of metrics that might be envisioned,
   but the reader should assume that any such discussion is intended
   only to shed light on points made in the framework document, and not
   to suggest any specific metrics.


4. Criteria for IP Performance Metrics



   The overarching goal of the IP Performance Metrics effort is to
   achieve a situation in which users and providers of Internet
   transport service have an accurate common understanding of the
   performance and reliability of the Internet component 'clouds' that
   they use/provide.

   To achieve this, performance and reliability metrics for paths
   through the Internet must be developed.  In several IETF meetings
   criteria for these metrics have been specified:

 +    The metrics must be concrete and well-defined,
 +    A methodology for a metric should have the property that it is
      repeatable: if the methodology is used multiple times under
      identical conditions, the same measurements should result in the
      same measurements.
 +    The metrics must exhibit no bias for IP clouds implemented with
      identical technology,
 +    The metrics must exhibit understood and fair bias for IP clouds
      implemented with non-identical technology,
 +    The metrics must be useful to users and providers in understanding
      the performance they experience or provide,







Paxson, et. al.              Informational                      [Page 3]

RFC 2330          Framework for IP Performance Metrics          May 1998


 +    The metrics must avoid inducing artificial performance goals.


5. Terminology for Paths and Clouds



   The following list defines terms that need to be precise in the
   development of path metrics.  We begin with low-level notions of
   'host', 'router', and 'link', then proceed to define the notions of
   'path', 'IP cloud', and 'exchange' that allow us to segment a path
   into relevant pieces.

   host A computer capable of communicating using the Internet
        protocols; includes "routers".

   link A single link-level connection between two (or more) hosts;
        includes leased lines, ethernets, frame relay clouds, etc.

   routerA host which facilitates network-level communication between
        hosts by forwarding IP packets.

   path A sequence of the form < h0, l1, h1, ..., ln, hn >, where n >=
        0, each hi is a host, each li is a link between hi-1 and hi,
        each h1...hn-1 is a router.  A pair <li, hi> is termed a 'hop'.
        In an appropriate operational configuration, the links and
        routers in the path facilitate network-layer communication of
        packets from h0 to hn.  Note that path is a unidirectional
        concept.

   subpath
        Given a path, a subpath is any subsequence of the given path
        which is itself a path.  (Thus, the first and last element of a
        subpath is a host.)

   cloudAn undirected (possibly cyclic) graph whose vertices are routers
        and whose edges are links that connect pairs of routers.
        Formally, ethernets, frame relay clouds, and other links that
        connect more than two routers are modelled as fully-connected
        meshes of graph edges.  Note that to connect to a cloud means to
        connect to a router of the cloud over a link; this link is not
        itself part of the cloud.

   exchange
        A special case of a link, an exchange directly connects either a
        host to a cloud and/or one cloud to another cloud.

   cloud subpath
        A subpath of a given path, all of whose hosts are routers of a
        given cloud.



Paxson, et. al.              Informational                      [Page 4]

RFC 2330          Framework for IP Performance Metrics          May 1998


   path digest
        A sequence of the form < h0, e1, C1, ..., en, hn >, where n >=
        0, h0 and hn are hosts, each e1 ... en is an exchange, and each
        C1 ... Cn-1 is a cloud subpath.

6. Fundamental Concepts




6.1. Metrics



   In the operational Internet, there are several quantities related to
   the performance and reliability of the Internet that we'd like to
   know the value of.  When such a quantity is carefully specified, we
   term the quantity a metric.  We anticipate that there will be
   separate RFCs for each metric (or for each closely related group of
   metrics).

   In some cases, there might be no obvious means to effectively measure
   the metric; this is allowed, and even understood to be very useful in
   some cases.  It is required, however, that the specification of the
   metric be as clear as possible about what quantity is being
   specified.  Thus, difficulty in practical measurement is sometimes
   allowed, but ambiguity in meaning is not.

   Each metric will be defined in terms of standard units of
   measurement.  The international metric system will be used, with the
   following points specifically noted:

 +    When a unit is expressed in simple meters (for distance/length) or
      seconds (for duration), appropriate related units based on
      thousands or thousandths of acceptable units are acceptable.
      Thus, distances expressed in kilometers (km), durations expressed
      in milliseconds (ms), or microseconds (us) are allowed, but not
      centimeters (because the prefix is not in terms of thousands or
      thousandths).
 +    When a unit is expressed in a combination of units, appropriate
      related units based on thousands or thousandths of acceptable
      units are acceptable, but all such thousands/thousandths must be
      grouped at the beginning.  Thus, kilo-meters per second (km/s) is
      allowed, but meters per millisecond is not.
 +    The unit of information is the bit.
 +    When metric prefixes are used with bits or with combinations
      including bits, those prefixes will have their metric meaning
      (related to decimal 1000), and not the meaning conventional with
      computer storage (related to decimal 1024).  In any RFC that
      defines a metric whose units include bits, this convention will be
      followed and will be repeated to ensure clarity for the reader.




Paxson, et. al.              Informational                      [Page 5]

RFC 2330          Framework for IP Performance Metrics          May 1998


 +    When a time is given, it will be expressed in UTC.

   Note that these points apply to the specifications for metrics and
   not, for example, to packet formats where octets will likely be used
   in preference/addition to bits.

   Finally, we note that some metrics may be defined purely in terms of
   other metrics; such metrics are call 'derived metrics'.


6.2. Measurement Methodology



   For a given set of well-defined metrics, a number of distinct
   measurement methodologies may exist.  A partial list includes:

 +    Direct measurement of a performance metric using injected test
      traffic.  Example: measurement of the round-trip delay of an IP
      packet of a given size over a given route at a given time.
 +    Projection of a metric from lower-level measurements.  Example:
      given accurate measurements of propagation delay and bandwidth for
      each step along a path, projection of the complete delay for the
      path for an IP packet of a given size.
 +    Estimation of a constituent metric from a set of more aggregated
      measurements.  Example: given accurate measurements of delay for a
      given one-hop path for IP packets of different sizes, estimation
      of propagation delay for the link of that one-hop path.
 +    Estimation of a given metric at one time from a set of related
      metrics at other times.  Example: given an accurate measurement of
      flow capacity at a past time, together with a set of accurate
      delay measurements for that past time and the current time, and
      given a model of flow dynamics, estimate the flow capacity that
      would be observed at the current time.

   This list is by no means exhaustive.  The purpose is to point out the
   variety of measurement techniques.

   When a given metric is specified, a given measurement approach might
   be noted and discussed.  That approach, however, is not formally part
   of the specification.

   A methodology for a metric should have the property that it is
   repeatable: if the methodology is used multiple times under identical
   conditions, it should result in consistent measurements.

   Backing off a little from the word 'identical' in the previous
   paragraph, we could more accurately use the word 'continuity' to
   describe a property of a given methodology: a methodology for a given
   metric exhibits continuity if, for small variations in conditions, it



Paxson, et. al.              Informational                      [Page 6]

RFC 2330          Framework for IP Performance Metrics          May 1998


   results in small variations in the resulting measurements.  Slightly
   more precisely, for every positive epsilon, there exists a positive
   delta, such that if two sets of conditions are within delta of each
   other, then the resulting measurements will be within epsilon of each
   other.  At this point, this should be taken as a heuristic driving
   our intuition about one kind of robustness property rather than as a
   precise notion.

   A metric that has at least one methodology that exhibits continuity
   is said itself to exhibit continuity.

   Note that some metrics, such as hop-count along a path, are integer-
   valued and therefore cannot exhibit continuity in quite the sense
   given above.

   Note further that, in practice, it may not be practical to know (or
   be able to quantify) the conditions relevant to a measurement at a
   given time.  For example, since the instantaneous load (in packets to
   be served) at a given router in a high-speed wide-area network can
   vary widely over relatively brief periods and will be very hard for
   an external observer to quantify, various statistics of a given
   metric may be more repeatable, or may better exhibit continuity.  In
   that case those particular statistics should be specified when the
   metric is specified.

   Finally, some measurement methodologies may be 'conservative' in the
   sense that the act of measurement does not modify, or only slightly
   modifies, the value of the performance metric the methodology
   attempts to measure.  {Comment: for example, in a wide-are high-speed
   network under modest load, a test using several small 'ping' packets
   to measure delay would likely not interfere (much) with the delay
   properties of that network as observed by others.  The corresponding
   statement about tests using a large flow to measure flow capacity
   would likely fail.}


6.3. Measurements, Uncertainties, and Errors



   Even the very best measurement methodologies for the very most well
   behaved metrics will exhibit errors.  Those who develop such
   measurement methodologies, however, should strive to:










Paxson, et. al.              Informational                      [Page 7]

RFC 2330          Framework for IP Performance Metrics          May 1998


 +    minimize their uncertainties/errors,
 +    understand and document the sources of uncertainty/error, and
 +    quantify the amounts of uncertainty/error.

   For example, when developing a method for measuring delay, understand
   how any errors in your clocks introduce errors into your delay
   measurement, and quantify this effect as well as you can.  In some
   cases, this will result in a requirement that a clock be at least up
   to a certain quality if it is to be used to make a certain
   measurement.

   As a second example, consider the timing error due to measurement
   overheads within the computer making the measurement, as opposed to
   delays due to the Internet component being measured.  The former is a
   measurement error, while the latter reflects the metric of interest.
   Note that one technique that can help avoid this overhead is the use
   of a packet filter/sniffer, running on a separate computer that
   records network packets and timestamps them accurately (see the
   discussion of 'wire time' below).  The resulting trace can then be
   analyzed to assess the test traffic, minimizing the effect of
   measurement host delays, or at least allowing those delays to be
   accounted for.  We note that this technique may prove beneficial even
   if the packet filter/sniffer runs on the same machine, because such
   measurements generally provide 'kernel-level' timestamping as opposed
   to less-accurate 'application-level' timestamping.

   Finally, we note that derived metrics (defined above) or metrics that
   exhibit spatial or temporal composition (defined below) offer
   particular occasion for the analysis of measurement uncertainties,
   namely how the uncertainties propagate (conceptually) due to the
   derivation or composition.


7. Metrics and the Analytical Framework



   As the Internet has evolved from the early packet-switching studies
   of the 1960s, the Internet engineering community has evolved a common
   analytical framework of concepts.  This analytical framework, or A-
   frame, used by designers and implementers of protocols, by those
   involved in measurement, and by those who study computer network
   performance using the tools of simulation and analysis, has great
   advantage to our work.  A major objective here is to generate network
   characterizations that are consistent in both analytical and
   practical settings, since this will maximize the chances that non-
   empirical network study can be better correlated with, and used to
   further our understanding of, real network behavior.





Paxson, et. al.              Informational                      [Page 8]

RFC 2330          Framework for IP Performance Metrics          May 1998


   Whenever possible, therefore, we would like to develop and leverage
   off of the A-frame.  Thus, whenever a metric to be specified is
   understood to be closely related to concepts within the A-frame, we
   will attempt to specify the metric in the A-frame's terms.  In such a
   specification we will develop the A-frame by precisely defining the
   concepts needed for the metric, then leverage off of the A-frame by
   defining the metric in terms of those concepts.

   Such a metric will be called an 'analytically specified metric' or,
   more simply, an analytical metric.

   {Comment: Examples of such analytical metrics might include:

propagation time of a link
     The time, in seconds, required by a single bit to travel from the
     output port on one Internet host across a single link to another
     Internet host.

bandwidth of a link for packets of size k
     The capacity, in bits/second, where only those bits of the IP
     packet are counted, for packets of size k bytes.

routeThe path, as defined in Section 5, from A to B at a given time.

hop count of a route
     The value 'n' of the route path.
     }

     Note that we make no a priori list of just what A-frame concepts
     will emerge in these specifications, but we do encourage their use
     and urge that they be carefully specified so that, as our set of
     metrics develops, so will a specified set of A-frame concepts
     technically consistent with each other and consonant with the
     common understanding of those concepts within the general Internet
     community.

     These A-frame concepts will be intended to abstract from actual
     Internet components in such a way that:

 +    the essential function of the component is retained,
 +    properties of the component relevant to the metrics we aim to
      create are retained,
 +    a subset of these component properties are potentially defined as
      analytical metrics, and







Paxson, et. al.              Informational                      [Page 9]

RFC 2330          Framework for IP Performance Metrics          May 1998


 +    those properties of actual Internet components not relevant to
      defining the metrics we aim to create are dropped.

   For example, when considering a router in the context of packet
   forwarding, we might model the router as a component that receives
   packets on an input link, queues them on a FIFO packet queue of
   finite size, employs tail-drop when the packet queue is full, and
   forwards them on an output link.  The transmission speed (in
   bits/second) of the input and output links, the latency in the router
   (in seconds), and the maximum size of the packet queue (in bits) are
   relevant analytical metrics.

   In some cases, such analytical metrics used in relation to a router
   will be very closely related to specific metrics of the performance
   of Internet paths.  For example, an obvious formula (L + P/B)
   involving the latency in the router (L), the packet size (in bits)
   (P), and the transmission speed of the output link (B) might closely
   approximate the increase in packet delay due to the insertion of a
   given router along a path.

   We stress, however, that well-chosen and well-specified A-frame
   concepts and their analytical metrics will support more general
   metric creation efforts in less obvious ways.

   {Comment: for example, when considering the flow capacity of a path,
   it may be of real value to be able to model each of the routers along
   the path as packet forwarders as above.  Techniques for estimating
   the flow capacity of a path might use the maximum packet queue size
   as a parameter in decidedly non-obvious ways.  For example, as the
   maximum queue size increases, so will the ability of the router to
   continuously move traffic along an output link despite fluctuations
   in traffic from an input link.  Estimating this increase, however,
   remains a research topic.}

   Note that, when we specify A-frame concepts and analytical metrics,
   we will inevitably make simplifying assumptions.  The key role of
   these concepts is to abstract the properties of the Internet
   components relevant to given metrics.  Judgement is required to avoid
   making assumptions that bias the modeling and metric effort toward
   one kind of design.

   {Comment: for example, routers might not use tail-drop, even though
   tail-drop might be easier to model analytically.}

   Finally, note that different elements of the A-frame might well make
   different simplifying assumptions.  For example, the abstraction of a
   router used to further the definition of path delay might treat the
   router's packet queue as a single FIFO queue, but the abstraction of



Paxson, et. al.              Informational                     [Page 10]

RFC 2330          Framework for IP Performance Metrics          May 1998


   a router used to further the definition of the handling of an RSVP-
   enabled packet might treat the router's packet queue as supporting
   bounded delay -- a contradictory assumption.  This is not to say that
   we make contradictory assumptions at the same time, but that two
   different parts of our work might refine the simpler base concept in
   two divergent ways for different purposes.

   {Comment: in more mathematical terms, we would say that the A-frame
   taken as a whole need not be consistent; but the set of particular
   A-frame elements used to define a particular metric must be.}


8. Empirically Specified Metrics



   There are useful performance and reliability metrics that do not fit
   so neatly into the A-frame, usually because the A-frame lacks the
   detail or power for dealing with them.  For example, "the best flow
   capacity achievable along a path using an RFC-2001-compliant TCP"
   would be good to be able to measure, but we have no analytical
   framework of sufficient richness to allow us to cast that flow
   capacity as an analytical metric.

   These notions can still be well specified by instead describing a
   reference methodology for measuring them.

   Such a metric will be called an 'empirically specified metric', or
   more simply, an empirical metric.

   Such empirical metrics should have three properties:

 +    we should have a clear definition for each in terms of Internet
      components,
 +    we should have at least one effective means to measure them, and
 +    to the extent possible, we should have an (necessarily incomplete)
      understanding of the metric in terms of the A-frame so that we can
      use our measurements to reason about the performance and
      reliability of A-frame components and of aggregations of A-frame
      components.













Paxson, et. al.              Informational                     [Page 11]

RFC 2330          Framework for IP Performance Metrics          May 1998


9. Two Forms of Composition




9.1. Spatial Composition of Metrics



   In some cases, it may be realistic and useful to define metrics in
   such a fashion that they exhibit spatial composition.

   By spatial composition, we mean a characteristic of some path
   metrics, in which the metric as applied to a (complete) path can also
   be defined for various subpaths, and in which the appropriate A-frame
   concepts for the metric suggest useful relationships between the
   metric applied to these various subpaths (including the complete
   path, the various cloud subpaths of a given path digest, and even
   single routers along the path).  The effectiveness of spatial
   composition depends:

 +    on the usefulness in analysis of these relationships as applied to
      the relevant A-frame components, and
 +    on the practical use of the corresponding relationships as applied
      to metrics and to measurement methodologies.

   {Comment: for example, consider some metric for delay of a 100-byte
   packet across a path P, and consider further a path digest <h0, e1,
   C1, ..., en, hn> of P.  The definition of such a metric might include
   a conjecture that the delay across P is very nearly the sum of the
   corresponding metric across the exchanges (ei) and clouds (Ci) of the
   given path digest.  The definition would further include a note on
   how a corresponding relation applies to relevant A-frame components,
   both for the path P and for the exchanges and clouds of the path
   digest.}

   When the definition of a metric includes a conjecture that the metric
   across the path is related to the metric across the subpaths of the
   path, that conjecture constitutes a claim that the metric exhibits
   spatial composition.  The definition should then include:















Paxson, et. al.              Informational                     [Page 12]

RFC 2330          Framework for IP Performance Metrics          May 1998


 +    the specific conjecture applied to the metric,
 +    a justification of the practical utility of the composition in
      terms of making accurate measurements of the metric on the path,
 +    a justification of the usefulness of the composition in terms of
      making analysis of the path using A-frame concepts more effective,
      and
 +    an analysis of how the conjecture could be incorrect.


9.2. Temporal Composition of Formal Models and Empirical Metrics



   In some cases, it may be realistic and useful to define metrics in
   such a fashion that they exhibit temporal composition.

   By temporal composition, we mean a characteristic of some path
   metric, in which the metric as applied to a path at a given time T is
   also defined for various times t0 < t1 < ... < tn < T, and in which
   the appropriate A-frame concepts for the metric suggests useful
   relationships between the metric applied at times t0, ..., tn and the
   metric applied at time T.  The effectiveness of temporal composition
   depends:

 +    on the usefulness in analysis of these relationships as applied to
      the relevant A-frame components, and
 +    on the practical use of the corresponding relationships as applied
      to metrics and to measurement methodologies.

   {Comment: for example, consider a  metric for the expected flow
   capacity across a path P during the five-minute period surrounding
   the time T, and suppose further that we have the corresponding values
   for each of the four previous five-minute periods t0, t1, t2, and t3.
   The definition of such a metric might include a conjecture that the
   flow capacity at time T can be estimated from a certain kind of
   extrapolation from the values of t0, ..., t3.  The definition would
   further include a note on how a corresponding relation applies to
   relevant A-frame components.

   Note: any (spatial or temporal) compositions involving flow capacity
   are likely to be subtle, and temporal compositions are generally more
   subtle than spatial compositions, so the reader should understand
   that the foregoing example is intentionally naive.}

   When the definition of a metric includes a conjecture that the metric
   across the path at a given time T is related to the metric across the
   path for a set of other times, that conjecture constitutes a claim
   that the metric exhibits temporal composition.  The definition should
   then include:




Paxson, et. al.              Informational                     [Page 13]

RFC 2330          Framework for IP Performance Metrics          May 1998


 +    the specific conjecture applied to the metric,
 +    a justification of the practical utility of the composition in
      terms of making accurate measurements of the metric on the path,
      and
 +    a justification of the usefulness of the composition in terms of
      making analysis of the path using A-frame concepts more effective.


10. Issues related to Time




10.1. Clock Issues



   Measurements of time lie at the heart of many Internet metrics.
   Because of this, it will often be crucial when designing a
   methodology for measuring a metric to understand the different types
   of errors and uncertainties introduced by imperfect clocks.  In this
   section we define terminology for discussing the characteristics of
   clocks and touch upon related measurement issues which need to be
   addressed by any sound methodology.

   The Network Time Protocol (NTP; RFC 1305) defines a nomenclature for
   discussing clock characteristics, which we will also use when
   appropriate [Mi92].  The main goal of NTP is to provide accurate
   timekeeping over fairly long time scales, such as minutes to days,
   while for measurement purposes often what is more important is
   short-term accuracy, between the beginning of the measurement and the
   end, or over the course of gathering a body of measurements (a
   sample).  This difference in goals sometimes leads to different
   definitions of terminology as well, as discussed below.

   To begin, we define a clock's "offset" at a particular moment as the
   difference between the time reported by the clock and the "true" time
   as defined by UTC.  If the clock reports a time Tc and the true time
   is Tt, then the clock's offset is Tc - Tt.

   We will refer to a clock as "accurate" at a particular moment if the
   clock's offset is zero, and more generally a clock's "accuracy" is
   how close the absolute value of the offset is to zero.  For NTP,
   accuracy also includes a notion of the frequency of the clock; for
   our purposes, we instead incorporate this notion into that of "skew",
   because we define accuracy in terms of a single moment in time rather
   than over an interval of time.

   A clock's "skew" at a particular moment is the frequency difference
   (first derivative of its offset with respect to true time) between
   the clock and true time.




Paxson, et. al.              Informational                     [Page 14]

RFC 2330          Framework for IP Performance Metrics          May 1998


   As noted in RFC 1305, real clocks exhibit some variation in skew.
   That is, the second derivative of the clock's offset with respect to
   true time is generally non-zero.  In keeping with RFC 1305, we define
   this quantity as the clock's "drift".

   A clock's "resolution" is the smallest unit by which the clock's time
   is updated.  It gives a lower bound on the clock's uncertainty.
   (Note that clocks can have very fine resolutions and yet be wildly
   inaccurate.)  Resolution is defined in terms of seconds.  However,
   resolution is relative to the clock's reported time and not to true
   time, so for example a resolution of 10 ms only means that the clock
   updates its notion of time in 0.01 second increments, not that this
   is the true amount of time between updates.

   {Comment: Systems differ on how an application interface to the clock
   reports the time on subsequent calls during which the clock has not
   advanced.  Some systems simply return the same unchanged time as
   given for previous calls.  Others may add a small increment to the
   reported time to maintain monotone-increasing timestamps.  For
   systems that do the latter, we do *not* consider these small
   increments when defining the clock's resolution.  They are instead an
   impediment to assessing the clock's resolution, since a natural
   method for doing so is to repeatedly query the clock to determine the
   smallest non-zero difference in reported times.}

   It is expected that a clock's resolution changes only rarely (for
   example, due to a hardware upgrade).

   There are a number of interesting metrics for which some natural
   measurement methodologies involve comparing times reported by two
   different clocks.  An example is one-way packet delay [AK97].  Here,
   the time required for a packet to travel through the network is
   measured by comparing the time reported by a clock at one end of the
   packet's path, corresponding to when the packet first entered the
   network, with the time reported by a clock at the other end of the
   path, corresponding to when the packet finished traversing the
   network.

   We are thus also interested in terminology for describing how two
   clocks C1 and C2 compare.  To do so, we introduce terms related to
   those above in which the notion of "true time" is replaced by the
   time as reported by clock C1.  For example, clock C2's offset
   relative to C1 at a particular moment is Tc2 - Tc1, the instantaneous
   difference in time reported by C2 and C1.  To disambiguate between
   the use of the terms to compare two clocks versus the use of the
   terms to compare to true time, we will in the former case use the
   phrase "relative".  So the offset defined earlier in this paragraph
   is the "relative offset" between C2 and C1.



Paxson, et. al.              Informational                     [Page 15]

RFC 2330          Framework for IP Performance Metrics          May 1998


   When comparing clocks, the analog of "resolution" is not "relative
   resolution", but instead "joint resolution", which is the sum of the
   resolutions of C1 and C2.  The joint resolution then indicates a
   conservative lower bound on the accuracy of any time intervals
   computed by subtracting timestamps generated by one clock from those
   generated by the other.

   If two clocks are "accurate" with respect to one another (their
   relative offset is zero), we will refer to the pair of clocks as
   "synchronized".  Note that clocks can be highly synchronized yet
   arbitrarily inaccurate in terms of how well they tell true time.
   This point is important because for many Internet measurements,
   synchronization between two clocks is more important than the
   accuracy of the clocks.  The is somewhat true of skew, too: as long
   as the absolute skew is not too great, then minimal relative skew is
   more important, as it can induce systematic trends in packet transit
   times measured by comparing timestamps produced by the two clocks.

   These distinctions arise because for Internet measurement what is
   often most important are differences in time as computed by comparing
   the output of two clocks.  The process of computing the difference
   removes any error due to clock inaccuracies with respect to true
   time; but it is crucial that the differences themselves accurately
   reflect differences in true time.

   Measurement methodologies will often begin with the step of assuring
   that two clocks are synchronized and have minimal skew and drift.
   {Comment: An effective way to assure these conditions (and also clock
   accuracy) is by using clocks that derive their notion of time from an
   external source, rather than only the host computer's clock.  (These
   latter are often subject to large errors.) It is further preferable
   that the clocks directly derive their time, for example by having
   immediate access to a GPS (Global Positioning System) unit.}

   Two important concerns arise if the clocks indirectly derive their
   time using a network time synchronization protocol such as NTP:

 +    First, NTP's accuracy depends in part on the properties
      (particularly delay) of the Internet paths used by the NTP peers,
      and these might be exactly the properties that we wish to measure,
      so it would be unsound to use NTP to calibrate such measurements.
 +    Second, NTP focuses on clock accuracy, which can come at the
      expense of short-term clock skew and drift.  For example, when a
      host's clock is indirectly synchronized to a time source, if the
      synchronization intervals occur infrequently, then the host will
      sometimes be faced with the problem of how to adjust its current,
      incorrect time, Ti, with a considerably different, more accurate
      time it has just learned, Ta.  Two general ways in which this is



Paxson, et. al.              Informational                     [Page 16]

RFC 2330          Framework for IP Performance Metrics          May 1998


      done are to either immediately set the current time to Ta, or to
      adjust the local clock's update frequency (hence, its skew) so
      that at some point in the future the local time Ti' will agree
      with the more accurate time Ta'.  The first mechanism introduces
      discontinuities and can also violate common assumptions that
      timestamps are monotone increasing.  If the host's clock is set
      backward in time, sometimes this can be easily detected.  If the
      clock is set forward in time, this can be harder to detect.  The
      skew induced by the second mechanism can lead to considerable
      inaccuracies when computing differences in time, as discussed
      above.

   To illustrate why skew is a crucial concern, consider samples of
   one-way delays between two Internet hosts made at one minute
   intervals.  The true transmission delay between the hosts might
   plausibly be on the order of 50 ms for a transcontinental path.  If
   the skew between the two clocks is 0.01%, that is, 1 part in 10,000,
   then after 10 minutes of observation the error introduced into the
   measurement is 60 ms.  Unless corrected, this error is enough to
   completely wipe out any accuracy in the transmission delay
   measurement.  Finally, we note that assessing skew errors between
   unsynchronized network clocks is an open research area.  (See [Pa97]
   for a discussion of detecting and compensating for these sorts of
   errors.) This shortcoming makes use of a solid, independent clock
   source such as GPS especially desirable.


10.2. The Notion of "Wire Time"



   Internet measurement is often complicated by the use of Internet
   hosts themselves to perform the measurement.  These hosts can
   introduce delays, bottlenecks, and the like that are due to hardware
   or operating system effects and have nothing to do with the network
   behavior we would like to measure.  This problem is particularly
   acute when timestamping of network events occurs at the application
   level.

   In order to provide a general way of talking about these effects, we
   introduce two notions of "wire time".  These notions are only defined
   in terms of an Internet host H observing an Internet link L at a
   particular location:

 +    For a given packet P, the 'wire arrival time' of P at H on L is
      the first time T at which any bit of P has appeared at H's
      observational position on L.






Paxson, et. al.              Informational                     [Page 17]

RFC 2330          Framework for IP Performance Metrics          May 1998


 +    For a given packet P, the 'wire exit time' of P at H on L is the
      first time T at which all the bits of P have appeared at H's
      observational position on L.

   Note that intrinsic to the definition is the notion of where on the
   link we are observing.  This distinction is important because for
   large-latency links, we may obtain very different times depending on
   exactly where we are observing the link.  We could allow the
   observational position to be an arbitrary location along the link;
   however, we define it to be in terms of an Internet host because we
   anticipate in practice that, for IPPM metrics, all such timing will
   be constrained to be performed by Internet hosts, rather than
   specialized hardware devices that might be able to monitor a link at
   locations where a host cannot.  This definition also takes care of
   the problem of links that are comprised of multiple physical
   channels.  Because these multiple channels are not visible at the IP
   layer, they cannot be individually observed in terms of the above
   definitions.

   It is possible, though one hopes uncommon, that a packet P might make
   multiple trips over a particular link L, due to a forwarding loop.
   These trips might even overlap, depending on the link technology.
   Whenever this occurs, we define a separate wire time associated with
   each instance of P seen at H's position on the link.  This definition
   is worth making because it serves as a reminder that notions like
   *the* unique time a packet passes a point in the Internet are
   inherently slippery.

   The term wire time has historically been used to loosely denote the
   time at which a packet appeared on a link, without exactly specifying
   whether this refers to the first bit, the last bit, or some other
   consideration.  This informal definition is generally already very
   useful, as it is usually used to make a distinction between when the
   packet's propagation delays begin and cease to be due to the network
   rather than the endpoint hosts.

   When appropriate, metrics should be defined in terms of wire times
   rather than host endpoint times, so that the metric's definition
   highlights the issue of separating delays due to the host from those
   due to the network.

   We note that one potential difficulty when dealing with wire times
   concerns IP fragments.  It may be the case that, due to
   fragmentation, only a portion of a particular packet passes by H's
   location.  Such fragments are themselves legitimate packets and have
   well-defined wire times associated with them; but the larger IP
   packet corresponding to their aggregate may not.




Paxson, et. al.              Informational                     [Page 18]

RFC 2330          Framework for IP Performance Metrics          May 1998


   We also note that these notions have not, to our knowledge, been
   previously defined in exact terms for Internet traffic.
   Consequently, we may find with experience that these definitions
   require some adjustment in the future.

   {Comment: It can sometimes be difficult to measure wire times.  One
   technique is to use a packet filter to monitor traffic on a link.
   The architecture of these filters often attempts to associate with
   each packet a timestamp as close to the wire time as possible.  We
   note however that one common source of error is to run the packet
   filter on one of the endpoint hosts.  In this case, it has been
   observed that some packet filters receive for some packets timestamps
   corresponding to when the packet was *scheduled* to be injected into
   the network, rather than when it actually was *sent* out onto the
   network (wire time).  There can be a substantial difference between
   these two times.  A technique for dealing with this problem is to run
   the packet filter on a separate host that passively monitors the
   given link.  This can be problematic however for some link
   technologies.  See [Pa97] for a discussion of the sorts of errors
   packet filters can exhibit.  Finally, we note that packet filters
   will often only capture the first fragment of a fragmented IP packet,
   due to the use of filtering on fields in the IP and transport
   protocol headers.  As we generally desire our measurement
   methodologies to avoid the complexity of creating fragmented traffic,
   one strategy for dealing with their presence as detected by a packet
   filter is to flag that the measured traffic has an unusual form and
   abandon further analysis of the packet timing.}


11. Singletons, Samples, and Statistics



   With experience we have found it useful to introduce a separation
   between three distinct -- yet related -- notions:

 +    By a 'singleton' metric, we refer to metrics that are, in a sense,
      atomic.  For example, a single instance of "bulk throughput
      capacity" from one host to another might be defined as a singleton
      metric, even though the instance involves measuring the timing of
      a number of Internet packets.
 +    By a 'sample' metric, we refer to metrics derived from a given
      singleton metric by taking a number of distinct instances
      together.  For example, we might define a sample metric of one-way
      delays from one host to another as an hour's worth of
      measurements, each made at Poisson intervals with a mean spacing
      of one second.






Paxson, et. al.              Informational                     [Page 19]

RFC 2330          Framework for IP Performance Metrics          May 1998


 +    By a 'statistical' metric, we refer to metrics derived from a
      given sample metric by computing some statistic of the values
      defined by the singleton metric on the sample.  For example, the
      mean of all the one-way delay values on the sample given above
      might be defined as a statistical metric.

   By applying these notions of singleton, sample, and statistic in a
   consistent way, we will be able to reuse lessons learned about how to
   define samples and statistics on various metrics.  The orthogonality
   among these three notions will thus make all our work more effective
   and more intelligible by the community.

   In the remainder of this section, we will cover some topics in
   sampling and statistics that we believe will be important to a
   variety of metric definitions and measurement efforts.


11.1. Methods of Collecting Samples



   The main reason for collecting samples is to see what sort of
   variations and consistencies are present in the metric being
   measured.  These variations might be with respect to different points
   in the Internet, or different measurement times.  When assessing
   variations based on a sample, one generally makes an assumption that
   the sample is "unbiased", meaning that the process of collecting the
   measurements in the sample did not skew the sample so that it no
   longer accurately reflects the metric's variations and consistencies.

   One common way of collecting samples is to make measurements
   separated by fixed amounts of time: periodic sampling.  Periodic
   sampling is particularly attractive because of its simplicity, but it
   suffers from two potential problems:

 +    If the metric being measured itself exhibits periodic behavior,
      then there is a possibility that the sampling will observe only
      part of the periodic behavior if the periods happen to agree
      (either directly, or if one is a multiple of the other).  Related
      to this problem is the notion that periodic sampling can be easily
      anticipated.  Predictable sampling is susceptible to manipulation
      if there are mechanisms by which a network component's behavior
      can be temporarily changed such that the sampling only sees the
      modified behavior.
 +    The act of measurement can perturb what is being measured (for
      example, injecting measurement traffic into a network alters the
      congestion level of the network), and repeated periodic
      perturbations can drive a network into a state of synchronization
      (cf. [FJ94]), greatly magnifying what might individually be minor
      effects.



Paxson, et. al.              Informational                     [Page 20]

RFC 2330          Framework for IP Performance Metrics          May 1998


   A more sound approach is based on "random additive sampling": samples
   are separated by independent, randomly generated intervals that have
   a common statistical distribution G(t) [BM92].  The quality of this
   sampling depends on the distribution G(t).  For example, if G(t)
   generates a constant value g with probability one, then the sampling
   reduces to periodic sampling with a period of g.

   Random additive sampling gains significant advantages.  In general,
   it avoids synchronization effects and yields an unbiased estimate of
   the property being sampled.  The only significant drawbacks with it
   are:

 +    it complicates frequency-domain analysis, because the samples do
      not occur at fixed intervals such as assumed by Fourier-transform
      techniques; and
 +    unless G(t) is the exponential distribution (see below), sampling
      still remains somewhat predictable, as discussed for periodic
      sampling above.


11.1.1. Poisson Sampling



   It can be proved that if G(t) is an exponential distribution with
   rate lambda, that is

       G(t) = 1 - exp(-lambda * t)

   then the arrival of new samples *cannot* be predicted (and, again,
   the sampling is unbiased).  Furthermore, the sampling is
   asymptotically unbiased even if the act of sampling affects the
   network's state.  Such sampling is referred to as "Poisson sampling".
   It is not prone to inducing synchronization, it can be used to
   accurately collect measurements of periodic behavior, and it is not
   prone to manipulation by anticipating when new samples will occur.

   Because of these valuable properties, we in general prefer that
   samples of Internet measurements are gathered using Poisson sampling.
   {Comment: We note, however, that there may be circumstances that
   favor use of a different G(t).  For example, the exponential
   distribution is unbounded, so its use will on occasion generate
   lengthy spaces between sampling times.  We might instead desire to
   bound the longest such interval to a maximum value dT, to speed the
   convergence of the estimation derived from the sampling.  This could
   be done by using

       G(t) = Unif(0, dT)





Paxson, et. al.              Informational                     [Page 21]

RFC 2330          Framework for IP Performance Metrics          May 1998


   that is, the uniform distribution between 0 and dT.  This sampling,
   of course, becomes highly predictable if an interval of nearly length
   dT has elapsed without a sample occurring.}

   In its purest form, Poisson sampling is done by generating
   independent, exponentially distributed intervals and gathering a
   single measurement after each interval has elapsed.  It can be shown
   that if starting at time T one performs Poisson sampling over an
   interval dT, during which a total of N measurements happen to be
   made, then those measurements will be uniformly distributed over the
   interval [T, T+dT].  So another way of conducting Poisson sampling is
   to pick dT and N and generate N random sampling times uniformly over
   the interval [T, T+dT].  The two approaches are equivalent, except if
   N and dT are externally known.  In that case, the property of not
   being able to predict measurement times is weakened (the other
   properties still hold).  The N/dT approach has an advantage that
   dealing with fixed values of N and dT can be simpler than dealing
   with a fixed lambda but variable numbers of measurements over
   variably-sized intervals.


11.1.2. Geometric Sampling



   Closely related to Poisson sampling is "geometric sampling", in which
   external events are measured with a fixed probability p.  For
   example, one might capture all the packets over a link but only
   record the packet to a trace file if a randomly generated number
   uniformly distributed between 0 and 1 is less than a given p.
   Geometric sampling has the same properties of being unbiased and not
   predictable in advance as Poisson sampling, so if it fits a
   particular Internet measurement task, it too is sound.  See [CPB93]
   for more discussion.


11.1.3. Generating Poisson Sampling Intervals



   To generate Poisson sampling intervals, one first determines the rate
   lambda at which the singleton measurements will on average be made
   (e.g., for an average sampling interval of 30 seconds, we have lambda
   = 1/30, if the units of time are seconds).  One then generates a
   series of exponentially-distributed (pseudo) random numbers E1, E2,
   ..., En.  The first measurement is made at time E1, the next at time
   E1+E2, and so on.

   One technique for generating exponentially-distributed (pseudo)
   random numbers is based on the ability to generate U1, U2, ..., Un,
   (pseudo) random numbers that are uniformly distributed between 0 and
   1.  Many computers provide libraries that can do this.  Given such



Paxson, et. al.              Informational                     [Page 22]

RFC 2330          Framework for IP Performance Metrics          May 1998


   Ui, to generate Ei one uses:

       Ei = -log(Ui) / lambda

   where log(Ui) is the natural logarithm of Ui.  {Comment: This
   technique is an instance of the more general "inverse transform"
   method for generating random numbers with a given distribution.}

   Implementation details:

   There are at least three different methods for approximating Poisson
   sampling, which we describe here as Methods 1 through 3.  Method 1 is
   the easiest to implement and has the most error, and method 3 is the
   most difficult to implement and has the least error (potentially
   none).

   Method 1 is to proceed as follows:

   1.  Generate E1 and wait that long.
   2.  Perform a measurement.
   3.  Generate E2 and wait that long.
   4.  Perform a measurement.
   5.  Generate E3 and wait that long.
   6.  Perform a measurement ...

   The problem with this approach is that the "Perform a measurement"
   steps themselves take time, so the sampling is not done at times E1,
   E1+E2, etc., but rather at E1, E1+M1+E2, etc., where Mi is the amount
   of time required for the i'th measurement.  If Mi is very small
   compared to 1/lambda then the potential error introduced by this
   technique is likewise small.  As Mi becomes a non-negligible fraction
   of 1/lambda, the potential error increases.

   Method 2 attempts to correct this error by taking into account the
   amount of time required by the measurements (i.e., the Mi's) and
   adjusting the waiting intervals accordingly:

   1.  Generate E1 and wait that long.
   2.  Perform a measurement and measure M1, the time it took to do so.
   3.  Generate E2 and wait for a time E2-M1.
   4.  Perform a measurement and measure M2 ..

   This approach works fine as long as E{i+1} >= Mi.  But if E{i+1} < Mi
   then it is impossible to wait the proper amount of time.  (Note that
   this case corresponds to needing to perform two measurements
   simultaneously.)





Paxson, et. al.              Informational                     [Page 23]

RFC 2330          Framework for IP Performance Metrics          May 1998


   Method 3 is generating a schedule of measurement times E1, E1+E2,
   etc., and then sticking to it:

   1.  Generate E1, E2, ..., En.
   2.  Compute measurement times T1, T2, ..., Tn, as Ti = E1 + ... + Ei.
   3.  Arrange that at times T1, T2, ..., Tn, a measurement is made.

   By allowing simultaneous measurements, Method 3 avoids the
   shortcomings of Methods 1 and 2.  If, however, simultaneous
   measurements interfere with one another, then Method 3 does not gain
   any benefit and may actually prove worse than Methods 1 or 2.

   For Internet phenomena, it is not known to what degree the
   inaccuracies of these methods are significant.  If the Mi's are much
   less than 1/lambda, then any of the three should suffice.  If the
   Mi's are less than 1/lambda but perhaps not greatly less, then Method
   2 is preferred to Method 1.  If simultaneous measurements do not
   interfere with one another, then Method 3 is preferred, though it can
   be considerably harder to implement.


11.2. Self-Consistency



   A fundamental requirement for a sound measurement methodology is that
   measurement be made using as few unconfirmed assumptions as possible.
   Experience has painfully shown how easy it is to make an (often
   implicit) assumption that turns out to be incorrect.  An example is
   incorporating into a measurement the reading of a clock synchronized
   to a highly accurate source.  It is easy to assume that the clock is
   therefore accurate; but due to software bugs, a loss of power in the
   source, or a loss of communication between the source and the clock,
   the clock could actually be quite inaccurate.

   This is not to argue that one must not make *any* assumptions when
   measuring, but rather that, to the extent which is practical,
   assumptions should be tested.  One powerful way for doing so involves
   checking for self-consistency.  Such checking applies both to the
   observed value(s) of the measurement *and the values used by the
   measurement process itself*.  A simple example of the former is that
   when computing a round trip time, one should check to see if it is
   negative.  Since negative time intervals are non-physical, if it ever
   is negative that finding immediately flags an error.  *These sorts of
   errors should then be investigated!* It is crucial to determine where
   the error lies, because only by doing so diligently can we build up
   faith in a methodology's fundamental soundness.  For example, it
   could be that the round trip time is negative because during the
   measurement the clock was set backward in the process of
   synchronizing it with another source.  But it could also be that the



Paxson, et. al.              Informational                     [Page 24]

RFC 2330          Framework for IP Performance Metrics          May 1998


   measurement program accesses uninitialized memory in one of its
   computations and, only very rarely, that leads to a bogus
   computation.  This second error is more serious, if the same program
   is used by others to perform the same measurement, since then they
   too will suffer from incorrect results.  Furthermore, once uncovered
   it can be completely fixed.

   A more subtle example of testing for self-consistency comes from
   gathering samples of one-way Internet delays.  If one has a large
   sample of such delays, it may well be highly telling to, for example,
   fit a line to the pairs of (time of measurement, measured delay), to
   see if the resulting line has a clearly non-zero slope.  If so, a
   possible interpretation is that one of the clocks used in the
   measurements is skewed relative to the other.  Another interpretation
   is that the slope is actually due to genuine network effects.
   Determining which is indeed the case will often be highly
   illuminating.  (See [Pa97] for a discussion of distinguishing between
   relative clock skew and genuine network effects.) Furthermore, if
   making this check is part of the methodology, then a finding that the
   long-term slope is very near zero is positive evidence that the
   measurements are probably not biased by a difference in skew.

   A final example illustrates checking the measurement process itself
   for self-consistency.  Above we outline Poisson sampling techniques,
   based on generating exponentially-distributed intervals.  A sound
   measurement methodology would include testing the generated intervals
   to see whether they are indeed exponentially distributed (and also to
   see if they suffer from correlation).  In the appendix we discuss and
   give C code for one such technique, a general-purpose, well-regarded
   goodness-of-fit test called the Anderson-Darling test.

   Finally, we note that what is truly relevant for Poisson sampling of
   Internet metrics is often not when the measurements began but the
   wire times corresponding to the measurement process.  These could
   well be different, due to complications on the hosts used to perform
   the measurement.  Thus, even those with complete faith in their
   pseudo-random number generators and subsequent algorithms are
   encouraged to consider how they might test the assumptions of each
   measurement procedure as much as possible.


11.3. Defining Statistical Distributions



   One way of describing a collection of measurements (a sample) is as a
   statistical distribution -- informally, as percentiles.  There are
   several slightly different ways of doing so.  In this section we
   define a standard definition to give uniformity to these
   descriptions.



Paxson, et. al.              Informational                     [Page 25]

RFC 2330          Framework for IP Performance Metrics          May 1998


   The "empirical distribution function" (EDF) of a set of scalar
   measurements is a function F(x) which for any x gives the fractional
   proportion of the total measurements that were <= x.  If x is less
   than the minimum value observed, then F(x) is 0.  If it is greater or
   equal to the maximum value observed, then F(x) is 1.

   For example, given the 6 measurements:

   -2, 7, 7, 4, 18, -5

   Then F(-8) = 0, F(-5) = 1/6, F(-5.0001) = 0, F(-4.999) = 1/6, F(7) =
   5/6, F(18) = 1, F(239) = 1.

   Note that we can recover the different measured values and how many
   times each occurred from F(x) -- no information regarding the range
   in values is lost.  Summarizing measurements using histograms, on the
   other hand, in general loses information about the different values
   observed, so the EDF is preferred.

   Using either the EDF or a histogram, however, we do lose information
   regarding the order in which the values were observed.  Whether this
   loss is potentially significant will depend on the metric being
   measured.

   We will use the term "percentile" to refer to the smallest value of x
   for which F(x) >= a given percentage.  So the 50th percentile of the
   example above is 4, since F(4) = 3/6 = 50%; the 25th percentile is
   -2, since F(-5) = 1/6 < 25%, and F(-2) = 2/6 >= 25%; the 100th
   percentile is 18; and the 0th percentile is -infinity, as is the 15th
   percentile.

   Care must be taken when using percentiles to summarize a sample,
   because they can lend an unwarranted appearance of more precision
   than is really available.  Any such summary must include the sample
   size N, because any percentile difference finer than 1/N is below the
   resolution of the sample.

   See [DS86] for more details regarding EDF's.

   We close with a note on the common (and important!) notion of median.
   In statistics, the median of a distribution is defined to be the
   point X for which the probability of observing a value <= X is equal
   to the probability of observing a value > X.  When estimating the
   median of a set of observations, the estimate depends on whether the
   number of observations, N, is odd or even:






Paxson, et. al.              Informational                     [Page 26]

RFC 2330          Framework for IP Performance Metrics          May 1998


 +    If N is odd, then the 50th percentile as defined above is used as
      the estimated median.
 +    If N is even, then the estimated median is the average of the
      central two observations; that is, if the observations are sorted
      in ascending order and numbered from 1 to N, where N = 2*K, then
      the estimated median is the average of the (K)'th and (K+1)'th
      observations.

   Usually the term "estimated" is dropped from the phrase "estimated
   median" and this value is simply referred to as the "median".


11.4. Testing For Goodness-of-Fit



   For some forms of measurement calibration we need to test whether a
   set of numbers is consistent with those numbers having been drawn
   from a particular distribution.  An example is that to apply a self-
   consistency check to measurements made using a Poisson process, one
   test is to see whether the spacing between the sampling times does
   indeed reflect an exponential distribution; or if the dT/N approach
   discussed above was used, whether the times are uniformly distributed
   across [T, dT].

   {Comment: There are at least three possible sets of values we could
   test: the scheduled packet transmission times, as determined by use
   of a pseudo-random number generator; user-level timestamps made just
   before or after the system call for transmitting the packet; and wire
   times for the packets as recorded using a packet filter.  All three
   of these are potentially informative: failures for the scheduled
   times to match an exponential distribution indicate inaccuracies in
   the random number generation; failures for the user-level times
   indicate inaccuracies in the timers used to schedule transmission;
   and failures for the wire times indicate inaccuracies in actually
   transmitting the packets, perhaps due to contention for a shared
   resource.}

   There are a large number of statistical goodness-of-fit techniques
   for performing such tests.  See [DS86] for a thorough discussion.
   That reference recommends the Anderson-Darling EDF test as being a
   good all-purpose test, as well as one that is especially good at
   detecting deviations from a given distribution in the lower and upper
   tails of the EDF.

   It is important to understand that the nature of goodness-of-fit
   tests is that one first selects a "significance level", which is the
   probability that the test will erroneously declare that the EDF of a
   given set of measurements fails to match a particular distribution
   when in fact the measurements do indeed reflect that distribution.



Paxson, et. al.              Informational                     [Page 27]

RFC 2330          Framework for IP Performance Metrics          May 1998


   Unless otherwise stated, IPPM goodness-of-fit tests are done using 5%
   significance.  This means that if the test is applied to 100 samples
   and 5 of those samples are deemed to have failed the test, then the
   samples are all consistent with the distribution being tested.  If
   significantly more of the samples fail the test, then the assumption
   that the samples are consistent with the distribution being tested
   must be rejected.  If significantly fewer of the samples fail the
   test, then the samples have potentially been doctored too well to fit
   the distribution.  Similarly, some goodness-of-fit tests (including
   Anderson-Darling) can detect whether it is likely that a given sample
   was doctored.  We also use a significance of 5% for this case; that
   is, the test will report that a given honest sample is "too good to
   be true" 5% of the time, so if the test reports this finding
   significantly more often than one time out of twenty, it is an
   indication that something unusual is occurring.

   The appendix gives sample C code for implementing the Anderson-
   Darling test, as well as further discussing its use.

   See [Pa94] for a discussion of goodness-of-fit and closeness-of-fit
   tests in the context of network measurement.


12. Avoiding Stochastic Metrics



   When defining metrics applying to a path, subpath, cloud, or other
   network element, we in general do not define them in stochastic terms
   (probabilities).  We instead prefer a deterministic definition.  So,
   for example, rather than defining a metric about a "packet loss
   probability between A and B", we would define a metric about a
   "packet loss rate between A and B".  (A measurement given by the
   first definition might be "0.73", and by the second "73 packets out
   of 100".)

   We emphasize that the above distinction concerns the *definitions* of
   *metrics*.  It is not intended to apply to what sort of techniques we
   might use to analyze the results of measurements.

   The reason for this distinction is as follows.  When definitions are
   made in terms of probabilities, there are often hidden assumptions in
   the definition about a stochastic model of the behavior being
   measured.  The fundamental goal with avoiding probabilities in our
   metric definitions is to avoid biasing our definitions by these
   hidden assumptions.







Paxson, et. al.              Informational                     [Page 28]

RFC 2330          Framework for IP Performance Metrics          May 1998


   For example, an easy hidden assumption to make is that packet loss in
   a network component due to queueing overflows can be described as
   something that happens to any given packet with a particular
   probability.  In today's Internet, however, queueing drops are
   actually usually *deterministic*, and assuming that they should be
   described probabilistically can obscure crucial correlations between
   queueing drops among a set of packets.  So it's better to explicitly
   note stochastic assumptions, rather than have them sneak into our
   definitions implicitly.

   This does *not* mean that we abandon stochastic models for
   *understanding* network performance! It only means that when defining
   IP metrics we avoid terms such as "probability" for terms like
   "proportion" or "rate".  We will still use, for example, random
   sampling in order to estimate probabilities used by stochastic models
   related to the IP metrics.  We also do not rule out the possibility
   of stochastic metrics when they are truly appropriate (for example,
   perhaps to model transmission errors caused by certain types of line
   noise).


13. Packets of Type P



   A fundamental property of many Internet metrics is that the value of
   the metric depends on the type of IP packet(s) used to make the
   measurement.  Consider an IP-connectivity metric: one obtains
   different results depending on whether one is interested in
   connectivity for packets destined for well-known TCP ports or
   unreserved UDP ports, or those with invalid IP checksums, or those
   with TTL's of 16, for example.  In some circumstances these
   distinctions will be highly interesting (for example, in the presence
   of firewalls, or RSVP reservations).

   Because of this distinction, we introduce the generic notion of a
   "packet of type P", where in some contexts P will be explicitly
   defined (i.e., exactly what type of packet we mean), partially
   defined (e.g., "with a payload of B octets"), or left generic.  Thus
   we may talk about generic IP-type-P-connectivity or more specific
   IP-port-HTTP-connectivity.  Some metrics and methodologies may be
   fruitfully defined using generic type P definitions which are then
   made specific when performing actual measurements.

   Whenever a metric's value depends on the type of the packets involved
   in the metric, the metric's name will include either a specific type
   or a phrase such as "type-P".  Thus we will not define an "IP-






Paxson, et. al.              Informational                     [Page 29]

RFC 2330          Framework for IP Performance Metrics          May 1998


   connectivity" metric but instead an "IP-type-P-connectivity" metric
   and/or perhaps an "IP-port-HTTP-connectivity" metric.  This naming
   convention serves as an important reminder that one must be conscious
   of the exact type of traffic being measured.

   A closely related note: it would be very useful to know if a given
   Internet component treats equally a class C of different types of
   packets.  If so, then any one of those types of packets can be used
   for subsequent measurement of the component.  This suggests we devise
   a metric or suite of metrics that attempt to determine C.


14. Internet Addresses vs. Hosts



   When considering a metric for some path through the Internet, it is
   often natural to think about it as being for the path from Internet
   host H1 to host H2.  A definition in these terms, though, can be
   ambiguous, because Internet hosts can be attached to more than one
   network.  In this case, the result of the metric will depend on which
   of these networks is actually used.

   Because of this ambiguity, usually such definitions should instead be
   defined in terms of Internet IP addresses.  For the common case of a
   unidirectional path through the Internet, we will use the term "Src"
   to denote the IP address of the beginning of the path, and "Dst" to
   denote the IP address of the end.


15. Standard-Formed Packets



   Unless otherwise stated, all metric definitions that concern IP
   packets include an implicit assumption that the packet is *standard
   formed*.  A packet is standard formed if it meets all of the
   following criteria:

 +    Its length as given in the IP header corresponds to the size of
      the IP header plus the size of the payload.
 +    It includes a valid IP header: the version field is 4 (later, we
      will expand this to include 6); the header length is >= 5; the
      checksum is correct.
 +    It is not an IP fragment.
 +    The source and destination addresses correspond to the hosts in
      question.








Paxson, et. al.              Informational                     [Page 30]

RFC 2330          Framework for IP Performance Metrics          May 1998


 +    Either the packet possesses sufficient TTL to travel from the
      source to the destination if the TTL is decremented by one at each
      hop, or it possesses the maximum TTL of 255.
 +    It does not contain IP options unless explicitly noted.
 +    If a transport header is present, it too contains a valid checksum
      and other valid fields.

   We further require that if a packet is described as having a "length
   of B octets", then 0 <= B <= 65535; and if B is the payload length in
   octets, then B <= (65535-IP header size in octets).

   So, for example, one might imagine defining an IP connectivity metric
   as "IP-type-P-connectivity for standard-formed packets with the IP
   TOS field set to 0", or, more succinctly, "IP-type-P-connectivity
   with the IP TOS field set to 0", since standard-formed is already
   implied by convention.

   A particular type of standard-formed packet often useful to consider
   is the "minimal IP packet from A to B" - this is an IP packet with
   the following properties:

 +    It is standard-formed.
 +    Its data payload is 0 octets.
 +    It contains no options.

   (Note that we do not define its protocol field, as different values
   may lead to different treatment by the network.)

   When defining IP metrics we keep in mind that no packet smaller or
   simpler than this can be transmitted over a correctly operating IP
   network.


16. Acknowledgements



   The comments of Brian Carpenter, Bill Cerveny, Padma Krishnaswamy
   Jeff Sedayao and Howard Stanislevic are appreciated.


17. Security Considerations



   This document concerns definitions and concepts related to Internet
   measurement.  We discuss measurement procedures only in high-level
   terms, regarding principles that lend themselves to sound
   measurement.  As such, the topics discussed do not affect the
   security of the Internet or of applications which run on it.





Paxson, et. al.              Informational                     [Page 31]

RFC 2330          Framework for IP Performance Metrics          May 1998


   That said, it should be recognized that conducting Internet
   measurements can raise both security and privacy concerns.  Active
   techniques, in which traffic is injected into the network, can be
   abused for denial-of-service attacks disguised as legitimate
   measurement activity.  Passive techniques, in which existing traffic
   is recorded and analyzed, can expose the contents of Internet traffic
   to unintended recipients.  Consequently, the definition of each
   metric and methodology must include a corresponding discussion of
   security considerations.


18. Appendix



   Below we give routines written in C for computing the Anderson-
   Darling test statistic (A2) for determining whether a set of values
   is consistent with a given statistical distribution.  Externally, the
   two main routines of interest are:

       double exp_A2_known_mean(double x[], int n, double mean)
       double unif_A2_known_range(double x[], int n,
                                  double min_val, double max_val)

   Both take as their first argument, x, the array of n values to be
   tested.  (Upon return, the elements of x are sorted.)  The remaining
   parameters characterize the distribution to be used: either the mean
   (1/lambda), for an exponential distribution, or the lower and upper
   bounds, for a uniform distribution.  The names of the routines stress
   that these values must be known in advance, and *not* estimated from
   the data (for example, by computing its sample mean).  Estimating the
   parameters from the data *changes* the significance level of the test
   statistic.  While [DS86] gives alternate significance tables for some
   instances in which the parameters are estimated from the data, for
   our purposes we expect that we should indeed know the parameters in
   advance, since what we will be testing are generally values such as
   packet sending times that we wish to verify follow a known
   distribution.

   Both routines return a significance level, as described earlier. This
   is a value between 0 and 1.  The correct use of the routines is to
   pick in advance the threshold for the significance level to test;
   generally, this will be 0.05, corresponding to 5%, as also described
   above.  Subsequently, if the routines return a value strictly less
   than this threshold, then the data are deemed to be inconsistent with
   the presumed distribution, *subject to an error corresponding to the
   significance level*.  That is, for a significance level of 5%, 5% of
   the time data that is indeed drawn from the presumed distribution
   will be erroneously deemed inconsistent.




Paxson, et. al.              Informational                     [Page 32]

RFC 2330          Framework for IP Performance Metrics          May 1998


   Thus, it is important to bear in mind that if these routines are used
   frequently, then one will indeed encounter occasional failures, even
   if the data is unblemished.

   Another important point concerning significance levels is that it is
   unsound to compare them in order to determine which of two sets of
   values is a "better" fit to a presumed distribution.  Such testing
   should instead be done using "closeness-of-fit metrics" such as the
   lambda^2 metric described in [Pa94].

   While the routines provided are for exponential and uniform
   distributions with known parameters, it is generally straight-forward
   to write comparable routines for any distribution with known
   parameters.  The heart of the A2 tests lies in a statistic computed
   for testing whether a set of values is consistent with a uniform
   distribution between 0 and 1, which we term Unif(0, 1).  If we wish
   to test whether a set of values, X, is consistent with a given
   distribution G(x), we first compute
       Y = G_inverse(X)
   If X is indeed distributed according to G(x), then Y will be
   distributed according to Unif(0, 1); so by testing Y for consistency
   with Unif(0, 1), we also test X for consistency with G(x).

   We note, however, that the process of computing Y above might yield
   values of Y outside the range (0..1).  Such values should not occur
   if X is indeed distributed according to G(x), but easily can occur if
   it is not.  In the latter case, we need to avoid computing the
   central A2 statistic, since floating-point exceptions may occur if
   any of the values lie outside (0..1).  Accordingly, the routines
   check for this possibility, and if encountered, return a raw A2
   statistic of -1.  The routine that converts the raw A2 statistic to a
   significance level likewise propagates this value, returning a
   significance level of -1.  So, any use of these routines must be
   prepared for a possible negative significance level.

   The last important point regarding use of A2 statistic concerns n,
   the number of values being tested.  If n < 5 then the test is not
   meaningful, and in this case a significance level of -1 is returned.

   On the other hand, for "real" data the test *gains* power as n
   becomes larger.  It is well known in the statistics community that
   real data almost never exactly matches a theoretical distribution,
   even in cases such as rolling dice a great many times (see [Pa94] for
   a brief discussion and references).  The A2 test is sensitive enough
   that, for sufficiently large sets of real data, the test will almost
   always fail, because it will manage to detect slight imperfections in
   the fit of the data to the distribution.




Paxson, et. al.              Informational                     [Page 33]

RFC 2330          Framework for IP Performance Metrics          May 1998


   For example, we have found that when testing 8,192 measured wire
   times for packets sent at Poisson intervals, the measurements almost
   always fail the A2 test.  On the other hand, testing 128 measurements
   failed at 5% significance only about 5% of the time, as expected.
   Thus, in general, when the test fails, care must be taken to
   understand why it failed.

   The remainder of this appendix gives C code for the routines
   mentioned above.

   /* Routines for computing the Anderson-Darling A2 test statistic.
    *
    * Implemented based on the description in "Goodness-of-Fit
    * Techniques," R. D'Agostino and M. Stephens, editors,
    * Marcel Dekker, Inc., 1986.
    */

   #include <stdio.h>
   #include <stdlib.h>
   #include <math.h>

   /* Returns the raw A^2 test statistic for n sorted samples
    * z[0] .. z[n-1], for z ~ Unif(0,1).
    */
   extern double compute_A2(double z[], int n);

   /* Returns the significance level associated with a A^2 test
    * statistic value of A2, assuming no parameters of the tested
    * distribution were estimated from the data.
    */
   extern double A2_significance(double A2);

   /* Returns the A^2 significance level for testing n observations
    * x[0] .. x[n-1] against an exponential distribution with the
    * given mean.
    *
    * SIDE EFFECT: the x[0..n-1] are sorted upon return.
    */
   extern double exp_A2_known_mean(double x[], int n, double mean);

   /* Returns the A^2 significance level for testing n observations
    * x[0] .. x[n-1] against the uniform distribution [min_val, max_val].
    *
    * SIDE EFFECT: the x[0..n-1] are sorted upon return.
    */
   extern double unif_A2_known_range(double x[], int n,
                       double min_val, double max_val);




Paxson, et. al.              Informational                     [Page 34]

RFC 2330          Framework for IP Performance Metrics          May 1998


   /* Returns a pseudo-random number distributed according to an
    * exponential distribution with the given mean.
    */
   extern double random_exponential(double mean);


   /* Helper function used by qsort() to sort double-precision
    * floating-point values.
    */
   static int
   compare_double(const void *v1, const void *v2)
   {
       double d1 = *(double *) v1;
       double d2 = *(double *) v2;

       if (d1 < d2)
           return -1;
       else if (d1 > d2)
           return 1;
       else
           return 0;
   }

   double
   compute_A2(double z[], int n)
   {
       int i;
       double sum = 0.0;

       if ( n < 5 )
           /* Too few values. */
           return -1.0;

       /* If any of the values are outside the range (0, 1) then
        * fail immediately (and avoid a possible floating point
        * exception in the code below).
        */
       for (i = 0; i < n; ++i)
           if ( z[i] <= 0.0 || z[i] >= 1.0 )
               return -1.0;

       /* Page 101 of D'Agostino and Stephens. */
       for (i = 1; i <= n; ++i) {
           sum += (2 * i - 1) * log(z[i-1]);
           sum += (2 * n + 1 - 2 * i) * log(1.0 - z[i-1]);
       }
       return -n - (1.0 / n) * sum;
   }



Paxson, et. al.              Informational                     [Page 35]

RFC 2330          Framework for IP Performance Metrics          May 1998


   double
   A2_significance(double A2)
   {
       /* Page 105 of D'Agostino and Stephens. */
       if (A2 < 0.0)
           return A2;    /* Bogus A2 value - propagate it. */

       /* Check for possibly doctored values. */
       if (A2 <= 0.201)
           return 0.99;
       else if (A2 <= 0.240)
           return 0.975;
       else if (A2 <= 0.283)
           return 0.95;
       else if (A2 <= 0.346)
           return 0.90;
       else if (A2 <= 0.399)
           return 0.85;

       /* Now check for possible inconsistency. */
       if (A2 <= 1.248)
           return 0.25;
       else if (A2 <= 1.610)
           return 0.15;
       else if (A2 <= 1.933)
           return 0.10;
       else if (A2 <= 2.492)
           return 0.05;
       else if (A2 <= 3.070)
           return 0.025;
       else if (A2 <= 3.880)
           return 0.01;
       else if (A2 <= 4.500)
           return 0.005;
       else if (A2 <= 6.000)
           return 0.001;
       else
           return 0.0;
   }

   double
   exp_A2_known_mean(double x[], int n, double mean)
   {
       int i;
       double A2;

       /* Sort the first n values. */
       qsort(x, n, sizeof(x[0]), compare_double);



Paxson, et. al.              Informational                     [Page 36]

RFC 2330          Framework for IP Performance Metrics          May 1998


       /* Assuming they match an exponential distribution, transform
        * them to Unif(0,1).
        */
       for (i = 0; i < n; ++i) {
           x[i] = 1.0 - exp(-x[i] / mean);
       }

       /* Now make the A^2 test to see if they're truly uniform. */
       A2 = compute_A2(x, n);
       return A2_significance(A2);
   }

   double
   unif_A2_known_range(double x[], int n, double min_val, double max_val)
   {
       int i;
       double A2;
       double range = max_val - min_val;

       /* Sort the first n values. */
       qsort(x, n, sizeof(x[0]), compare_double);

       /* Transform Unif(min_val, max_val) to Unif(0,1). */
       for (i = 0; i < n; ++i)
           x[i] = (x[i] - min_val) / range;

       /* Now make the A^2 test to see if they're truly uniform. */
       A2 = compute_A2(x, n);
       return A2_significance(A2);
   }

   double
   random_exponential(double mean)
   {
       return -mean * log1p(-drand48());
   }















Paxson, et. al.              Informational                     [Page 37]

RFC 2330          Framework for IP Performance Metrics          May 1998


19. References



   [AK97] G. Almes and S. Kalidindi, "A One-way Delay Metric for IPPM",
   Work in Progress, November 1997.

   [BM92] I. Bilinskis and A. Mikelsons, Randomized Signal Processing,
   Prentice Hall International, 1992.

   [DS86] R. D'Agostino and M. Stephens, editors, Goodness-of-Fit
   Techniques, Marcel Dekker, Inc., 1986.

   [CPB93] K. Claffy, G. Polyzos, and H-W. Braun, "Application of
   Sampling Methodologies to Network Traffic Characterization," Proc.
   SIGCOMM '93, pp. 194-203, San Francisco, September 1993.

   [FJ94] S. Floyd and V. Jacobson, "The Synchronization of Periodic
   Routing Messages," IEEE/ACM Transactions on Networking, 2(2), pp.
   122-136, April 1994.

   [Mi92] Mills, D., "Network Time Protocol (Version 3) Specification,
   Implementation and Analysis", RFC 1305, March 1992.

   [Pa94] V. Paxson, "Empirically-Derived Analytic Models of Wide-Area
   TCP Connections," IEEE/ACM Transactions on Networking, 2(4), pp.
   316-336, August 1994.

   [Pa96] V. Paxson, "Towards a Framework for Defining Internet
   Performance Metrics," Proceedings of INET '96,
   ftp://ftp.ee.lbl.gov/papers/metrics-framework-INET96.ps.Z

   [Pa97] V. Paxson, "Measurements and Analysis of End-to-End Internet
   Dynamics," Ph.D. dissertation, U.C. Berkeley, 1997,
   ftp://ftp.ee.lbl.gov/papers/vp-thesis/dis.ps.gz.


















Paxson, et. al.              Informational                     [Page 38]

RFC 2330          Framework for IP Performance Metrics          May 1998


20. Authors' Addresses



   Vern Paxson
   MS 50B/2239
   Lawrence Berkeley National Laboratory
   University of California
   Berkeley, CA  94720
   USA

   Phone: +1 510/486-7504
   EMail: vern@ee.lbl.gov


   Guy Almes
   Advanced Network & Services, Inc.
   200 Business Park Drive
   Armonk, NY  10504
   USA

   Phone: +1 914/765-1120
   EMail: almes@advanced.org


   Jamshid Mahdavi
   Pittsburgh Supercomputing Center
   4400 5th Avenue
   Pittsburgh, PA  15213
   USA

   Phone: +1 412/268-6282
   EMail: mahdavi@psc.edu


   Matt Mathis
   Pittsburgh Supercomputing Center
   4400 5th Avenue
   Pittsburgh, PA  15213
   USA

   Phone: +1 412/268-3319
   EMail: mathis@psc.edu










Paxson, et. al.              Informational                     [Page 39]

RFC 2330          Framework for IP Performance Metrics          May 1998


21. Full Copyright Statement



   Copyright (C) The Internet Society (1998).  All Rights Reserved.

   This document and translations of it may be copied and furnished to
   others, and derivative works that comment on or otherwise explain it
   or assist in its implementation may be prepared, copied, published
   and distributed, in whole or in part, without restriction of any
   kind, provided that the above copyright notice and this paragraph are
   included on all such copies and derivative works.  However, this
   document itself may not be modified in any way, such as by removing
   the copyright notice or references to the Internet Society or other
   Internet organizations, except as needed for the purpose of
   developing Internet standards in which case the procedures for
   copyrights defined in the Internet Standards process must be
   followed, or as required to translate it into languages other than
   English.

   The limited permissions granted above are perpetual and will not be
   revoked by the Internet Society or its successors or assigns.

   This document and the information contained herein is provided on an
   "AS IS" basis and THE INTERNET SOCIETY AND THE INTERNET ENGINEERING
   TASK FORCE DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING
   BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION
   HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF
   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
























Paxson, et. al.              Informational                     [Page 40]