One document matched: draft-briscoe-tsvwg-relax-fairness-00.xml


<?xml version="1.0" encoding="US-ASCII"?>
<!DOCTYPE rfc SYSTEM "rfc2629.dtd">
<!-- Alterations to I-D/RFC boilerplate -->
<?rfc private=""?>
<!-- Default private="" Produce an internal memo 2.5pp shorter than an I-D or RFC -->
<?rfc topblock="yes" ?>
<!-- Default topblock="yes" put the famous header block on the first page -->
<?rfc footer="" ?>
<!-- Default footer="Expires <date>" override the center footer string -->
<?rfc header="" ?>
<!-- Default header="Internet-Draft" override the leftmost header string -->
<?rfc authorship="yes" ?>
<!-- Default authorship="yes" Render authors' addresses section -->
<?rfc rfcprocack="yes" ?>
<!-- Default rfcprocack="no" add a short sentence acknowledging xml2rfc -->
<?rfc strict="no" ?>
<!-- Default strict="no" Don't check I-D nits -->
<?rfc rfcedstyle="no" ?>
<!-- Default rfcedstyle="yes" attempt to closely follow finer details from the latest observable RFC-Editor style -->
<?rfc colonspace="no" ?>
<!-- Default colonspace="no" put two spaces instead of one after each colon (``:'') in txt or nroff files -->
<?rfc notedraftinprogress="yes" ?>
<!-- Default notedraftinprogress="yes" generates "(work in progress)", as appropriate -->
<?rfc refparent="References" ?>
<!-- Default refparent="References" title of the top-level section containing all references -->
<!-- IETF process -->
<?rfc iprnotified="no" ?>
<!-- Default iprnotified="no" I haven't disclosed existence of IPR to IETF -->
<!-- ToC format -->
<?rfc toc="yes" ?>
<!-- Default toc="no" No Table of Contents -->
<?rfc tocappendix="yes" ?>
<!-- Default tocappendix="yes" control whether the word `Appendix' appears in the table-of-content -->
<?rfc tocdepth="3" ?>
<!-- Default tocdepth="3" if toc is "yes", then this determines the depth of the table-of-contents -->
<?rfc tocindent="yes" ?>
<!-- Default tocindent="yes" if toc is "yes", indent subsections in the table-of-contents -->
<?rfc tocnarrow="yes" ?>
<!-- Default tocnarrow="yes" affects horizontal spacing in the table-of-content -->
<?rfc tocompact="yes" ?>
<!-- Default tocompact="yes" if toc is "yes", then setting this to "no" will make it a little less compact -->
<!-- Cross referencing, footnotes, comments -->
<?rfc symrefs="yes" ?>
<!-- Default symrefs="no" Don't use anchors, but use numbers for refs -->
<?rfc sortrefs="yes"?>
<!-- Default sortrefs="no" Don't sort references into order -->
<?rfc comments="yes" ?>
<!-- Default comments="no" Don't render comments -->
<?rfc inline="no"?>
<!-- Default inline="no" if comments is "yes", then render comments inline; otherwise render them in an `Editorial Comments' section -->
<?rfc editing="no" ?>
<!-- Default editing="no" Don't insert editing marks for ease of discussing draft versions -->
<!-- Pagination control -->
<?rfc compact="yes"?>
<!-- Default compact="no" Start sections on new pages -->
<?rfc subcompact="no"?>
<!-- Default subcompact="(as compact setting)" yes/no is not quite as compact as yes/yes -->
<?rfc autobreaks="yes" ?>
<!-- Default autobreaks="yes" avoid widows and orphans (not perfect) -->
<!-- HTML formatting control -->
<?rfc emoticonic="yes" ?>
<!-- Default emoticonic="no" Doesn't prettify HTML format -->
<?rfc background="" ?>
<!-- Default background="" when producing a html file, use this image -->
<?rfc useobject="no" ?>
<!-- Default useobject="no" use <object> not <src> when outputting HTML -->
<?rfc linkmailto="yes" ?>
<!-- Default linkmailto="yes" generate mailto: URL, as appropriate -->
<?rfc docmapping="no" ?>
<!-- Default docmapping="no" use hierarchical tags (e.g., <h1>, <h2>, etc.) for (sub)section titles -->
<?rfc slides="no" ?>
<!-- Default slides="no" when producing a html file, produce multiple files for a slide show -->
<rfc category="info" docName="draft-briscoe-tsvwg-relax-fairness-00"
     ipr="full3978">
  <front>
    <title abbrev="We Don't Have To Do Fairness">Problem Statement: We Don't
    Have To Do Fairness Ourselves</title>

    <author fullname="Bob Briscoe" initials="B." surname="Briscoe">
      <organization>BT & UCL</organization>

      <address>
        <postal>
          <street>B54/77, Adastral Park</street>

          <street>Martlesham Heath</street>

          <city>Ipswich</city>

          <code>IP5 3RE</code>

          <country>UK</country>
        </postal>

        <phone>+44 1473 645196</phone>

        <email>bob.briscoe@bt.com</email>

        <uri>http://www.cs.ucl.ac.uk/staff/B.Briscoe/</uri>
      </address>
    </author>

    <author fullname="Toby Moncaster" initials="T." surname="Moncaster">
      <organization>BT</organization>

      <address>
        <postal>
          <street>B54/70, Adastral Park</street>

          <city>Martlesham Heath</city>

          <region>Ipswich</region>

          <code>IP5 3RE</code>

          <country>UK</country>
        </postal>

        <phone>+44 1473 645196</phone>

        <email>toby.moncaster@bt.com</email>

        <uri>http://research.bt.com/networks/TobyMoncaster.html</uri>
      </address>
    </author>

    <author fullname="Louise Burness" initials="L." surname="Burness">
      <organization>BT</organization>

      <address>
        <postal>
          <street>B54/77, Adastral Park</street>

          <street>Martlesham Heath</street>

          <city>Ipswich</city>

          <code>IP5 3RE</code>

          <country>UK</country>
        </postal>

        <phone>+44 1473 646504</phone>

        <email>Louise.Burness@bt.com</email>

        <uri>http://research.bt.com/networks/LouiseBurness.html</uri>
      </address>
    </author>

    <date day="12" month="November" year="2007" />

    <area>Transport</area>

    <workgroup>Transport Area Working Group</workgroup>

    <keyword>Accountability</keyword>

    <keyword>Fairness</keyword>

    <keyword>Resource Sharing</keyword>

    <keyword>Congestion Control</keyword>

    <keyword>Quality of Service</keyword>

    <keyword>QoS</keyword>

    <keyword>Denial of Service</keyword>

    <keyword>Architecture</keyword>

    <abstract>
      <t>Nowadays resource sharing on the Internet is largely a result of what
      applications, users and operators do at run-time, rather than what the
      IETF designs into transport protocols at design-time. The IETF now needs
      to recognise this trend and consider how to allow resource sharing to be
      properly controlled at run-time.</t>
    </abstract>

    <note title="Requirements Language">
      <t>The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
      "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
      document are to be interpreted as described in <xref
      target="RFC2119"></xref>.</t>
    </note>
  </front>

  <middle>
    <!-- ================================================================ -->

    <section anchor="cacc_Introduction" title="Introduction">
      <t>The strength of the Internet is that any of the thousand million or
      so hosts may use nearly any network resource on the whole public
      Internet without asking, whether in access or core networks, wireless or
      fixed, local or remote. The question of how each resource is shared is
      generally delegated to the congestion control algorithms available on
      each endpoint, most often TCP.</t>

      <t>We (the IETF) aim to ensure reasonably fair sharing of the congested
      resources of the Internet <xref target="RFC2914"></xref>. Specifically,
      the TCP algorithm aims to ensure every flow gets a roughly equal share
      of each bottleneck, measured in packets per round trip time <xref
      target="RFC2581"></xref>. But our efforts have become distorted by
      unfair use of protocols we intended to be fair, and further by the
      attempts of operators to correct the situation. The problem is we aim to
      control fairness at protocol design-time, but resource shares are now
      primarily determined at run-time—as the outcome of a tussle
      between users, app developers and operators.</t>

      <t>For instance, about 35% of total traffic currently seen (Sep'07) at a
      core node on the public wireline Internet is p2p file-sharing {ToDo: Add
      ref}. Even though file-sharing generally uses TCP, it uses the
      well-known trick of opening multiple connections—currently around
      100 actively transferring over different paths is not uncommon. A
      competing Web application might open a couple of flows at a time, but
      perhaps only actively transfer data 1-10% of the time (its activity
      factor). Combining 50x less flows and 10-100x lower activity factor
      means the traffic intensity from the Web app can be 500-5,000x less.
      However, despite being so much lighter on the network, it gets 50x less
      bit rate through the bottleneck.</t>

      <t>The design-time approach worked well enough during the early days of
      the Internet, because most users' activity factors and numbers of flows
      were in proportion to their access link rate. But, now the Internet has
      to support a jostling mix of different attitudes to resource sharing:
      carelessness, unwitting self-interest, active self-interest, malice and
      sometimes even a little consideration for others. So although TCP sets
      an important baseline, it is no longer the main determinant of how
      resources are shared between users at run-time.</t>

      <t>Just because we can no longer control resource sharing at design
      time, we aren't saying it isn't important. In <xref
      target="cacc_Concrete_Consequences"></xref>, we show that badly skewed
      resource sharing has serious concrete knock-on effects that are of great
      concern to the health of the Internet.</t>

      <t>And we are not saying the IETF is powerless to do anything to help.
      However, our role must now be to create the run-time <spanx
      style="emph">policy framework</spanx> within which users and operators
      can control relative resource shares. So the debate is not about the
      IETF choosing between TCP-friendliness, max-min fairness, cost fairness
      or any other sort of fairness, because whatever we decide at design-time
      won't be strong enough to change what happens at run-time. We need to
      focus on giving principled and enforceable control to users and
      operators, so they can agree between themselves which fair use policy
      they want locally <xref target="Rate_fair_Dis"></xref>.</t>

      <t>The requirements for this resource sharing framework will be the
      subject of a future document, but the most important role of the IETF is
      to promote <spanx style="emph">understanding</spanx> of the sorts of
      resource sharing that users and operators will want to use at run-time
      and to resolve the misconceptions and differences between them (<xref
      target="cacc_Incompatible_Partial_Worlds"></xref>).</t>

      <t>We are in an era where new congestion control requirements often
      involve starting more aggressively than TCP or going faster than TCP, or
      not responding to congestion as quickly as TCP. By shifting control of
      fairness from design to run-time, we will free up all our new congestion
      control design work, so that it can first and foremost meet the
      objectives of these more demanding applications. But we can still
      quantify, minimise and constrain the effect on others due to faster
      average rate and different dynamics (<xref
      target="cacc_Dynamics_Design-Time"></xref>). We can say now that the
      framework will have to encompass and endorse the practice of opening
      multiple flows, for instance. But alongside recognition of such freedoms
      will come constraints, in order to balance the side-effects on other
      users over time.</t>
    </section>

    <!-- ================================================================ -->

    <section anchor="cacc_What_Problem" title="What Problem?">
      <t></t>

      <!-- ________________________________________________________________ -->

      <section anchor="cacc_Incompatible_Partial_Worlds"
               title="Two Incompatible Partial Worldviews">
        <t>When looking at the current Internet, some people see a massive
        fairness problem, while others think there's hardly a problem at all.
        This is because two divergent ways of reasoning about resource sharing
        have developed in the industry:<list style="symbols">
            <t>IETF guidelines on fair sharing of congested resources <xref
            target="RFC2357"></xref>,<xref target="RFC2309"></xref>,<xref
            target="RFC2914"></xref> have recommended that flows experiencing
            the same congested path should aim to achieve broadly equal window
            sizes, as TCP does <xref target="RFC2581"></xref>. We will
            characterise this as the "flow rate equality" worldview, shared by
            the IETF and large parts of the networking research
            community.<cref anchor="Note_Window">Within the flow rate equality
            worldview, there are differences in views over whether window
            sizes should be compared in packets or bytes, and whether a longer
            round trip time (RTT) should reduce the target rate or merely slow
            down how quickly the rate changes in order to reach a target rate
            that is independent of RTT [FAST]. However, although these details
            are important, they are merely minor internal differences within
            the flow rate equality worldview when compared against the
            differences with volume accounting.</cref></t>

            <t>Network operators and Internet users tend to reason about the
            problem of resources sharing very differently. Nowadays they do
            not generally concern themselves with the rates of individual
            flows. Instead they think in terms of the volume of data that
            different users transfer over a period <xref
            target="Res_p2p"></xref>. We will term this the "volume
            accounting" worldview. They do not believe volume over a period
            (traffic intensity) is a measure of unfairness in itself, but they
            believe it should be <spanx style="emph">taken into
            account</spanx> when deciding whether relative bit rates are
            fair.<!--{ToDo: Summarise Cho06 stats on volume distribution}--></t>
          </list></t>

        <t>The most obvious distinction between the two worldviews is that
        flow rate equality is between <spanx style="emph">flows</spanx>,
        whereas volume accounting shares resources between <spanx
        style="emph">users</spanx>. The IETF understands well that fairness is
        actually between users, but generally considers flow fairness to be a
        reasonable approximation as long as users aren't opening too many
        flows.</t>

        <t>However, there is a second much more subtle distinction. The flow
        rate equality worldview discusses fair resource sharing in terms of
        bit rates, but operators and users reason about fair bit rates in the
        context of byte volume over a period. Given bit rate is an
        instantaneous metric, it may aid understanding to convert 'volume over
        a period' into an instantaneous metric too. The relevant metric is
        traffic intensity, which like traffic rate is an instantaneous metric,
        but it takes account of likely activity <spanx style="emph">over
        time</spanx>. The traffic intensity from one user is the product of
        two metrics: i) the user's desired bit rate when active and ii) how
        often they are active over a period (their activity factor). </t>

        <t>Operators have to provision capacity based on the aggregate traffic
        intensity from all users over the busy period. And many users think in
        terms of how much volume they can transfer over a period. So, because
        traffic intensity is equivalent to 'volume over a period', both
        operators and users often effectively share the same worldview.</t>

        <t>To further aid understanding, <xref
        target="cacc_Example_Scenario"></xref> presents an example scenario
        where heavy users compete for a bottleneck with light users. It has
        enough similarities to the current Internet to be relevant, but it has
        been stripped to its bare essentials to allow the main issues to be
        grasped.</t>

        <t>The base scenario in <xref target="cacc_Base_Scenario"></xref>
        starts with the light users having TCP connections open for less of
        the time than heavy users (a lower activity factor). But, when they
        are active, they open as many connections as heavy users. It shows
        that users with a lower activity factor transfer less volume of
        traffic through the bottleneck over a period because, even though TCP
        gives roughly equal rate to each flow, the heavy users' flows are
        present more of the time. </t>

        <t>The volume accounting view is <spanx style="emph">not</spanx> that
        it is unfair for some users to transfer more volume than
        others—afterall the lighter users have less traffic that they
        want to send. However, they believe it is reasonable for users who put
        a heavier load on the system to be given less bottleneck bit rate than
        lighter users. </t>

        <t><xref target="cacc_Compounding_Overlooked_Dimensions"></xref>
        continues the example, giving the heavy users the added advantage of
        using 50x multiple flows, just as they do on the current Internet.
        When multiple flows are compounded with their higher activity factors,
        they can get 500-2,000x greater traffic intensity through the
        bottleneck.</t>

        <t>Certainly, the flow rate equality worldview recognises that opening
        50x more flows than other users starts to become a serious fairness
        problem, because some users get 50x more bit rate through a bottleneck
        than others. But the volume accounting worldview sees this as a much
        bigger problem. They first see 2,000x heavier use of the bottleneck
        over time, then they judge that <spanx style="emph">also</spanx>
        getting 50x greater bit rate seems seriously unfair.</t>

        <!--Add ref to experiment on colleague's unaltered Windows XP machine running Azureus-->

        <t>But are these numbers realistic? Attended use of something like the
        Web might typically have an activity factor of 1-10%, while unattended
        apps approach 100%. A Web browser might typically open two TCPs when
        active <xref target="RFC2616"></xref>, while a p2p file-sharing app on
        a 512kbps upstream DSL line actively uses anything from 40-500
        connections <xref target="az-calc"></xref>. Heavy users generally
        compound the two factors together (10-100x greater activity factor and
        20-250x more connections), achieving anything from 200x to 25,000x
        greater traffic intensity through a bottleneck than light users.</t>

        <t>The above question of what size the different worldviews think
        resource shares <spanx style="emph">should</spanx> be is separate from
        the question of whether to <spanx style="emph">enforce</spanx> them
        and how to (see <xref target="cacc_Losing_Voluntarism"></xref>).
        Within the volume accounting worldview, many operators (particularly
        in Europe) already limit the bit rate of their heaviest users at peak
        times in order to protect the experience of the majority of their
        customers.<cref anchor="Note_Neutral">Enforcement of /overall/ traffic
        limits within an agreed acceptable use policy is a completely
        different question to that of whether operators should disciminate
        against /specific/ applications or service providers (but they are
        confusible—see the section on DPI.</cref> But, enforcement is a
        separate question. Although prevalent use of TCP seems to be
        continuing without any enforcement, even the flow rate equality
        worldview generally accepts that opening excessive multiple
        connections can't be solved voluntarily. Quoting RFC2914,
        "…instead of a spiral of increasingly aggressive transport
        protocols, we … have a spiral of increasingly …
        aggressive applications").</t>

        <t>To summarise so far, one industry worldview aims for equal flow
        rates, while the other prefers an outcome with very unequal flow
        rates. Even though they both share the same intentions of fairer
        resource sharing, the two worldviews have developed subgoals that are
        fundamentally at odds.</t>

        <section anchor="cacc_Overlooked_Dimensions"
                 title="Overlooked Degrees of Freedom">
          <t>So which worldview is correct? Actually, our reason for pointing
          out these divergent worldviews is to show that both contain valuable
          insights, but that each also highlights weaknesses in the other.
          Given our audience is the IETF, we have tried to explain the volume
          accounting worldview in terms of flow rate equality, but volume
          accounting is by no means rigorous or complete itself. <xref
          target="cacc_Table_Overlooked_Dimensions"></xref> identifies the
          three degrees of freedom of resource sharing that are missing in one
          or the other of the two worldviews.</t>

          <texttable anchor="cacc_Table_Overlooked_Dimensions"
                     title="Resource Sharing Degrees of Freedom Encompassed by Different Worldviews; Y = yes and X = no.">
            <preamble></preamble>

            <ttcol>Degree of Freedom</ttcol>

            <ttcol align="center">Flow Rate Equality</ttcol>

            <ttcol align="center">Volume Accounting</ttcol>

            <c>Activity factor</c>

            <c>X</c>

            <c>Y</c>

            <c>Multiple flows</c>

            <c>X</c>

            <c>Y</c>

            <c>Congestion variation</c>

            <c>Y</c>

            <c>X</c>

            <postamble></postamble>
          </texttable>

          <t><list style="hanging">
              <t hangText="Activity factor:">We have already pointed out how
              flow rate equality does not take different activity factors into
              account. On the other hand, volume accounting naturally takes
              the on-off activity of flows into account, because in the
              process of counting volume over time, the off periods are
              naturally excluded.</t>

              <t hangText="Multiple flows:">Similarly, it is well-known <xref
              target="RFC2309"></xref> <xref target="RFC2914"></xref> that
              flow rate equality does not make allowance for multiple flows,
              whereas counting volume naturally includes all flows from a
              user, whether they terminate at the same remote endpoint or many
              different ones.</t>

              <t hangText="Congestion variation:">Flow rate equality, of
              course, takes full account of how congested different
              bottlenecks are at different times, ensuring that the same
              volume must be squeezed out over a longer duration, the more
              flows it competes with. However, volume accounting doesn't
              recognise that congestion can vary by orders of magnitude,
              making it fairly useless for encouraging congestion control. The
              best it can do is only count volume during a 'peak period',
              effectively considering congestion as either 1 everywhere during
              this time or 0 everywhere otherwise.</t>
            </list>These clearly aren't just problems of detail. Having each
          overlooked whole dimensions of the problem, both worldviews seem to
          require a fundamental rethink. In a future document defining the
          requirements for a new resource sharing framework, we plan to unify
          both worldviews. But, in the present problem statement, it is
          sufficient to register that we need to reconcile the fundamentally
          contradictory worldviews that the industry has developed about
          resource sharing.</t>
        </section>
      </section>

      <!-- ________________________________________________________________ -->

      <section anchor="cacc_Average_Rates_Run-Time"
               title="Average Rates are a Run-Time Issue">
        <t>A less obvious difference between the two worldviews is that flow
        rate equality tries to control resource shares at design-time, while
        volume accounting controls resource shares once the run-time situation
        is known. Also the volume accounting worldview actually involves two
        separate functions: passive monitoring and active intervention. So,
        importantly, the run-time questions of whether to and how to intervene
        can depend on policy.</t>

        <t>The "spiral of increasingly aggressive applications" <xref
        target="RFC2914"></xref> has shifted the resource sharing problem out
        of the IETF's design-time space, making flow rate equality
        insufficient (or perhaps even inappropriate) in technical and in
        policy terms:<list style="hanging">
            <t hangText="Technical:">At design time, it is impossible to know
            whether a congestion control will be fair at run-time without
            knowing more about the run-time situation it will be used
            in—how long flow durations will be and whether users will
            open multiple flows.</t>

            <t hangText="Policy:">At design time, we cannot (and should not)
            prejudge the 'fair use' policy that has been agreed between users
            and their network operators.</t>
          </list>A transport protocol can no longer be made 'fair' at design
        time—it all now depends how 'unfairly' it is used at run-time,
        and what has been agreed as 'unfair'.</t>

        <t>However, we are not saying that volume accounting is the answer. It
        just gives us the insight that resource sharing has to be controlled
        at run-time by policy, not at design-time by the IETF. Volume
        accounting would be more useful if it took a more precise approach to
        congestion than either 'everything is congested' or 'nothing is
        congested'.</t>

        <t>What operators and users need from the IETF is a framework to judge
        and to control resource sharing at run-time. It needs to work across
        all a user's flows (like volume accounting). It needs to take account
        of idle periods over time (like volume accounting). And it needs to
        take account of congestion variation (like flow rate equality).</t>
      </section>

      <!-- ________________________________________________________________ -->

      <section anchor="cacc_Dynamics_Design-Time"
               title="Protocol Dynamics is the Design-Time Issue">
        <t>Although fairness is a run-time issue, at protocol design-time it
        requires more from the IETF than just a policy control framework.
        Policy can control the <spanx style="emph">average</spanx> amount of
        congestion that a particular application causes, but the Internet also
        needs the collective expertise of the IETF and the IRTF to standardise
        best practice in the <spanx style="emph">dynamics</spanx> of transport
        protocols. The IETF has a duty to provide standard transports with a
        response to congestion that is always safe and robust. But the hard
        part is to keep the network safe while still meeting the needs of more
        demanding applications (e.g. high speed transfer of data objects or
        media streaming that can adapt its rate but not too abruptly).</t>

        <t>If we assume for a moment that we will have a framework to judge
        and control <spanx style="emph">average</spanx> rates, we will still
        need a framework to assess which proposed congestion controls make the
        trade-off between achieving the task effectively and minimising
        congestion caused to others, during <spanx
        style="emph">dynamics</spanx>:<list style="symbols">
            <t>The faster a new flow accelerates the more packets it will have
            in flight when it detects its first loss, potentially leading many
            other flows to experience a long burst of losses as queues
            overrun. When is a fast start fast enough? Or too fast <xref
            target="RFC3742"></xref>?</t>

            <t>One way for a small number of high speed flows to better
            utilise a high speed link is to respond more smoothly to
            congestion events than TCP's rate-halving saw-tooth does
            [proprietary fast TCPs] <xref target="FAST"></xref>,<xref
            target="RFC3649"></xref>. But then new flows will take much longer
            to 'push-in' and reach a high rate themselves.</t>

            <t>Transports like TCP-friendly rate control [proprietary media
            players], <xref target="RFC3448"></xref>, <xref
            target="RFC4828"></xref> are designed to respond more smoothly to
            congestion than TCP. But even if a TFRC flow has the same average
            bit rate as a TCP flow, the more sluggish it is, the more
            congestion it will cause <xref target="Rate_fair_Dis"></xref>. How
            do we decide how much smoother we should go? How large a
            proportion of Internet traffic could we allow to be unresponsive
            to congestion over long durations, before we were at risk of
            causing growing periods of congestion collapse <xref
            target="RFC2914"></xref>?<cref anchor="Note_Collapse">Some would
            say that it is not a congestion collapse if congestion control
            automatically recovers the situation after a while. However, even
            though lack of autorecovery would be truly devastating, it isn't
            part of the definition [RFC2914].</cref></t>

            <t>TFRC has been proposed as a possible way for aggregates of
            flows crossing the public Internet to respond to congestion
            (pseudo-wire emulations may contain flows that cannot, or do not
            want to respond quickly to congestion themselves) <xref
            target="I-D.rosen-pwe3-congestion"></xref>, <xref
            target="I-D.ietf-capwap-protocol-specification"></xref>, <xref
            target="TSV_CAPWAP_issues"></xref>. But it doesn't make any sense
            to insist that, wherever flows are aggregated together into one
            identifiable bundle, the whole bundle of perhaps hundreds of flows
            must keep to the same mean rate as a single TCP flow.</t>
          </list></t>

        <t>In view of the continual demand for alternate congestion controls,
        the IETF has recently agreed a new process for standardising them
        <xref target="ion-tsv-alt-cc"></xref>. The IETF will use the expertise
        of the IRTF Internet congestion control research group, governed by
        agreed general guidelines for the design of new congestion controls
        <xref target="RFC5033"></xref>. However, in writing those guidelines
        it proved very difficult to give any specific guidance on where a line
        could be drawn between fair and unfair protocols. The best we could do
        were phrases like, "Alternate congestion controllers that have a
        significantly negative impact on traffic using standard congestion
        control may be suspect..." and "In environments with multiple
        competing flows all using the same alternate congestion control
        algorithm, the proposal should explore how bandwidth is shared among
        the competing flows."</t>

        <t>Once we have agreed that average behaviour should be a policy
        issue, we can focus on the dynamic behaviour of congestion controls,
        which is where the important standards issues lie, such as preventing
        congestion collapse or preventing new flows causing bursts of
        congestion by unnecessarily overrunning as they seek out the operating
        point of their path.</t>

        <t>As always, the IETF will not want to standardise aspects where
        implementers can gain an edge over their competitors, but we must set
        standards to prevent serious harm to the stability and usefulness of
        the Internet, and to make transports available that avoid causing
        <spanx style="emph">unnecessary</spanx> congestion in the course of
        achieving any particular application objective.</t>
      </section>
    </section>

    <!-- ================================================================ -->

    <section anchor="cacc_Concrete_Consequences"
             title="Concrete Consequences of Unfairness">
      <t>People have different levels of tolerance for unfairness. Even when
      we agree how to measure fairness, there are a range of views on how
      unfair the situation needs to get before the IETF should do anything
      about it. Nonetheless, lack of fairness can lead to more concretely
      pathological knock-on effects. Even if we don't particularly care if
      some users get more than their fair share and others less, we should
      care about the more concrete knock-on effects below.</t>

      <!-- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -  -->

      <section anchor="cacc_Invest_Risk" title="Higher Investment Risk">
        <t>Some users want more Internet capacity to transfer large volumes of
        data, while others want more capacity to be able to interact more
        quickly with other sites and other users. We have called these heavy
        and light users, although of course, many users are mix of the two in
        differing proportions.</t>

        <t>We have shown that heavy users can use applications that open
        multiple connections, so that TCP gives the light users very little of
        a bottleneck. But unfortunately, upgrading capacity does little for
        the light users unless the heavy users run out of data to send (which
        doesn't tend to happen often). In the reasonably realistic example in
        <xref target="cacc_Upgrading_Worse"></xref>, the light users start off
        only being able to use 10kbps of their 2Mbps line because heavy users
        are skewing the sharing of the bottleneck by using multiple flows. But
        a 4x upgrade to the bottleneck, which should add 500kbps per user if
        shared equally, only gives the light users 30kbps extra. </t>

        <t>But, the upgrade has to be paid for. A commercial ISP will
        generally pass on the cost equally to all its customers through its
        monthly fees. So, to rub salt in the wound, the light users end up
        paying the cost of this 500kbps upgrade but we have seen they only get
        30kbps. Ultimately, extreme unfairness in the sharing of capacity
        tends to drive operators to stop investing in capacity. Because all
        the light users, who experience so little of the benefit, won't be
        prepared to pay an equal share to recover the costs—the ISP
        risks losing them to a 'fairer' competitor.</t>

        <t>But there seems to be plenty of evidence that operators around the
        world are still investing in capacity growth despite the prevalence of
        TCP. How can this be, if flow rate equality makes investment so risky?
        One explanation, particularly in parts of Asia, is that some
        investments are Governernment subsidised. In the US, the explanation
        is probably more down to weak competition. In Europe, the main
        explanation is that many commercial operators haven't allowed their
        networks to become as unfair as the above example—they have made
        resource sharing fairer by <spanx style="emph">overriding</spanx>
        TCP's flow rate equality.</t>

        <t>Competitive operators in many countries limit the volume
        transferred by heavy users, particularly at peak times. They have
        effectively overriden flow rate equality to achieve a different
        allocation of resources that they believe is better for the majority
        of their customers (and consequently better for their competitive
        position). Typically these operators use a combination of tiered
        pricing of volume caps and throttling of the heaviest so-called
        'unlimited' users at peak times. In this way they have removed some of
        the investment risk that would otherwise have resulted if flow rate
        equality had been relied on to share congested resources.</t>
      </section>

      <!-- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -  -->

      <section anchor="cacc_Losing_Voluntarism" title="Losing Voluntarism">
        <t>Throughout the early years of the Internet, flow rate equality
        resulted in approximate fairness that most people considered
        sufficient. This was because most users' traffic during peak hours
        tended to correlate with their access rate. Those who bought high
        capacity access also generally sent more traffic at peak times (e.g.
        heavy users or server farms).</t>

        <t>As higher access rates have become more affordable, this happy
        coincidence has been eroded. Some people only require their higher
        access rate occasionally, while others require it more continuously.
        But once they all have more access capacity, even those who don't
        really require it all the time often fill it anyway—as long as
        there's nothing to dissuade them. People tend to use what they desire,
        not just what they require.</t>

        <t>Of course, more access traffic requires more shared capacity at
        relevant network bottlenecks. But if we rely on TCP to share out these
        bottlenecks, we have seen how those who just desire more can swamp
        those who require more (<xref target="cacc_Invest_Risk"></xref>).</t>

        <t>Some operators have continued to provision sufficiently excessive
        shared capacity and just passed the cost on to all their customers.
        But many operators have found that those customers who don't actually
        require all that shared infrastructure would rather not have to pay
        towards its cost. So, to avoid losing customers, they have introduced
        tiered volume limits (this hasn't happened in the US yet though). It
        is well known that many users are averse to unpredictable charges
        <xref target="PMP"></xref> (§5), so many now choose ISPs who
        limit their volume (with suitable override facilities) rather than
        charge more when they use more.</t>

        <t>Thus, we are seeing a move away from voluntary restraint (within
        peak access rates) towards a preference for enforced fairness, as long
        as the user stays in overall control. This has implications on the
        Internet infrastructure that the IETF needs to recognise and address.
        Effectively, parts of the best effort Internet are becoming like the
        other Diffserv classes, with traffic policers and traffic conditioning
        agreements (TCAs <xref target="RFC2475"></xref>), albeit volume-based
        rather than rate and burst-based TCAs. (In fact, the addition of
        congestion accounting or policing need not be confined to just the
        best effort class.)</t>

        <t>We are not saying that the Internet <spanx
        style="emph">requires</spanx> fairness enforcement, merely that it has
        become prevalent. We need to acknowledge the trend towards enforcement
        to ensure that it does not introduce unnecessary complexity into the
        basic functioning of the Internet, and that our current approach to
        fairness (embedded in endpoint congestion control) remains compatible
        with this changing world. For instance, when a rate policer introduces
        drops, are they equivalent to drops due to congestion? are they
        equivalent to drops when you exceed your own access rate? do we need
        to tell the difference?</t>
      </section>

      <!-- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -  -->

      <section anchor="cacc_DPI"
               title="Networks using DPI to make Choices for Users">
        <t>We have seen how network operators might well believe it is in
        their customers' interests to override the resource sharing decisions
        of TCP. They seem to have sound reasons for throttling their heaviest
        users at peak times. But this is leading to a far more controversial
        side-effect: network operators have started making performance choices
        between <spanx style="emph">applications</spanx> on behalf of their
        customers.</t>

        <t>Once operators start throttling heavy users, they hit a problem.
        Most heavy volume users are actually a mix of the two types
        characterised in our example scenario (<xref
        target="cacc_Example_Scenario"></xref>). Some of their traffic is
        attended and some is unattended. If the operator throttles all traffic
        from a heavy user indiscriminately, it will severely degrade the
        customer's attended applications, but it actually only needs to
        throttle the unattended applications to protect the traffic of
        others.</t>

        <t>Ideally, the threat of heavy throttling of all a user's traffic
        would encourage the user to self-throttle the traffic she least
        valued, in order to avoid the operator's indiscriminate throttling.
        But many users these days have neither the expertise nor the software
        to do this. Instead, operators have generally decided to infer what
        they think the user would do, using readily available deep packet
        inspection (DPI) equipment. </t>

        <t>An operator may infer customer priorities with honourable
        intentions, but such activity is easily confusible with attempts to
        discriminate against certain applications that the operator happens
        not to like. Also customers get understandably upset every time the
        operator guesses their priorities wrongly.</t>

        <t>It is well documented (but less well-known) that user priorities
        are task-specific, not application-specific <xref
        target="AppVsTask"></xref>. P2p filesharing can be used for
        downloading music with some vague intent to listen to it some day
        soon, or to download a critical security patch. User intent cannot be
        inferred at the network layer just by working out what the application
        is. The end-to-end design principle <xref target="RFC1958"></xref>
        warns that a function should only be implemented at a lower layer
        after trying really hard to implement it at a higher layer. Otherwise,
        the network layer gradually becomes specialised around the functions
        and priorities of the moment—the middlebox problem <xref
        target="RFC3234"></xref>.</t>

        <t>To address this problem of feature creep into the network layer, we
        need to understand whether there are valid reasons why this DPI is
        being deployed to override TCP's decisions. We shouldn't deny the
        existence of a problem just because one solution to it breaks a
        fundamental Internet design principle. We should instead find a better
        solution.</t>
      </section>

      <!-- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -  -->

      <section anchor="cacc_Starvation_Anomalies_Emergencies"
               title="Starvation during Anomalies and Emergencies">
        <t>The problems due to unfairness that we have outlined so far all
        arise when the Internet is working normally. However, fairness
        concerns become far more acute when a part of the Internet
        infrastructure becomes extremely stressed, either because there's much
        more traffic than expected (e.g. flash crowds), or much less capacity
        than expected (e.g. physical attack, accident, disaster).</t>

        <t>Under non-disaster conditions, we have already said that fair
        sharing of congested resources is a matter that should be decided
        between users and their providers at run-time. Often that will mean
        "you get what you've paid for" becomes the rule, at least in
        commercial parts of the Internet. But during really acute emergencies
        many people would expect such commercial concerns to be set aside
        <xref target="I-D.floyd-tsvwg-besteffort"></xref>.</t>

        <t>We agree that users shouldn't be able to squeeze out others during
        emergencies. But the mechanisms we have in place at the moment don't
        allow anyone to control whether this happens or not, because they can
        be overriden at run-time by using the extra degress of freedom
        available to get round TCP. It could equally be argued that each user
        (not each flow) should get an equal share of remaining capacity in an
        emergency. Indeed, it would seem wrong for one user to expect 100
        continuously running flows downloading music & videos to take 100
        times more capacity than other users sending brief flows containing
        messages trying to contact loved ones or the emergency services <xref
        target="Hengchun_quake"></xref>.<cref anchor="Note_Earthquake">On 26
        Dec 2006, the Hengchun earthquake caused faults on 12 of the 18
        undersea cables passing between Taiwan and the Philippines. The
        Internet was virtually unusable for those trying to make their
        emergency arrangements over these cables (as well as for much of Asia
        generally). Each of these flows was still having to compete with the
        multiple flows of video downloads for remote users who were presumably
        oblivious to the fact they were consuming much of the surviving
        capacity. When the Singaporean ISP, SingNet, announced restoration of
        service before the cables were repaired, it revealed that it had
        achieved this at the expense of video downloads and gaming traffic
        .</cref></t>

        <t>We argue that fairness during emergencies is, more than anything
        else, a policy matter to be decided at run-time (either before or
        during an anomaly) by users, operators, regulators and
        governments—not at design time by the IETF. The IETF should
        however provide the framework within which typical policies can be
        enforced. And the IETF should ensure that the Internet is still likely
        to utilise resources <spanx style="emph">efficiently</spanx> under
        extreme stress, assuming a reasonable mix of likely policies,
        including none.</t>

        <t>The main take-away point from this section is that the IETF should
        not, and need not, make such life-and-death decisions. It should
        provide protocols that allow any of these policy options to be chosen
        at the time of need or by making contingencies beforehand. The
        congestion accountability framework in {ToDo: ref sister doc} provides
        such control, while also allowing different controls (including no
        control at all) in normal circumstances. For instance an ISP might
        normally allow its customers to pay to override any usage limits. But
        during a disaster it might suspend this right. Then users would get
        only the shares they had established before the disaster broke out
        (the ISP would thus also avoid accusations of profiteering from
        people's misery). Whatever, it is not for the IETF to embed answers to
        questions like these in our protocols.</t>
      </section>
    </section>

    <!-- ================================================================ -->

    <section anchor="cacc_Sec_Consider" title="Security Considerations">
      <t>{ToDo:}</t>
    </section>

    <!-- ================================================================ -->

    <section anchor="cacc_Conclusions" title="Conclusions">
      <t>{ToDo:}</t>
    </section>

    <!-- ================================================================ -->

    <section anchor="cacc_Acknowledgements" title="Acknowledgements">
      <t>Arnaud Jacquet, Phil Eardley.</t>
    </section>

    <!-- ================================================================ -->

    <section anchor="cacc_Comments_Solicited" title="Comments Solicited">
      <t>Comments and questions are encouraged and very welcome. They can be
      addressed to the IETF Transport Area working group mailing list
      <tsvwg@ietf.org>, and/or to the authors.</t>
    </section>
  </middle>

  <back>
    <!-- ================================================================ -->

    <references title="Normative References">
      <?rfc include="reference.RFC.2119" ?>

      <?rfc include="reference.RFC.2309" ?>

      <?rfc include="reference.RFC.2914" ?>

      <?rfc include="reference.RFC.2581" ?>
    </references>

    <references title="Informative References">
      <?rfc include="reference.RFC.2357" ?>

      <?rfc include="reference.RFC.2616" ?>

      <?rfc include="reference.RFC.3448" ?>

      <?rfc include="reference.RFC.4828" ?>

      <?rfc include="reference.RFC.5033" ?>

      <?rfc include="reference.RFC.2475" ?>

      <?rfc include="reference.RFC.1958" ?>

      <?rfc include="reference.RFC.3234" ?>

      <?rfc include="reference.RFC.3649" ?>

      <?rfc include="reference.RFC.3742" ?>

      <?rfc include="localref.Jin04.FAST_TCP" ?>

      <?rfc include="localref.Odlyzko97.PMP" ?>

      <?rfc include="localref.Infinite-Source07.Azureus_calculator" ?>

      <?rfc include="localref.Briscoe07.Rate_fair_Dis" ?>

      <?rfc include="localref.Cho06.Res_p2p" ?>

      <?rfc include="localref.Bouch00.Of_packets_and_people" ?>

      <?rfc include="localref.Wikipedia06.Hengchun_quake" ?>

      <?rfc include="localref.ION.tsv-alt-cc" ?>

      <?rfc include="reference.I-D.rosen-pwe3-congestion" ?>

      <?rfc include="localref.Borman07.TSV_CAPWAP_issues" ?>

      <?rfc include="reference.I-D.ietf-capwap-protocol-specification" ?>

      <?rfc include="reference.I-D.floyd-tsvwg-besteffort" ?>
    </references>

    <section anchor="cacc_Example_Scenario" title="Example Scenario">
      <t></t>

      <section anchor="cacc_Base_Scenario" title="Base Scenario">
        <t>We will consider 100 users all sharing a link from the Internet
        with 2Mbps downstream access capacity. Eighty bought their line for
        occasional flurries of activity like browsing the Web, booking their
        travel arrangements or reading their email. The other twenty bought it
        mainly for unattended volume transfer of large files. We will call
        these two types of use attended (or light) and unattended (or heavy).
        Ignoring the odd UDP packet, we will assume all these applications use
        TCP congestion control, and that all flows have approximately equal
        round trip times.</t>

        <t>Imagine the network operator has provisioned the shared link for a
        contention ratio of 20:1, ie 100x2Mbps/20 = 10Mbps. For simplicity, we
        assume a 16hr 'day' and that the attended use is only in the 'day',
        while unattended use is always present, having the night to
        itself.</t>

        <t>During the 'day', flows from the sixty attended users come and go
        with about 1 in 10 actively downloading flows at any one time (a
        downstream activity factor of 10%). To start with, we will further
        assume that, when active, every user has approximately the same number
        of flows open, whether attended or unattended. So, once all flows have
        stabilised, at any instant TCP will ensure every user (when active)
        gets about 10Mbps/(80*10% + 20*100%) = 357kbps of the bottleneck.</t>

        <t><xref target="cacc_Table_Base_Scenario"></xref> tabulates the
        salient features of this scenario. Also the rightmost column shows the
        volume transferred per user during the day, and for completeness the
        bottom row shows the aggregate.</t>

        <texttable anchor="cacc_Table_Base_Scenario"
                   title="Base Scenario assuming 100% utilisation of 10Mbps bottleneck and each user runs approx. equal numbers of flows with equal RTTs.">
          <preamble></preamble>

          <ttcol align="right">Type of use</ttcol>

          <ttcol align="right">No. of users</ttcol>

          <ttcol align="right">Activ- ity factor</ttcol>

          <ttcol align="right">Day rate /user (16hr)</ttcol>

          <ttcol align="right">Day volume /user (16hr)</ttcol>

          <c>Attended</c>

          <c>80</c>

          <c>10%</c>

          <c>357kbps</c>

          <c>257MB</c>

          <c>Unattended</c>

          <c>20</c>

          <c>100%</c>

          <c>357kbps</c>

          <c>2570MB</c>

          <c> </c>

          <c></c>

          <c></c>

          <c></c>

          <c></c>

          <c>Aggregate</c>

          <c>100</c>

          <c></c>

          <c>10Mbps</c>

          <c>72GB</c>

          <postamble></postamble>
        </texttable>

        <t>This scenario is not meant to be an accurate model of the current
        Internet, for instance:<list style="symbols">
            <t>Utilisation is never 100%.</t>

            <t>Upstream not downstream constrains most p2p apps on DSL (but
            not all fixed & wireless access technologies).</t>

            <t>The activity factor of 10% in our base example scenario is
            perhaps an optimistic estimate for attended use over a 16hr peak
            period. 1% is just as likely for many users (before file-sharing
            became popular, DSL networks were provisioned for a contention
            ratio of about 25:1, aiming to handle a peak average activity
            factor of 4% across all user types).</t>

            <t>And rather than falling into two neat categories, users sit on
            a wide spectrum that extends to far more extreme types in both
            directions, while in between there are users who mix both types in
            different proportions <xref target="Res_p2p"></xref>.</t>
          </list>But the scenario has merely been chosen because it makes it
        simple to grasp the main issues while still retaining some similarity
        to the real Internet. We will also develop the scenario as we go, to
        add more realism (e.g. adding mixed user types).</t>
      </section>

      <section anchor="cacc_Compounding_Overlooked_Dimensions"
               title="Compounding Overlooked Degrees of Freedom">
        <t><xref target="cacc_Table_Compounded_Scenario"></xref> extends the
        base scenario of <xref target="cacc_Example_Scenario"></xref> to
        compound differences in average activity factor with differences in
        average numbers of active flows.</t>

        <t>During the 'day' at any instant we assume on average that attended
        use results in 2 flows per user (which are still only open 10% of the
        time), while unattended use results in 100 flows per user open
        continuously. So at any one time 2016 flows are active, 16 from
        attended use (10%*80=8 users at any one time * 2 flows) and 2000 from
        unattended use (20 users * 100 flows). TCP will ensure each of the 8
        users who are active at any one time gets about 2*10Mbps/2016 =
        9.9kbps of the bottleneck, while each of the 20 unattended users gets
        about 100*10Mbps/2016 = 496kbps. This ignores flow start up effects,
        which will tend to make matters even worse for attended use, given
        briefer flows start more often.</t>

        <texttable anchor="cacc_Table_Compounded_Scenario"
                   title="Compounded scenario with attentive users less frequently active and running less flows than unattentive users,  assuming 100% utilisation of 10Mbps bottleneck and all equal RTTs.">
          <preamble></preamble>

          <ttcol align="right">Type of use</ttcol>

          <ttcol align="right">No. of users</ttcol>

          <ttcol align="right">Activ- ity factor</ttcol>

          <ttcol align="right">Ave simultaneous flows /user</ttcol>

          <ttcol align="right">Day rate /user (16hr)</ttcol>

          <ttcol align="right">Day volume /user (16hr)</ttcol>

          <c>Attended</c>

          <c>80</c>

          <c>10%</c>

          <c>2</c>

          <c>9.9kbps</c>

          <c>7.1MB</c>

          <c>Unattended</c>

          <c>20</c>

          <c>100%</c>

          <c>100</c>

          <c>496kbps</c>

          <c>3.6GB</c>

          <c> </c>

          <c></c>

          <c></c>

          <c></c>

          <c></c>

          <c></c>

          <c>Aggregate</c>

          <c>100</c>

          <c></c>

          <c>2016</c>

          <c>10Mbps</c>

          <c>72GB</c>

          <postamble></postamble>
        </texttable>
      </section>

      <section anchor="cacc_Hybrid_Users" title="Hybrid Users">
        <t>{ToDo:}</t>
      </section>

      <section anchor="cacc_Upgrading_Worse"
               title="Upgrading Makes Most Users Worse Off">
        <t>Now that the light users are only getting 9.9kbps from their 2Mbps
        lines, the operator needs to consider upgrading their bottleneck (and
        all the other access bottlenecks for its other customers), so it does
        a market survey. The operator finds that fifty of the eighty light
        users and ten of the twenty heavy users are willing to pay more to get
        an extra 500kbps each at the bottleneck. (Note that by making a
        smaller proportion of the heavy users willing to pay more we haven't
        weighted the argument in our favour—in fact our argument would
        have been even stronger the other way round.)</t>

        <t>To satisfy the sixty users who are willing to pay for a 500kbps
        upgrade will require a 60*500kbps = 30Mbps upgrade to the bottleneck
        and proportionate upgrades deeper into the network, which will cost
        the ISP an extra $120 per month (say). The outcome is shown in <xref
        target="cacc_Table_Upgrade1_Scenario"></xref>. Because the bottleneck
        has grown from 10Mbps to 40Mbps, the bit rates in the whole scenario
        essentially scale up by 4x. However, also notice that the total volume
        sent by the light users has not grown by 4x. Although they can send at
        4x the bit rate, which means they get more done and therefore transfer
        more volume, they don't have 4x more volume to transfer—they let
        their machines idle for longer between transfers reflected in their
        activity factor having reduced from 10% to 4%. More bit rate was what
        they wanted, not more volume particularly.</t>

        <t>Let's assume the operator increases the monthly fee of all 100
        customers by $1.20 to pay for the $120 upgrade. The light users had a
        9.9kbps share of the bottleneck. They've all paid their share of the
        upgrade, but they've only got 30kbps more than they had—nothing
        like the 500kbps upgrade most of them wanted and thought they were
        paying for. TCP has caused each heavy user to increase the bit rate of
        its flows by 4x too, and each has 50x more flows for 25x more of the
        time, so they use up most of the newly provisioned capacity even
        though only half of them were willing to pay for it.</t>

        <t>But the operator knew from its marketing that 30 of the light users
        and 10 of the heavy ones didn't want to pay any more anyway. Over
        time, the extra $1.20/month is likely to make them drift away to a
        competitor who runs a similar network but who decided not to upgrade
        its 10Mbps bottlenecks. Then the cost of the upgrade on our example
        network will have to be shared over 60 not 100 customers, requiring
        each to pay $2/month extra, rather than $1.20.</t>

        <texttable anchor="cacc_Table_Upgrade1_Scenario"
                   title="Scenario with bottleneck upgraded to 40Mbps, but otherwise unchanged from compounded scenario.">
          <preamble></preamble>

          <ttcol align="right">Type of use</ttcol>

          <ttcol align="right">No. of users</ttcol>

          <ttcol align="right">Activ- ity factor</ttcol>

          <ttcol align="right">Ave simultaneous flows /user</ttcol>

          <ttcol align="right">Day rate /user (16hr)</ttcol>

          <ttcol align="right">Day volume /user (16hr)</ttcol>

          <c>Attended</c>

          <c>80</c>

          <c>4%</c>

          <c>2</c>

          <c>40kbps</c>

          <c>11MB</c>

          <c>Unattended</c>

          <c>20</c>

          <c>100%</c>

          <c>100</c>

          <c>2.0Mbps</c>

          <c>14GB</c>

          <c> </c>

          <c></c>

          <c></c>

          <c></c>

          <c></c>

          <c></c>

          <c>Aggregate</c>

          <c>100</c>

          <c></c>

          <c>2006.4</c>

          <c>40Mbps</c>

          <c>288GB</c>

          <postamble></postamble>
        </texttable>

        <t>But perhaps losing a greater proportion of the heavy users will
        help? <xref target="cacc_Table_Upgrade2_Scenario"></xref> shows the
        resulting shares of the bottleneck once all the cost sensitive
        customers have drifted away. Bit rates have increased by another 2x,
        mainly because there are 2x fewer heavy users. But that still only
        gives the light users 80kbps when they wanted 500kbps—and, to
        rub salt in their wounds, their monthly fees have increased by $2 in
        all. The remaining 10 heavy users are probably happy enough though.
        For the extra $2/month they get to transfer 8x more volume each (and
        they still have the night to themselves).</t>

        <t>We have shown how the operator might lose those customers who
        didn't want to pay. But it also risks losing all fifty of those
        valuable light customers who were willing to pay, and who did pay, but
        who hardly got any benefit. In this situation, a rational operator
        will eventually have no choice but to stop investing in capacity,
        otherwise it will only be left with ten customers.</t>

        <texttable anchor="cacc_Table_Upgrade2_Scenario"
                   title="Scenario with bottleneck upgraded to 40Mbps, but having lost customers due to extra cost; otherwise unchanged from compounded scenario.">
          <preamble></preamble>

          <ttcol align="right">Type of use</ttcol>

          <ttcol align="right">No. of users</ttcol>

          <ttcol align="right">Activ- ity factor</ttcol>

          <ttcol align="right">Ave simultaneous flows /user</ttcol>

          <ttcol align="right">Day rate /user (16hr)</ttcol>

          <ttcol align="right">Day volume /user (16hr)</ttcol>

          <c>Attended</c>

          <c>50</c>

          <c>2.5%</c>

          <c>2</c>

          <c>80kbps</c>

          <c>14MB</c>

          <c>Unattended</c>

          <c>10</c>

          <c>100%</c>

          <c>100</c>

          <c>4.0Mbps</c>

          <c>29GB</c>

          <c> </c>

          <c></c>

          <c></c>

          <c></c>

          <c></c>

          <c></c>

          <c>Aggregate</c>

          <c>60</c>

          <c></c>

          <c>1002.5</c>

          <c>40Mbps</c>

          <c>288GB</c>

          <postamble></postamble>
        </texttable>

        <t>We hope the above examples have clearly illustrated two main
        points:<list style="symbols">
            <t>Rate equality at design time doesn't prevent extreme unfairness
            at run time;</t>

            <t>If extreme unfairness is not corrected, capacity investment
            tends to stop—a concrete consequence of unfairness that
            affects everyone.</t>
          </list></t>

        <t>Finally, note that configuration guidelines for typical p2p
        applications (e.g. BitTorrent calculator <xref
        target="az-calc"></xref>), advise a maximum number of open connections
        that increases roughly linearly with upstream capacity.</t>
      </section>
    </section>
  </back>
</rfc>

PAFTECH AB 2003-20262026-04-22 16:46:54