draft-dolson-sfc-hierarchical.xml

<?xml version="1.0" encoding="US-ASCII"?>
<!DOCTYPE rfc SYSTEM "rfc2629.dtd" [
<!-- One method to get references from the online citation libraries.
     There has to be one entity for each item to be referenced. 
     An alternate method (rfc include) is described in the references. -->

<!ENTITY RFC2119 SYSTEM "http://xml.resource.org/public/rfc/bibxml/reference.RFC.2119.xml">
<!ENTITY I-D.draft-homma-sfc-forwarding-methods-analysis SYSTEM 
"http://xml.resource.org/public/rfc/bibxml3/reference.I-D.draft-homma-sfc-forwarding-methods-analysis-01.xml">
<!ENTITY I-D.draft-ietf-sfc-nsh SYSTEM 
"http://xml.resource.org/public/rfc/bibxml3/reference.I-D.draft-ietf-sfc-nsh-00.xml">
<!ENTITY I-D.draft-ietf-sfc-architecture SYSTEM 
"http://xml.resource.org/public/rfc/bibxml3/reference.I-D.draft-ietf-sfc-architecture-07.xml">
<!ENTITY I-D.draft-ietf-sfc-dc-use-cases SYSTEM 
"http://xml.resource.org/public/rfc/bibxml3/reference.I-D.draft-ietf-sfc-dc-use-cases-02.xml">
]>
<?xml-stylesheet type='text/xsl' href='rfc2629.xslt' ?>
<!-- used by XSLT processors -->
<!-- For a complete list and description of processing instructions (PIs), 
     please see http://xml.resource.org/authoring/README.html. -->
<!-- Below are generally applicable Processing Instructions (PIs) that most I-Ds might want to use.
     (Here they are set differently than their defaults in xml2rfc v1.32) -->
<?rfc strict="yes" ?>
<!-- give errors regarding ID-nits and DTD validation -->
<!-- control the table of contents (ToC) -->
<?rfc toc="yes"?>
<!-- generate a ToC -->
<?rfc tocdepth="4"?>
<!-- the number of levels of subsections in ToC. default: 3 -->
<!-- control references -->
<?rfc symrefs="yes"?>
<!-- use symbolic references tags, i.e, [RFC2119] instead of [1] -->
<?rfc sortrefs="yes" ?>
<!-- sort the reference entries alphabetically -->
<!-- control vertical white space 
     (using these PIs as follows is recommended by the RFC Editor) -->
<?rfc compact="yes" ?>
<!-- do not start each main section on a new page -->
<?rfc subcompact="no" ?>
<!-- keep one blank line between list items -->
<!-- end of list of popular I-D processing instructions -->
<rfc category="info" docName="draft-dolson-sfc-hierarchical-03" ipr="trust200902">

  <!-- ***** FRONT MATTER ***** -->

  <front>

    <title>Hierarchical Service Function Chaining</title>

    <!-- add 'role="editor"' below for the editors if appropriate -->

    <author fullname="David Dolson" initials="D." surname="Dolson">
      <organization>Sandvine</organization>
      <address>
        <postal>
          <street>408 Albert Street</street>
          <!-- Reorder these if your country does things differently -->
          <city>Waterloo</city>
          <region>ON</region>
          <code>N2L 3V3</code>
          <country>Canada</country>
        </postal>
        <phone>+1 519 880 2400</phone>
        <email>ddolson@sandvine.com</email>
        <!-- uri and facsimile elements may also be added -->
      </address>
    </author>

    <author fullname="Shunsuke Homma" initials="S." surname="Homma">
      <organization abbrev="NTT">NTT, Corp.</organization>
      <address>
        <postal>
          <street>3-9-11, Midori-cho</street>
          <city>Musashino-shi</city>
          <region>Tokyo</region>
          <code>180-8585</code>
          <country>Japan</country>
        </postal>
        <email>homma.shunsuke@lab.ntt.co.jp</email>
      </address>
    </author>

    <author fullname="Diego R. Lopez" initials="D. R." surname="Lopez">
      <organization>Telefonica I+D</organization>
      <address>
        <postal>
          <street>Don Ramon de la Cruz, 82</street>
          <city>Madrid</city>
          <region></region>
          <code>28006</code>
          <country>Spain</country>
        </postal>
        <phone>+34 913 129 041</phone>
        <email>diego.r.lopez@telefonica.com</email>
    <!-- uri and facsimile elements may also be added -->
      </address>
    </author>

    <author fullname="Mohamed Boucadair" initials="M." surname="Boucadair">
      <organization>Orange Group</organization>
      <address>
        <postal>
          <street></street>
          <city>Rennes</city>
          <code>35000</code>
          <country>France</country>
        </postal>
        <email> mohamed.boucadair@orange.com </email>
      </address>
    </author>

    <author fullname="Dapeng Liu" initials="D." surname="Liu">
      <organization>Alibaba Group</organization>
      <address>
        <postal>
          <street></street>
          <city>Beijing</city>
          <code>100022</code>
          <country>China</country>
        </postal>
        <email> max.ldp@alibaba-inc.com </email>
      </address>
    </author>


    <date month="Sep" year="2015" />

    <!-- Meta-data Declarations -->

    <area>Routing Area</area>

    <workgroup>Service Function Chaining</workgroup>

    <keyword>sfc</keyword>
    <keyword>hierarchical</keyword>

    <abstract>
      <t>
        Hierarchical Service Function Chaining (hSFC) is a network architecture
        allowing an organization to compartmentalize a large-scale network into
        multiple domains of administration.
      </t>
      <t>
        The goals of hSFC are to make a large-scale network easier to reason 
        about, simpler to control and to support independent functional groups 
        within large operators.
      </t>
    </abstract>
  </front>

  <middle>
    <section title="Introduction">
      <t>
        Service Function Chaining (SFC) is a technique for prescribing 
        differentiated traffic forwarding policies within the SFC domain.
        SFC is described in detail in the
        <xref target="I-D.ietf-sfc-architecture">
         SFC architecture document
        </xref>,
        and is not repeated here.
      </t>
      <t>
        In this document we consider the difficult problem of implementing SFC
        across a large, geographically dispersed network comprised of millions 
        of hosts and thousands of network forwarding elements, involving 
        multiple operational teams (with varying functional responsibilities). 
        We expect asymmetrical routing is inherent in the network, while 
        recognizing that some Service Functions (SFs) require bidirectional 
        traffic for transport-layer sessions (e.g., NATs, firewalls). We assume  
        that some paths need to be selected on the basis of application-specific 
        data visible to the network, with 5-tuple stickiness to specific Service
        Function instances.
      </t>
      <t>
        Note: in this document, the notion of the "path" of a packet is the
        series of SF instances traversed by a packet. The means of delivering 
        packets between SFs (the forwarding mechanisms of the underlay network)
        is not relevant to the current discussion.
      </t>
      <t>
        Difficult problems are often made easier by decomposing them in a 
        hierarchical (nested) manner. So instead of considering an omniscient 
        SFC Control Plane that can manage (create, withdraw, supervise, etc.) 
        complete paths from one end of the network to the other, we decompose 
        the network into smaller sub-domains. Each sub-domain may support a 
        subset of the network applications or a subset of the users. The 
        criteria for determining decomposition into SFC-enabled sub-domains are 
        beyond the scope of this document.
      </t>
      <t>
        Note that decomposing a network into multiple SFC-enabled domains should 
        permit end-to-end visibility of Service Functions and Service Function 
        Paths. Decomposition should also be implemented with special care to 
        ease monitoring and troubleshooting of the network as a whole.
      </t>
      <t>
        An example of simplifying a network by using multiple SF domains is
        further discussed in <xref target="I-D.ietf-sfc-dc-use-cases"/>.
      </t>
      <t>
        We assume the SF technology uses 
        <xref target="I-D.ietf-sfc-nsh">NSH</xref> or a similar labeling 
        mechanism.
      </t>
      <t>
        The "domains" discussed in this document are assumed to be under control 
        of a single organization, such that here is a strong trust relationship 
        between the domains. The intention of creating multiple domains is to 
        improve the ability to operate a network. It is outside of the scope of 
        the document to consider domains operated by different organizations.
      </t>

      <section 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">RFC 2119</xref>.</t>
      </section>
    </section>

    <section title="Hierarchical Service Function Chaining (hSFC)">
      <t>
        A hierarchy has multiple levels. The top-most level encompasses the 
        entire network domain to be managed, and lower levels encompass
        portions of the network.
      </t>
      
      <section title="Top Level">
        <t>
          Considering the example in <xref target="fig_hierarchical_top"/>, a 
          top-level network domain includes SFC components distributed over a 
          wide area, including:
          <list style="symbols">
            <t>classifiers (CFs),</t>
            <t>Service Function Forwarders (SFFs) and</t>
            <t>Sub-domains.</t>
          </list>
          For the sake of clarity, components of the underlay network are not 
          shown; an underlay network is assumed to provide connectivity between 
          SFC components.
        </t>
        <t>
          Top-level service function paths carry packets from classifiers 
          through a series of SFFs and sub-domains, with the operations within
          sub-domains being opaque to the higher levels.
        </t>
        <t>
          We expect the system to include a top-level control-plane having
          responsibility for configuring forwarding and classification.
          The top-level Service Chaining control-plane manages end-to-end 
          service chains and associated service function paths from network edge
          points to sub-domains and configuring top-level classifiers at a
          coarse level (e.g., based on source or destination host) to forward
          traffic along paths that will transit appropriate sub-domains. The
          figure shows one possible service chain passing from edge, through two
          sub-domains, to network egress. The top-level control plane does NOT
          configure classification or forwarding within the sub-domains.
        </t>
        <t>
          At this network-wide level, the number of SFPs required is a linear 
          function of the number of ways in which a packet is required to 
          traverse different sub-domains and egress the network. Note that the 
          various paths which may be taken within a sub-domain are not 
          represented by distinct network-wide SFPs; specific policies at the 
          ingress nodes of each sub-domain bind flows to sub-domain paths.
        </t>
        <t>
          Packets are classified at the edge of the network to select the paths 
          by which sub-domains are to be traversed. At the ingress of each 
          sub-domain, paths are reclassified to select the paths by which SFs in 
          the sub-domain are to be traversed. At the egress of each sub-domain, 
          packets are returned to the top-level paths. Contrast this with an 
          approach requiring the top-level classifier to select paths to specify 
          all of the SFs in each sub-domains.
        </t>
        <t>
          It should be assumed that some service functions in the network 
          require bidirectional symmetry of paths (see more in 
          <xref target="section_classifier"/>). Therefore the classifiers at the 
          top level must be configured with policies ensuring server-to-client
          packets take the reverse path of client-to-server packet through
          sub-domains. (Recall the "path" denotes the series of service 
          functions; the precise physical network path within the underlay 
          network is not relevant here.)
        </t>
        <figure anchor="fig_hierarchical_top"
                title="Network-wide view of Top Level of Hierarchy">
          <artwork><![CDATA[
                 +------------+
                 |Sub-domain#1|
                 |  in DC1    | 
                 +----+-------+
                      |
               .---- SFF1 ------.   +--+
       +--+   /     /  |         \--|CF|
   --->|CF|--/---->'   |          \ +--+
       +--+ /  SC#1    |           \
            |          |            |
            |          V    .------>|--->
            |         /    /        |
            \         |   /        /
       +--+  \        |  /        /  +--+
       |CF|---\       | /        /---|CF|
       +--+    '---- SFF2 ------'    +--+
                      |
                 +----+-------+
                 |Sub-domain#2|
                 |   in DC2   |
                 +------------+
        ]]> </artwork>
          <postamble>
            One path is shown from edge classifier to SFF1 to Sub-domain#1 
            (residing in data-center1) to SFF1 to SFF2 (residing in data-center 
            2) to Sub-domain#2 to SFF2 to network egress.
          </postamble>
       </figure>

      </section>

      <section title="Lower Levels">
        <t>
          Each of the sub-domains in <xref target="fig_hierarchical_top"/> is an 
          SFC domain.
        </t>
        <t>
          Unlike the top level, however, data packets entering the sub-domain 
          are already encapsulated within SFC transport.
          <xref target="fig_hierarchical_lower"/> shows a sub-domain interfaced
          with a higher-level domain by means of an Internal Boundary Node 
          (IBN). It is the purpose of the IBN to remove packets from the SFC
          encapsulation, apply Classification rules, and direct the packets to
          the selected local service function paths terminating at an egress 
          IBN. The egress SFC Domain Gateway finally restores packets to the
          original SFC transport and hands them off to SFFs.
        </t>
        <t>
          Each sub-domain intersects a subset of the total paths that are 
          possible in the higher-level domain. An IBN is concerned with
          higher-level paths, but only those traversing the sub-domain. A
          top-level controller may configure the IBN as an SF (the IBN plays the
          SF role in the top-level domain).
        </t>
        <t>
          We expect each sub-domain to have a control-plane that can operate 
          independently of the top-level control-plane. The sub-domain 
          control-plane configures the classification and forwarding rules in 
          the sub-domain. The classification rules reside in the IBN, where 
          packets are moved from SFC encapsulation of the top-level domain to 
          and from SFC encapsulation of the lower-level domain.
        </t>
        <figure anchor="fig_hierarchical_lower"
                title="Sub-domain within a higher-level domain">
          <artwork><![CDATA[
  +----+    +-----+  +----------------------+   +-----+
  |    |SC#1| SFF |  |       IBN 1          |   | SFF |
->|    |================*                *===============>
  |    |    +-----+  |  # (in DC 1)      #  |   +-----+
  | CF |             |  V                #  |
  |    |             |+---+            +---+|   Top domain
  |    |    * * * * *||CF | * * * * * *|SFF|| * * * * *
  |    |    *        |+---+            +-+-+|         *
  +----+    *        | | |              | | |    Sub  *
            *        +-o-o--------------o-o-+   domain*
            *     SC#2 | |SC#1          ^ ^       #1  *
            *    +-----+ |              | |           *
            *    |       V              | |           *
            *    |     +---+  +------+  | |           *
            *    |     |SFF|->|SF#1.1|--+ |           *
            *    |     +---+  +------+    |           *
            *    V                        |           *
            *  +---+  +------+  +---+  +------+       *
            *  |SFF|->|SF#2.1|->|SFF|->|SF#2.2|       *
            *  +---+  +------+  +---+  +------+       *
            * * * * * * * * * * * * * * * * * * * * * *
       ]]> </artwork>
          <postamble>
         *** Sub-domain boundary; === top-level chain; --- low-level chain.
          </postamble>
        </figure>

        <t>
          If desired, the pattern can be applied recursively. For example, 
          SF#1.1 in <xref target="fig_hierarchical_lower"/> could be a 
          sub-domain of the sub-domain.
        </t>
        
      </section>
    </section>

    <section title="Internal Boundary Node (IBN)">
      <t>
        A network element termed "Internal Boundary Node" (IBN) bridges packets 
        between domains. It looks like an SF to the higher level, and looks like 
        a classifier and end-of-chain to the lower level.
      </t>
      <t>
        To achieve the benefits of hierarchy, the IBN should be applying more
        granular traffic classification rules at the lower level than the
        traffic passed to it. This means that the number of SF Paths within the
        lower level is greater than the number of SF Paths arriving to the
        IBN.
      </t>
      <t>
        The IBN is also the termination of lower-level SF paths. This is because
        the packets exiting lower-level SF paths must be returned to the
        higher-level SF paths and forwarded to the next hop in the higher-level
        domain.
      </t>

      <section title="IBN Path Configuration">
        <t>
          An operator of a lower-level SF Domain may be aware of which 
          high-level paths transit their domain, or they may wish to accept any
          paths. 
        </t>
        <t>
          When packets enter the sub-domain, the Path Identifier and Path Index 
          are re-marked according to the path selected by the classifier.
        </t>
        <t>
          After exiting a path in the sub-domain, packets can be restored to an 
          upper-level SF path by these methods:
        <list style="numbers">
          <t>
            Stateful per flow,
          </t>
          <t>
            Pushing path identifier into metadata,
          </t>
          <t>
            Using unique lower-level paths per upper-level path.
          </t>
        </list>
        </t>

        <section title="Flow-Stateful IBN">
          <t>
            An IBN can be flow-aware, returning packets to the correct
            higher-level SF path on the basis of 5-tuple of packets exiting the
            lower-level SF paths.
          </t>
          <t>
            When packets are received by the IBN on a higher-level path, the
            encapsulated packets are parsed for IP and transport-layer (TCP or
            UDP) coordinates. State is created, indexed by the 5-tuple of
            {source-IP, destination-IP, source-port, destination-port and
            transport protocol}. The state contains critical fields of the
            encapsulating SFC header (or perhaps the entire header).
          </t>
          <t>
            The simplest approach has the packets return to the same IBN at the
            end of the chain that classified the packet at the start of the
            chain. This is because the required 5-tuple state is rapidly
            changing and most efficiently kept locally. If the packet is 
            returned to a different IBN for egress, 5-tuple state must be
            synchronized between the IBNs.
          </t>
          <t>
            When a packet returns to the IBN at the end of a chain, the SFC
            header is removed, the packet is parsed for IP and transport-layer
            coordinates, and state is retrieved by the 5-tuple of the packet.
            The state contains the information required to forward the packet
            within the higher-level service chain.
          </t>
          <t>
            State cannot be created by packets arriving from the lower-level 
            chain; when state cannot be found for such packets, they MUST be 
            dropped.
          </t>
          <t>
            This stateful approach is limited to use with SFs that retain the 
            5-tuple of the packet. This approach cannot be used with SFs that
            modify the 5-tuple (e.g., as done by a NAT) or otherwise create
            packets for new 5-tuples other than those received (e.g., as an HTTP
            cache might do to retrieve content on behalf of the original flow).
            In both cases, the fundamental problem is the inability to forward
            packets when state cannot be found for the packet 5-tuples.
          </t>
          <t>
            In the stateful approach, there are issues caused by the state, such 
            as how long the state should be maintained (it MUST time out 
            eventually), as well as whether the state needs to be replicated to 
            other devices to create a highly available network.
          </t>
          <t>
            It is valid to consider the state disposable after failure, since it
            can be re-created by each new packet arriving from the higher-level
            domain. For example, if an IBN loses all flow state, the state is
            re-created by an end-point retransmitting a TCP packet.
          </t>
          <t>
            If an SFC domain handles multiple network regions (e.g., multiple 
            private networks), the 5-tuple may be augmented with a 6th 
            parameter, perhaps using some metadata to identify the network 
            region.
          </t>
          <t>
            In this stateful approach, it is not necessary for the sub-domain's 
            control-plane to modify paths when higher-level paths are changed.
            The complexity of the higher-level domain does not cause complexity
            in the lower-level domain.
          </t>
        </section>

        <section title="Encoding Upper-Level Paths in Metadata">
          <t>
            An IBN can push the upper-level service path identifier (SPI) and
            service index (SI) (or encoding thereof) into a metadata field of
            the lower-level encapsulation (e.g., placing upper-level path
            information into a metadata field of NSH). When packets exit the
            lower-level path, the upper-level SPI and SI can be restored from
            the metadata retrieved from the packet.
          </t>
          <t>
            This approach requires the SFs in the path to be capable of 
            forwarding the metadata and appropriately attaching metadata to any 
            packets injected for a flow.
          </t>
          <t>
            Using new metadata may inflate packet size when variable-length
            metadata (type 2 from <xref target="I-D.ietf-sfc-nsh">NSH</xref>)
            is used.
          </t>
          <t>
            It is conceivable that the MD-type 1 Mandatory Context Header fields 
            of
            <xref target="I-D.ietf-sfc-nsh">NSH</xref> are not all relevant to
            the lower-level domain. In this case, one of the metadata slots of
            the Mandatory Context Header could be repurposed within the 
            lower-level domain, and restored when leaving.
          </t>
          <t>
            In this metadata approach, it is not necessary for the sub-domain's 
            controller to modify paths when higher-level paths are changed.
            The complexity of the higher-level domain does not cause complexity
            in the lower-level domain.
          </t>
        </section>

        <section title="Using Unique Paths per Upper-Level Path">
          <t>
            In this approach, paths within the sub-domain are constrained so 
            that a path identifier (of the sub-domain) unambiguously indicates 
            the egress path (of the upper domain). This allows the original path 
            information to be restored at sub-domain egress from a look-up table 
            using the sub-domain path identifier.
          </t>
          <t>
            Whenever the upper-level domain provisions a path via the 
            lower-level domain, the lower-level domain controller must provision 
            corresponding paths to traverse the lower-level domain.
          </t>
          <t>
            A down-side of this approach is that the number of paths in the 
            lower-level domain is multiplied by the number of paths in the 
            higher-level domain that traverse the lower-level domain.
            I.e., a sub-path must be created for each combination of upper Path 
            identifier and lower path.
          </t>
        </section>

      </section>
      
      <section title="Gluing Levels Together">
        <t>
          The path identifier or metadata on a packet received by the IBN may be
          used as input to reclassification and path selection within the
          lower-level domain.
        </t>
        <t>
          In some cases the meanings of the various path IDs and metadata must 
          be coordinated between domains.
        </t>
        <t>
          One approach is to use well-known identifier values in metadata, 
          communicated by some organizational registry.
        </t>
        <t>
          Another approach is to use well-known labels for path identifiers or 
          metadata, as an indirection to the actual identifiers. The actual 
          identifiers can be assigned by control-plane systems. For example, a 
          sub-domain classifier could have a policy, "if pathID=classA then 
          chain packet to path 1234"; the higher-level controller would be 
          expected to configure the concrete higher-level pathID for classA.
        </t>

      </section>
    </section>

    <section title="Sub-domain Classifier" anchor="section_classifier">
      <t>
        Within the sub-domain (referring to <xref target="fig_hierarchical_lower"/>),
        after the IBN removes higher-level encapsulation from incoming packets,
        it sends the packets to the classifier, which selects the encapsulation
        for the packet within the sub-domain.
      </t>
      <t>
        One of the goals of the hierarchical approach is to make it easy to 
        have transport-flow-aware service chaining with bidirectional paths. For
        example, it is desired that for each TCP flow, the client-to-server 
        packets traverse the same SFs as the server-to-client packets, but in 
        the opposite sequence. We call this bidirectional symmetry. If 
        bidirectional symmetry is required, it is the responsibility of the 
        control-plane to be aware of symmetric paths and configure the
        classifier to chain the traffic in a symmetric manner.
      </t>
      <t>
        Another goal of the hierarchical approach is to simplify the mechanisms 
        of scaling in and scaling out service functions.
        All of the complexities of load-balancing among multiple SFs can be 
        handled within a sub-domain, under control of the classifier, allowing 
        the higher-level domain to be oblivious to the existence of multiple SF 
        instances.
      </t>
      <t>
        Considering the requirements of bidirectional symmetry and 
        load-balancing, it is useful to have all packets entering a sub-domain 
        to be received by the same classifier or a coordinated cluster of 
        classifiers. There are both stateful and stateless approaches to 
        ensuring bidirectional symmetry.
      </t>
    </section>

    <section title="Control Plane Elements">
      <t>
        Controllers have been mentioned in this document without much
        explanation. Although control protocols have not yet been standardized, 
        from the point of view of hierarchical service function chaining we have 
        these expectations:
        <list style="symbols">
          <t>
            Each control-plane instance manages a single level of hierarchy of a
            single domain.
          </t>
          <t>
            Each control-plane is agnostic about other levels of hierarchy. This 
            aspect allows humans to reason about the system within a single 
            domain and allows control-plane algorithms to use only domain-local 
            inputs. Top-level control does not need visibility to sub-domain 
            policies, nor does sub-domain control need visibility to 
            higher-level policies.
          </t>
          <t>
            Sub-domain control-planes are agnostic about control-planes of other
            sub-domains. This allows both humans and machines to manipulate 
            sub-domain policy without considering policies of other domains.
          </t>
        </list>
      </t>
      <t>
        Recall that the IBN acts as an SF in the higher-level domain (receiving
        SF instructions from the higher-level control-plane) and as a classifier
        in the lower-level domain (receiving classification rules from the
        sub-domain control-plane). In this view, it is the IBN that glues the
        layers together.
      </t>
      <t>
        The above expectations are not intended to prohibit network-wide 
        control. A control hierarchy can be envisaged to distribute information 
        and instructions to multiple domains and sub-domains. Control hierarchy 
        is outside the scope of this document.
      </t>

    </section>

    <section anchor="Acknowledgements" title="Acknowledgements">
      <t>The concept of Hierarchical Service Path Domains was introduced in
      <xref target="I-D.homma-sfc-forwarding-methods-analysis">
      draft-homma-sfc-forwarding-methods-analysis-01</xref>
      as a means to improve scalability of service chaining in large networks. 
      </t>
      <t>
        The authors would like to thank the following individuals for taking the
        time to read and provide valuable feedback:
      </t>
      <t>
        <list style="hanging">
          <t> Ron Parker </t>
          <t> Christian Jacquenet </t>
          <t> Jie Cao </t>
        </list>
      </t>
    </section>

    <!-- Possibly a 'Contributors' section ... -->

    <section anchor="IANA" title="IANA Considerations">
      <t>This memo includes no request to IANA.</t>
    </section>

    <section anchor="Security" title="Security Considerations">
      <t>
        Hierarchical service function chaining makes use of service chaining 
        architecture, and hence inherits the security considerations described 
        in the architecture document.
      </t>
      <t>
        Furthermore, hierarchical service function chaining inherits security 
        considerations of the data-plane protocols (e.g., NSH) and control-plane 
        protocols used to realize the solution.
      </t>
      <t>
        The systems described in this document bear responsibility for 
        forwarding internet traffic. In some cases the systems are responsible 
        for maintaining separation of traffic in private networks.
      </t>
      <t>
        This document describes systems within different domains of
        administration that must have consistent configurations in order to 
        properly forward traffic and to maintain private network separation.
        Any protocol designed to distribute the configurations must be secure 
        from tampering.
      </t>
      <t>
        All of the systems and protocols must be secure from modification by 
        untrusted agents.
      </t>
    </section>
  </middle>

  <!--  *****BACK MATTER ***** -->

  <back>

    <references title="Normative References">
      &RFC2119;


    </references>

    <references title="Informative References">

      &I-D.draft-homma-sfc-forwarding-methods-analysis;

      &I-D.draft-ietf-sfc-nsh;

      &I-D.draft-ietf-sfc-architecture;

      &I-D.draft-ietf-sfc-dc-use-cases;

    </references>

    <!-- Appendix -->
    <section title="Examples of Hierarchical Service Function Chaining">
      <t>
        The advantage of hierarchical service function chaining compared with
        normal or flat service function chaining is that it can reduce the
        management complexity significantly. This section discusses examples
        that show the advantage of hierarchical service function chaining. 
      </t>

      <section title="Reducing the Number of Service Function Paths">
        <t>
          In this use case, hierarchical service function chaining is used to 
          simplify service function chaining management by reducing the number 
          of Service Function Paths.
        </t>

        <t>
          As shown in <xref target="fig_example_flat"/>,
          there are two domains each with different concerns: a Security Domain 
          that selects Service Functions based on network conditions and an 
          Optimization Domain that selects Service Functions based on traffic 
          protocol.
        </t>
        <t>
          There are five security functions deployed in the Security Domain. The
          Security Domain operator wants to enforce the five different security
          policies, and the Optimization Domain operator wants to apply
          different optimization (either cache or video optimization to each of 
          these two types of traffic. If we use flat SFC (normal branching), 10 
          SFPs are needed in each domain. In contrast, if we use hierarchical 
          SFC, only 5 SFPs in Security Domain and 2 SFPs in Optimization 
          Domain will be required, as shown in
          <xref target="fig_example_hsfc"/>.
        </t>
        <t>
          In the flat model, the number of SFPs is the product of the number of 
          functions in all of the domains. In the hSFC model, the number of SFPs 
          is the sum of the number of functions. For example, adding a "bypass" 
          path in the Optimization Domain would cause the flat model to require 
          15 paths (5 more), but cause the hSFC model to require one more path 
          in the Optimization Domain.
        </t>
        <figure anchor="fig_example_flat"
                title="Flat SFC (Normal Branching)">
          <artwork><![CDATA[
           . . . . . . . . . . . .   . . . . . . . . . . . . .
           . Security Domain     .   .  Optimization Domain  .
           .                     .   .                       .
           .    +-1---[     ]----------------->[Cache  ]------->
           .    |     [ WAF ]    .   .                       .
           .    +-2-->[     ]----------------->[Video Opt.]---->
           .    |                .   .                       .    
           .    +-3---[Anti ]----------------->[Cache  ]------->
           .    |     [Virus]    .   .                       .
           .    +-4-->[     ]----------------->[Video Opt.]---->
           .    |                .   .                       .
           .    +-5-->[     ]----------------->[Cache  ]------->
[DPI]--->[CF]---|     [ IPS ]    .   .                       .
           .    +-6-->[     ]----------------->[Video Opt.]---->
           .    |                .   .                       .
           .    +-7-->[     ]----------------->[Cache  ]------->
           .    |     [ IDS ]    .   .                       .
           .    +-8-->[     ]----------------->[Video Opt.]---->
           .    |                .   .                       .
           .    +-9-->[Traffic]--------------->[Cache  ]------->
           .    |     [Monitor]  .   .                       .
           .    +-10->[       ]--------------->[Video Opt.]---->
           . . . . . . . . . . . .   . . . . . . . . . . . . .
       ]]> </artwork>
       <postamble>
         The classifier must select paths that determine the combination of 
         Security and Optimization concerns. 1:WAF+Cache, 2:WAF+VideoOpt, 
         3:AntiVirus+Cache, 4:AntiVirus+VideoOpt, 5: IPS+Cache, 6:IPS+VideoOpt,
         7:IDS+Cache, 8:IDS+VideoOpt, 9:TrafficMonitor+Cache, 
         10:TrafficMonitor+VideoOpt
       </postamble>
        </figure>

        <figure anchor="fig_example_hsfc"
                title="Simplified Path Management with Hierarchical SFC">
          <artwork><![CDATA[
     . . . . . . . . . . . . . . .    . . . . . . . . . . . . . . .
     .     Security Domain       .    .   Optimization Domain     .
     .                           .    .                           .
[CF]---->[  [CF]    IBN      ]---------->[  [CF]   IBN         ]---->
     .    |                  ^   .    .  |                     ^  .
     .    +----->[ WAF ]-----+   .    .  +-->[ Cache ]---------+  .
     .    |                  |   .    .  |                     |  .
     .    +-->[Anti-Virus]---+   .    .  +-->[Video Opt]-------+  .
     .    |                  |   .    .                           .
     .    +----->[ IPS ]-----+   .    . . . . . . . . . . . . . . .
     .    |                  |   .
     .    +----->[ IDS ]-----+   .
     .    |                  |   .
     .    +-->[ Traffic ]----+   .
     .        [ Monitor ]        .
     . . . . . . . . . . . . . . .
       ]]> </artwork>
        </figure>
      </section>

      <section title="Managing a Distributed Data-Center Network">
        <t>
          Hierarchical service function chaining can be used to simplify 
          inter-data-center SFC management. In the example of 
          <xref target="fig_example_hsfc_inter_dc"/>,
          shown below, there is a central data center (Central DC)
          and multiple local data centers (Local DC#1, #2, #3)
          that are deployed in a geographically distributed manner.
          All of the data centers are under a single administrative domain. 
        </t>
        <t>
          The central DC may have some service functions that the local DC needs,
          such that the local DC needs to chain traffic via the central DC.
          This could be because:
          <list style="symbols">
            <t>
              Some service functions are deployed as dedicated hardware 
              appliances, and there is a desire to avoid the capital and 
              operation expenses (CAPEX and OPEX respectively) of deploying such 
              service functions in all data centers.
            </t>
            <t>
              Consider the case when service functions are being trialed,
              introduced or otherwise handle a relatively small amount of 
              traffic. It may be cheaper to manage these service functions in a 
              single central data center and steer packets to the central data 
              center than to manage these service functions in all data centers.
            </t>
            
          </list>
        </t>

        <figure anchor="fig_example_hsfc_inter_dc"
                title="Simplify Inter-DC SFC Management">
          <artwork><![CDATA[
                +-----------+
                |Central DC |
                +-----------+
                   ^  ^   ^
                   |  |   |
               .---|--|---|----.    
              /   /   |   |      \
             /   /    |    \      \ 
  +-----+   /   /     |     \      \    +-----+
  |Local|  |   /      |      \     |    |Local|
  |DC#1 |--|--.       |       .----|----|DC#3 |
  +-----+  |          |            |    +-----+
            \         |            /
             \        |           /  
              \       |          /   
               '----------------'    
                      |      
                   +-----+
                   |Local|
                   |DC#2 |
                   +-----+                          
                   ]]> 
          </artwork>
        </figure>
        
        <t>
          For large data center operators, one local DC may have tens of thousands of servers
          and hundred of thousands of virtual machines. SFC can be used to manage user traffic. 
          For example, SFC can be used to classify user traffic based on service type, DDoS state etc.
        </t>

        <t>
          In such large scale data center, using flat SFC
          is very complex, requiring a super-controller to configure all data centers.
          For example, any changes 
          to Service Functions or Service Function Paths in the central DC (e.g., deploying a new SF) would
          require updates to all of the Service Function Paths in the local DCs accordingly.
          Furthermore, requirements for symmetric paths add additional complexity when
          flat SFC is used in this scenario.
        </t>
        <t>
          Conversely, if using hierarchical SFC, each data center can be managed independently
          and the management complexity could be reduced significantly. Service Function Paths between 
          data centers can represent abstract notions without regard to details within data centers.
          Independent controllers can be used for the top level (getting packets to pass the correct
          data centers) and local levels (getting packets to specific SF instances).
        </t>


      </section>
      
    </section>

  </back>
</rfc>