BESS                                                            Z. Zhang
Internet-Draft                                          Juniper Networks
Intended status: Standards Track                               R. Raszuk
Expires: May 21, 2020                                       Bloomberg LP
                                                              D. Pacella
                                                                A. Gulko
                                                         Thomson Reuters
                                                       November 18, 2019

                Controller Based BGP Multicast Signaling


   This document specifies a way that one or more centralized
   controllers can use BGP to set up a multicast distribution tree in a
   network.  In the case of labeled tree, the labels are assigned by the
   controllers either from the controllers' local label spaces, or from
   a common Segment Routing Global Block (SRGB), or from each routers
   Segment Routing Local Block (SRLB) that the controllers learn.  In
   case of labeled unidirectional tree and label allocation from the
   common SRGB or from the controllers' local spaces, a single common
   label can be used for all routers on the tree to send and receive
   traffic with.  Since the controllers calculate the trees, they can
   use sophisticated algorithms and constraints to achieve traffic

Requirements Language

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

Status of This Memo

   This Internet-Draft is submitted in full conformance with the
   provisions of BCP 78 and BCP 79.

   Internet-Drafts are working documents of the Internet Engineering
   Task Force (IETF).  Note that other groups may also distribute
   working documents as Internet-Drafts.  The list of current Internet-
   Drafts is at

Zhang, et al.             Expires May 21, 2020                  [Page 1]

Internet-Draft            bgp-mcast-controller             November 2019

   Internet-Drafts are draft documents valid for a maximum of six months
   and may be updated, replaced, or obsoleted by other documents at any
   time.  It is inappropriate to use Internet-Drafts as reference
   material or to cite them other than as "work in progress."

   This Internet-Draft will expire on May 21, 2020.

Copyright Notice

   Copyright (c) 2019 IETF Trust and the persons identified as the
   document authors.  All rights reserved.

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

Table of Contents

   1.  Overview  . . . . . . . . . . . . . . . . . . . . . . . . . .   3
     1.1.  Introduction  . . . . . . . . . . . . . . . . . . . . . .   3
     1.2.  Resilience  . . . . . . . . . . . . . . . . . . . . . . .   4
     1.3.  Signaling . . . . . . . . . . . . . . . . . . . . . . . .   5
     1.4.  Label Allocation  . . . . . . . . . . . . . . . . . . . .   5
       1.4.1.  Using a Common per-tree Label for All Routers . . . .   6
       1.4.2.  Upstream-assignment from Controller's Local Label
               Space . . . . . . . . . . . . . . . . . . . . . . . .   7
     1.5.  Determining Root/Leaves . . . . . . . . . . . . . . . . .   8
       1.5.1.  PIM-SSM/Bidir or mLDP P2MP  . . . . . . . . . . . . .   9
       1.5.2.  PIM ASM . . . . . . . . . . . . . . . . . . . . . . .   9
   2.  Specification . . . . . . . . . . . . . . . . . . . . . . . .   9
     2.1.  Additional Tunnel Types for TEA . . . . . . . . . . . . .   9
       2.1.1.  Any-Encapsulation Tunnel  . . . . . . . . . . . . . .   9
       2.1.2.  Load-balancing Tunnel . . . . . . . . . . . . . . . .  10
     2.2.  RPF Label Stack Sub-TLV . . . . . . . . . . . . . . . . .  10
     2.3.  Context Label Wide Community  . . . . . . . . . . . . . .  10
     2.4.  Procedures  . . . . . . . . . . . . . . . . . . . . . . .  10
   3.  Security Considerations . . . . . . . . . . . . . . . . . . .  10
   4.  IANA Considerations . . . . . . . . . . . . . . . . . . . . .  11
   5.  Acknowledgements  . . . . . . . . . . . . . . . . . . . . . .  11
   6.  References  . . . . . . . . . . . . . . . . . . . . . . . . .  11
     6.1.  Normative References  . . . . . . . . . . . . . . . . . .  11
     6.2.  Informative References  . . . . . . . . . . . . . . . . .  12

Zhang, et al.             Expires May 21, 2020                  [Page 2]

Internet-Draft            bgp-mcast-controller             November 2019

   Authors' Addresses  . . . . . . . . . . . . . . . . . . . . . . .  12

1.  Overview

1.1.  Introduction

   [I-D.zzhang-bess-bgp-multicast] describes a way to use BGP as a
   replacement signaling for PIM [RFC7761] or mLDP [RFC6388].  The BGP-
   based multicast signaling described there provides a mechanism for
   setting up both (s,g)/(*,g) multicast trees (as PIM does, but
   optionally with labels) and labeled (MPLS) multicast tunnels (as mLDP
   does).  Each router on a tree performs essentially the same
   procedures as it would perform if using PIM or mLDP, but all the
   inter-router signaling is done using BGP.

   These procedures allow the routers to set up a separate tree for each
   individual multicast (x,g) flow where the 'x' could be either 's' or
   '*', but they also allow the routers to set up trees that are used
   for more than one flow.  In the latter case, the trees are often
   referred to as "multicast tunnels" or "multipoint tunnels", and
   specifically in this document they are mLDP tunnels (except that they
   are set up with BGP signaling).  While it actually does not have to
   be restricted to mLDP tunnels, mLDP FEC is conveniently borrowed to
   identify the tunnel.  In the rest of the document, the term tree and
   tunnel are used interchangeably.

   The trees/tunnels are set up using the "receiver-initiated join"
   technique of PIM/mLDP, hop by hop from downstream routers towards the
   root.  The BGP messages are either sent hop by hop between downstream
   routers and their upstream neighbors, or can be reflected by Route
   Reflectors (RRs).

   As an alternative to each hop independently determining its upstream
   router and signaling upstream towards the root (following PIM/mLDP
   model), the entire tree can be calculated by a centralized
   controller, and the signaling can be entirely done from the
   controller, using the same BGP messages as defined in
   [I-D.zzhang-bess-bgp-multicast].  For that, some additional
   procedures and optimizations are specified in this document.

   While it is outside the scope of this document, signaling from the
   controllers could be done via other means as well, like Netconf or
   any other SDN methods.

Zhang, et al.             Expires May 21, 2020                  [Page 3]

Internet-Draft            bgp-mcast-controller             November 2019

1.2.  Resilience

   Each router could establish direct BGP sessions with one or more
   controllers, or it could establish BGP sessions with RRs who in turn
   peer with controllers.  For the same tree/tunnel, each controller may
   independently calculate the tree/tunnel and signal the routers on the
   tree/tunnel using MCAST-TREE S-PMSI/Leaf A-D routes
   [I-D.zzhang-bess-bgp-multicast].  How the tree/tunnel roots/leaves
   are discovered and how the calculation is done are outside the scope
   of this document.

   On each router, BGP route selection rules will lead to one
   controller's route for the tree/tunnel being selected as the active
   route and used for setting up forwarding state.  As long as all the
   routers on a tree/tunnel consistently pick the same controller's
   routes for the tree/tunnel, the setup should be consistent.  If the
   tree/tunnel is labeled, different labels will be used from different
   controllers so there is no traffic loop issue even if the routers do
   not consistently select the same controlle's routes.  In the
   unlabeled case, to ensure the consistency the selection SHOULD be
   solely based on the identifier of the controller, which could be
   carried in an Address Specific Extended Community (EC).

   Another consistency issue is when a bidirectional tree/tunnel needs
   to be re-routed.  Because this is no longer triggered hop-by-hop from
   downstream to upstream, it is possible that the upstream change
   happens before the downstream, causing traffic loop.  In the
   unlabeled case, there is no good solution (other than that the
   controller issues upstream change only after it gets acknowledgement
   from downstream).  In the labeled case, as long as a new label is
   used there should be no problem.

   Besides the traffic loop issue, there could be transient traffic loss
   before both the upstream and downstream's forwarding state are
   updated.  This could be mitigated if the upstream keep sending
   traffic on the old path (in addition to the new path) and the
   downstream keep accepting traffic on the old path (but not on the new
   path) for some time.  It is a local matter when for the downstream to
   switch to the new path - it could be data driven (e.g., after traffic
   arrives on the new path) or timer driven.

   For each tree, multiple disjoint instances could be calculated and
   signaled for live-live protection.  Different labels are used for
   different instances, so that the leaves can differentiate incoming
   traffic on different instances.  As far as transit routers are
   concerned, the instances are just independent.  Note that the two
   instances are not expected to share common transit routers (it is
   otherwise outside the scope of this document/revision).

Zhang, et al.             Expires May 21, 2020                  [Page 4]

Internet-Draft            bgp-mcast-controller             November 2019

1.3.  Signaling

   Each router only receives S-PMSI/Leaf A-D routes from the controllers
   but does not originate or re-advertise those routes.  The re-
   advertisement of a received route can be blocked based on the fact
   that a configured import RT matches the RT of the route, which
   indicates that this router is the target and consumer of the route
   hence it should not be re-advertised further.  The routes includes
   the outgoing forwarding information in the form of Tunnel
   Encapsulation Attributes (TEA) [I-D.ietf-idr-tunnel-encaps], with
   optional enhancements specified in this document.  The router infers
   the incoming forwarding information from the Upstream Router's IP
   Address field in the NLRI in case of an unlabeled tree.

   Suppose that for a particular tree, there are two downstream routers
   D1 and D2 for a particular upstream router U.  A controller C may
   send two Leaf A-D routes to U, as if the two routes were originated
   by D1 and D2 but reflected by the controller.  As an alternative in
   case of a labeled tree, C could just send one route to U, with a TEA
   specifying both downstreams.  In this case, the Originating Router's
   Address field of the Leaf A-D route is set to the controller's
   address.  Note that for a TEA attached to a unicast NLRI, only one of
   the tunnels in a TEA is used for forwarding a particular packet,
   while all the tunnels in a TEA are used to reach multiple endpoints
   when it is attached to a multicast NLRI.

   Note that, in case of labeled trees, the (x,g) or mLDP FEC signaling
   is actually not needed to transit routers but only needed on tunnel
   root/leaves.  However, for consistency, the same signaling is used to
   all routers.

1.4.  Label Allocation

   In the case of labeled multicast signaled hop by hop towards the
   root, whether it's (x,g) multicast or "mLDP" tunnel, labels are
   assigned by a downstream router and advertised to its upstream router
   (from traffic direction point of view).  In the case of controller
   based signaling, routers do not originate tree join (S-PMSI/Leaf A-D)
   routes anymore, so the controllers have to assign labels on behalf of
   routers, and there are three options for label assignment:

   o  From each router's SRLB that the controller learns

   o  From the common SRGB that the controller learns

   o  From the controller's local label space

Zhang, et al.             Expires May 21, 2020                  [Page 5]

Internet-Draft            bgp-mcast-controller             November 2019

   Assignment from each router's SRLB is no different from each router
   assigning labels from its own local label space in the hop-by-hop
   signaling case.  The assignments for a router is independent of
   assignments for another router, even for the same tree.

   Assignment from the controller's local label space is upstream-
   assigned [RFC5331].  It is used if the controller does not learn the
   common SRGB or each router's SRLB.  Assignment from the SRGB
   [RFC8402] is only meaningful if all SRGBs are the same and a single
   common label is used for all the routers on a tree in case of
   unidirectional tree/tunnel (Section 1.4.1).  Otherwise, assignment
   from SRLB is preferred.

   The choice of which of the options to use depends on many factors.
   An operator may want to use a single common label per tree for ease
   of monitoring and debugging, but that requires explicit RPF checking
   and either SRGB or upstream assigned labels, which may not be
   supported due to either the software or hardware limitations (e.g.
   label imposition/disposition limits).  In an SR network, assignment
   from the common SRGB if it's required to use a single common label
   per unidirectional tree, or otherwise assignment from SRLB is a good
   choice because it does not require support for context label spaces.

1.4.1.  Using a Common per-tree Label for All Routers

   MPLS labels only have local significance.  For an LSP that goes
   through a series of routers, each router allocates a label
   independently and it swaps the incoming label (that it advertised to
   its upstream) to an outgoing label (that it received from its
   downstream) when it forwards a labeled packet.  Even if the incoming
   and outgoing labels happen to be the same on a particular router,
   that is just incidental.

   With Segment Routing, it is becoming a common practice that all
   routers use the same SRGB so that a SID maps to the same label on all
   routers.  This makes it easier for operators to monitor and debug
   their network.  The same concept applies to multicast trees as well -
   a common per-tree label is used for a router to receive traffic from
   its upstream neighbor and replicate traffic to all its downstream

   However, a common per-tree label can only be used for unidirectional
   trees.  Additionally, it requires each router to do explicit RPF
   check, so that only packets from its expected upstream neighbor are
   accepted.  Otherwise, traffic loop may form during topology changes,
   because the forwarding state update is no longer ordered.

Zhang, et al.             Expires May 21, 2020                  [Page 6]

Internet-Draft            bgp-mcast-controller             November 2019

   Traditionally, p2mp mpls forwarding does not require explicit RPF
   check as a downstream router advertises a label only to its upstream
   router and all traffic with that incoming label is presumed to be
   from the upstream router and accepted.  When a downstream router
   switches to a different upstream router a different label will be
   advertised, so it can determine if traffic is from its expected
   upstream neighbor purely based on the label.  Now with a single
   common label used for all routers on a tree to send and receive
   traffic with, a router can no longer determine if the traffic is from
   its expected neighbor just based on that common tree label.
   Therefore, explicit RPF check is needed.  Instead of interface based
   RPF checking as in PIM case, neighbor based RPF checking is used - a
   label identifying the upstream neighbor precedes the tree label and
   the receiving router checks if that preceding neighbor label matches
   its expected upstream neighbor.  Notice that this is similar to
   what's described in Section "9.1.1 Discarding Packets from Wrong PE"
   of RFC 6513 (an egress PE discards traffic sent from a wrong ingress
   PE).  The only difference is one is used for label based forwarding
   and the other is used for (s,g) based forwarding. [note: for
   bidirectional trees, we may be able to use two labels per tree - one
   for upstream traffic and one for downstream traffic.  This needs
   further verification].

   Both the common per-tree label and the neighbor label are allocated
   either from the common SRGB or from the controller's local label
   space.  In the latter case, an additional label identifying the
   controller's label space is needed, as described in the following

1.4.2.  Upstream-assignment from Controller's Local Label Space

   In this case in the multicast packet's label stack the tree label and
   upstream neighbor label (if used in case of single common-label per
   tree) are preceded by a downstream-assigned "context label".  The
   context label identifies a context-specific label space (the
   controller's local label space), and the upstream-assigned label that
   follows it is looked up in that space.

   This specification requires that, in case of upstream-assignment from
   a controller's local label space, each router D to assign,
   corresponding to each controller C, a context label that identifies
   the upstream-assigned label space used by that controller.  This
   label, call it Lc-D, is communicated by D to C.

   Suppose a controller is setting up unidirectional tree T.  It assigns
   that tree the label Lt, and assigns label Lu to identify router U
   which is the upstream of router D on tree T.  C needs to tell U: "to
   send a packet on the given tree/tunnel, one of the things you have to

Zhang, et al.             Expires May 21, 2020                  [Page 7]

Internet-Draft            bgp-mcast-controller             November 2019

   do is push Lt onto the packet's label stack, then push Lu, then push
   Lc-D onto the packet's label stack, then unicast the packet to D".
   Controller C also needs to inform router D of the correspondence
   between <Lc-D, Lu, Lt> and tree T.

   To achieve that, when C sends an S-PMSI/Leaf A-D route, for each
   tunnel in the TEA, it includes a label stack Sub-TLV
   [I-D.ietf-idr-tunnel-encaps], with the outer label being the context
   label Lc-D (received by the controller from the corresponding
   downstream), the next label being the upstream neighbor label Lu, and
   the inner label being the label Lt assigned by the controller for the
   tree.  The router receiving the route will use the label stacks to
   send traffic to its downstreams.

   For C to signal the expected label stack for D to receive traffic
   with, we overload a tunnel TLV in the TEA of the Leaf A-D route sent
   to D - if the remote endpoint of that tunnel TLV matches the Upstream
   Router field in the Leaf A-D route, then it indicates that this is
   actually for receiving traffic from the upstream.  If a common tree
   label is used, then the TLV contains a variant of the Label Stack
   Sub-TLV because the D needs to treat the second inner most label as
   the upstream neighbor label and set up forwarding state accordingly
   for explicit RPF check.  This variant is referred to as RPF Label
   Stack Sub-TLV (Section 2.2).

   Note that the use of TEA to specify downstream and upstream
   forwarding information also apply to label assignment from the common
   SRGB or each router's SRLB, with the differences that the context
   label is not needed in the SRGB/SRLB case, and that in SRLB case only
   a Label Stack Sub-TLV with a single SRLB label is used for upstream
   and downstream forwarding information (no RPF Label Stack Sub-TLV is
   needed) in the SRLB case.

1.5.  Determining Root/Leaves

   For the controller to calculate a tree, it needs to determine the
   root and leaves of the tree.  This may be based on provisioning
   (static or dynamically programmed), or based on BGP signaling using
   the BGP multicast messages defined in
   [I-D.zzhang-bess-bgp-multicast], as described in the following two

   In both cases, the BGP updates are targeted at the controller, via an
   address specific Route Target with Global Administration Field set to
   the controller's address and the Local Administration Field set to 0.

Zhang, et al.             Expires May 21, 2020                  [Page 8]

Internet-Draft            bgp-mcast-controller             November 2019

1.5.1.  PIM-SSM/Bidir or mLDP P2MP

   In this case, the PIM Last Hop Routers (LHRs) with interested
   receivers or mLDP P2MP tunnel leaves encode a Leaf A-D route with the
   Upstream Router's IP Address field set to the controller's address
   and the Originating Router's IP Address set to the address of the LHR
   or the P2MP tunnel leaf.  The encoded PIM SSM source or mLDP FEC
   provides root information and the Originating Router's IP Address
   provides leaves information.

1.5.2.  PIM ASM

   In this case, the First Hop Routers (FHRs) originate Source Active
   routes which provides root information, and the LHRs originate Leaf
   A-D routes, encoded as in the PIM-SSM case except that it is (*,G)
   instead of (S,G).  The Leaf A-D routes provide leaf information.

2.  Specification

2.1.  Additional Tunnel Types for TEA

   This document specifies two new Tunnel Types.  The type codes will be
   assigned by IANA from the "BGP Tunnel Encapsulation Attribute Tunnel

2.1.1.  Any-Encapsulation Tunnel

   When a multicast packet needs to be sent from an upstream node to a
   downstream node, it may not matter how it is sent - natively when the
   two nodes are directly connected or tunneled otherwise.  In case of
   tunneling, it may not matter what kind of tunnel is used - MPLS, GRE,
   IPinIP, or whatever.

   To support this, an "Any-Encapsulation" tunnel type is defined.  This
   tunnel MUST have a Tunnel Endpoint Sub-TLV and SHOULD NOT have any
   other Sub-TLVs.  The Tunnel Endpoint Sub-TLV specifies an IP address,
   which could be any of the following:

   o  An interface's local address - when a packet needs to sent out of
      the corresponding interface natively.

   o  An interface's remote address - when a packet needs to sent to the
      address natively.

   o  An address that is not directly connected - when a packet needs to
      be tunneled to the address (any tunnel type/instance can be used).

Zhang, et al.             Expires May 21, 2020                  [Page 9]

Internet-Draft            bgp-mcast-controller             November 2019

2.1.2.  Load-balancing Tunnel

   Consider that a multicast packet needs to be sent to a downstream
   node, which could be reached via four paths P1~P4.  If it does not
   matter which of path is taken, an "Any-Encapsulation" tunnel with the
   Tunnel Endpoint Sub-TLV specifying the downstream node's loopback
   address works well.  If the controller wants to specify that only
   P1~P2 should be used, then a "Load-balancing" tunnel needs to be
   used, listing P1 and P2 as member tunnels of the "Load-balancing"

   A load-balancing tunnel has one "Member Tunnels" Sub-TLV defined in
   this document.  The Sub-TLV is a list of tunnels, each specifying a
   way to reach the downstream.  A packet will be sent out of one of the
   tunnels listed in the Member Tunnels Sub-TLV of the load-balancing

2.2.  RPF Label Stack Sub-TLV

   This is almost identical to Label Stack Sub-TLV.  The only difference
   is that the second inner most label in the stack identifies the
   expected upstream neighbor and explicit RPF checking needs to be set
   up for the tree label accordingly.

2.3.  Context Label Wide Community

   For a router to signal the context label that it assigns for a
   controller (or any label allocator that assigns labels that will be
   seen by this router), it attaches a Context Label Wide Community
   [I-D.ietf-idr-wide-bgp-communities] to the host route for its own
   address used in its BGP session towards the controllers (directly or
   via RRs).  This is a new wide community that specifies the (Label
   Allocator, Context Label) tuple, and the exact format will be
   specified in a future revision.

2.4.  Procedures

   Details to be added.  The general idea is described in the
   introduction section.

3.  Security Considerations

   This document does not introduce new security risks.

Zhang, et al.             Expires May 21, 2020                 [Page 10]

Internet-Draft            bgp-mcast-controller             November 2019

4.  IANA Considerations

   This document makes the following IANA requests:

   o  "Any-Encapsulation" and "Load-balancing" tunnel types from the
      "BGP Tunnel Encapsulation Attribute Tunnel Types" registry

   o  "Member Tunnels" and "RPF Label Stack" sub-TLV types from the "BGP
      Tunnel Encapsulation Attribute Sub-TLVs" registry


5.  Acknowledgements

   The authors Eric Rosen for his questions, suggestions, and help
   finding solutions to some issues like the neighbor based explicit RPF
   checking.  The authors also thank Lenny Giuliano, Sanoj Vivekanandan
   and IJsbrand Wijnands for their review and comments.

6.  References

6.1.  Normative References

              Patel, K., Velde, G., and S. Ramachandra, "The BGP Tunnel
              Encapsulation Attribute", draft-ietf-idr-tunnel-encaps-14
              (work in progress), September 2019.

              Raszuk, R., Haas, J., Lange, A., Decraene, B., Amante, S.,
              and P. Jakma, "BGP Community Container Attribute", draft-
              ietf-idr-wide-bgp-communities-05 (work in progress), July

              Zhang, Z., Giuliano, L., Patel, K., Wijnands, I., mishra,
              m., and A. Gulko, "BGP Based Multicast", draft-zzhang-
              bess-bgp-multicast-03 (work in progress), October 2019.

   [RFC2119]  Bradner, S., "Key words for use in RFCs to Indicate
              Requirement Levels", BCP 14, RFC 2119,
              DOI 10.17487/RFC2119, March 1997,

   [RFC8174]  Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
              2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
              May 2017, <>.

Zhang, et al.             Expires May 21, 2020                 [Page 11]

Internet-Draft            bgp-mcast-controller             November 2019

6.2.  Informative References

   [RFC6388]  Wijnands, IJ., Ed., Minei, I., Ed., Kompella, K., and B.
              Thomas, "Label Distribution Protocol Extensions for Point-
              to-Multipoint and Multipoint-to-Multipoint Label Switched
              Paths", RFC 6388, DOI 10.17487/RFC6388, November 2011,

   [RFC6513]  Rosen, E., Ed. and R. Aggarwal, Ed., "Multicast in MPLS/
              BGP IP VPNs", RFC 6513, DOI 10.17487/RFC6513, February
              2012, <>.

   [RFC7761]  Fenner, B., Handley, M., Holbrook, H., Kouvelas, I.,
              Parekh, R., Zhang, Z., and L. Zheng, "Protocol Independent
              Multicast - Sparse Mode (PIM-SM): Protocol Specification
              (Revised)", STD 83, RFC 7761, DOI 10.17487/RFC7761, March
              2016, <>.

   [RFC8402]  Filsfils, C., Ed., Previdi, S., Ed., Ginsberg, L.,
              Decraene, B., Litkowski, S., and R. Shakir, "Segment
              Routing Architecture", RFC 8402, DOI 10.17487/RFC8402,
              July 2018, <>.

Authors' Addresses

   Zhaohui Zhang
   Juniper Networks


   Robert Raszuk
   Bloomberg LP


   Dante Pacella


   Arkadiy Gulko
   Thomson Reuters


Zhang, et al.             Expires May 21, 2020                 [Page 12]