Internet DRAFT - draft-ravi-icnrg-ccn-notification

draft-ravi-icnrg-ccn-notification







ICN Research Group                                          R. Ravindran
Internet-Draft                                            A. Chakraborti
Intended status: Informational                                   S. Amin
Expires: January 17, 2018                            Huawei Technologies
                                                                 J. Chen
                                              Winlab, Rutgers University
                                                           July 16, 2017


                    Support for Notifications in CCN
                  draft-ravi-icnrg-ccn-notification-01

Abstract

   This draft proposes a new packet primitive called Notification for
   CCN.  Notification is a PUSH primitive and can be unicast or
   multicast to multiple listening points.  Notifications do not expect
   a Content Object response hence only requires the use of FIB state in
   the CCN forwarder.  Emulating Notification as a PULL has performance
   and routing implications.  The draft first discusses the design
   choices associated with using current Interest/Data abstraction for
   achieving push and challenges associated with them.  We follow this
   by proposing a new fixed header primitive called Notification and a
   CCN message encoding using Content Object primitive to transport
   Notifications.  This discussion are presented in the context of
   CCNx1.0 [1] proposal.  The draft also provides discussions on various
   aspects related to notification such as flow and congestion control,
   routing and reliability considerations, and use case scenarios.

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 http://datatracker.ietf.org/drafts/current/.

   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 January 17, 2018.






Ravindran, et al.       Expires January 17, 2018                [Page 1]

Internet-Draft            Interest Notification                July 2017


Copyright Notice

   Copyright (c) 2017 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
   (http://trustee.ietf.org/license-info) in effect on the date of
   publication of this document.  Please review these documents
   carefully, as they describe your rights and restrictions with respect
   to this document.  Code Components extracted from this document must
   include Simplified BSD License text as described in Section 4.e of
   the Trust Legal Provisions and are provided without warranty as
   described in the Simplified BSD License.

Table of Contents

   1.  Introduction  . . . . . . . . . . . . . . . . . . . . . . . .   2
   2.  Notification Requirements in CCN  . . . . . . . . . . . . . .   3
   3.  Using Interest/Data Abstraction for PUSH  . . . . . . . . . .   4
   4.  Proposed Notification Primitive in CCN  . . . . . . . . . . .   9
   5.  Notification Message Encoding . . . . . . . . . . . . . . . .  10
   6.  Notification Processing . . . . . . . . . . . . . . . . . . .  12
   7.  Security Considerations . . . . . . . . . . . . . . . . . . .  12
   8.  Annex . . . . . . . . . . . . . . . . . . . . . . . . . . . .  13
     8.1.  Flow and Congestion Control . . . . . . . . . . . . . . .  13
       8.1.1.  Issues with Basic Notifications . . . . . . . . . . .  13
       8.1.2.  Flow and Congestion Control Mechanims . . . . . . . .  14
         8.1.2.1.  End-to-End Approaches . . . . . . . . . . . . . .  14
         8.1.2.2.  Hybrid Approaches . . . . . . . . . . . . . . . .  15
       8.1.3.  Receiver Reliability  . . . . . . . . . . . . . . . .  17
     8.2.  Routing Notifications . . . . . . . . . . . . . . . . . .  18
     8.3.  Notification reliability  . . . . . . . . . . . . . . . .  18
     8.4.  Use Case Scenarios  . . . . . . . . . . . . . . . . . . .  19
       8.4.1.  Realizing PUB/SUB System  . . . . . . . . . . . . . .  19
   9.  Informative References  . . . . . . . . . . . . . . . . . . .  20
   Authors' Addresses  . . . . . . . . . . . . . . . . . . . . . . .  22

1.  Introduction

   Notification is a PUSH primitive used in the Internet today by many
   IoT and social applications.  The nature of notifications varies with
   the application scenario, ranging from being mission critical to one
   that is best effort.  Notifications can be unicast or multicast
   depending on whether the notification service is aware of all the
   consumers or not.  A notification service is preceded by a consumer
   subscribing to a specific event such as, subscription to hash-tag
   feeds, health emergency notification service, or temperature sensor



Ravindran, et al.       Expires January 17, 2018                [Page 2]

Internet-Draft            Interest Notification                July 2017


   reading from a room in a building; following this subscription the
   service pushes notifications to consuming entities.  It has to be
   noted that certain IoT applications expects notification end-to-end
   latency of few milliseconds [2].  Industrial IoT applications have
   more stringent requirement in terms of QoS, timeliness, and
   reliability of message delivery.  Though we term it as a
   Notification, this primitive can also be used for transactional
   exchange between two points.

   CCN optimizes networking around efficiently distributing already
   published content which the consumers learn through mechanisms like
   manifests containing the names of published content chunks and their
   locations.  Applications relying on notifications requires event
   driven data to be pushed from multiple producers to multiple
   subscribers for which the current Interest/Data primitive is
   inefficient.  This draft proposes to extend CCN's current primitives
   set with a new notification primitive that can be processed in a new
   way by the CCN forwarder to serve notification objectives.
   Notification here implies a PUSH semantic that is available with IP
   today and supported by other FIA architectures like MobilityFirst [3]
   and XIA [4].

2.  Notification Requirements in CCN

   General notification requirements and features have been discussed
   have been discussed in protocols such as CoAP's Observe proposal [5]
   to push notifications from the server to the clients.  Here we
   discuss basic notification requirements from CCN's network layer
   perspective.  Other requirements related to reliability, low latency,
   flow control can be engineered by the application or through more
   network layer state once the following requirements are met.

   o  Supporting PUSH Intent: CCN should provide efficient and scalable
      support for PUSH, where application's intent is to PUSH content to
      listening application without expecting any data in return.
      Efficiency relates to minimimizing control and forwarding overhead
      and scalability refers to support arbitrary number of producers
      and consumers participating in a general pub/sub or multicast
      service.

   o  Multicast Support: CCN network should be able to handle multicast
      notifications from a producer to multiple consumers.

   o  Security: Just as a content object in the context of Interest/Data
      primitive provides data authentication and privacy, similar
      features should also be offered by notification objects too.





Ravindran, et al.       Expires January 17, 2018                [Page 3]

Internet-Draft            Interest Notification                July 2017


   o  Routing/Forwarding Support: Name prefixes over which multicast
      notifications are managed should be handled in a different manner
      from the name prefixes over which Interest/Data primitive is used
      for content distribution in order to support the PUSH intent.
      This differentiation applies to the control as well as the
      forwarding plane.

   o  Minimizing Processing: Notification processing in the forwarder
      should be minimized considering the application's intent to PUSH
      data to listening consumers.

3.  Using Interest/Data Abstraction for PUSH

   Recent CCN and NDN research [6][7] have studied the problem of
   handling notifications and have proposed several solutions to handle
   this.  Here, we discuss several of them and point out their benefits
   and issues:

   Long-lived Interest v.1:  The most intuitive solution makes the
      assumption that the consumers know exactly the names of the
      contents that will be published in the future.  Yet, it is not
      easy since the providers can give arbitrary names to each piece of
      content, even though the contents might share a common prefix
      (i.e., GROUP_PREFIX).  To make it feasible, the providers can
      publish the contents with sequential ID, e.g., /GROUP_PREFIX/
      SENQUENTIAL_ID[/SEGMENT_ID], so that the consumers can query the
      contents with names /GROUP_ID/item_1, /GROUP_ID/item_2, ... (each
      name represents a content item).  The consumers can pipeline the
      requests (always keep some unsatisfied requests in flight, similar
      to TCP) to better utilize the network capacity.

      However, this solution has several issues, especially in the
      multi-provider scenario:

      *  Since it is unknown to the consumer (and the network) which
         provider will use which sequential ID, each request has to be
         forwarded to all the possible providers.  This solution might
         use up a large amount of state (PIT entries) in the network, as
         each consumer can keep tens of requests (to all providers) in
         flight for each group.

      *  Since each sequential ID should only be used by one provider,
         many PIT entries will not be consumed until timeout (if there
         is a timeout mechanism).  E.g., P1 and P2 are 2 providers of a
         group (/GROUP), the consumers have to send requests /GROUP/
         item_1, and /GROUP/item_2 to both providers.  Assume that P1
         publishes first so he uses the name /GROUP/item_1.  The PIT




Ravindran, et al.       Expires January 17, 2018                [Page 4]

Internet-Draft            Interest Notification                July 2017


         entries for /GROUP/item_1 towards P2 will not be consumed since
         P2 should now publish with name /GROUP/item_2.

      *  When the PIT entries form loops in the network (it can happen
         quite often in the multi-provider, multi-consumer scenario),
         the data packets can waste network traffic while following the
         loops and get discarded when redundancy happens.

      *  Other than the inefficiencies mentioned above, one major issue
         with this solution is the difficulty of provider
         synchronization.  It is not easy to make sure that different
         providers would use different sequential IDs especially when
         the providers are publishing contents at the same time.

   Polling v.1:  To eliminate the requirement for a sequential ID when
      publishing (to address the synchronization issue), the solution
      Polling v.1 makes the providers publish contents with name format:
      /GROUP_ID/TIMESTAMP.  While querying the contents, the consumer
      query using name /GROUP_ID/ with "exclude" field <Earliest version
      after Tx>, where Tx is the latest version the consumer has
      received.  E.g., after receiving a content with name /GROUP_ID/
      v_1234 (v_1234 is the timestamp of the publication time), the
      consumer would send a query with name /GROUP_ID/<Earliest after
      v_1234>.  He might get the next piece with name /GROUP_ID/v_2345
      (assuming that there is no content published between these two
      time stamps) without the need to know the exact names of the
      contents.  The content providers do not have to be synchronized on
      the sequential IDs and use the timestamp instead.

      While this solution is similar to the one used in NDN for getting
      the "latest" version under a prefix, it has several issues when we
      need to get "all" versions under a prefix:

      *  Ambiguity contents will appear when two providers of a same
         group publish at the same time.

      *  Consumers might miss messages when the clocks are not
         synchronized on the providers.  E.g., one provider (with faster
         clock) might publish a content with name /GROUP_ID/v_2345 after
         v_1234.  When the consumer queries for the earliest version
         after v_1234, he will get the content.  Yet, another provider
         (with slower clock) would publish a content with name
         /GROUP_ID/v_2234 after the consumer gets v_2345.  The consumer
         would miss the content with v_2234 as he will query for
         <Earliest after v_2345>.

      *  Consumers might miss messages due to different delivery latency
         (e.g., cache hit vs. no cache hit) even when the clocks on the



Ravindran, et al.       Expires January 17, 2018                [Page 5]

Internet-Draft            Interest Notification                July 2017


         providers are perfectly synchronized (e.g., via GPS signals).
         E.g., when a client queries for content /GROUP_ID/<Earliest
         after v_1234>, and there are two pieces of content exist in the
         network (v_2234, and v_2345).  It can happen that v_2345 is
         returned earlier (either due to a cache hit or because the
         provider is closer).  The consumer would then query for
         <Earliest after v_2345> and miss v_2234 with this solution.

      *  Also just as with the previous approach, this mechanism also
         requires the producers to sync so that they don't produce
         content using the same name.

   Long-lived Interest v.2:  To completely address the issues with
      multiple providers sharing a same prefix (e.g., synchronization in
      Long-lived Interest v.1, and clock synchronization in Polling
      v.1), Long-lived Interest v.2 gives a prefix to each provider.
      The providers in this solution provide contents with name
      /GROUP_ID/PROVIDER_ID/SEQUENTIAL_ID, and the consumers query the
      full names accordingly (similar to Long-lived Interest v.1 but
      with an extra prefix PROVIDER_ID).  The consumer can still use
      pipelining to improve the throughput.

      While this solution can avoid packet losses in the previous
      solution, it has several other issues:

      *  Consumers have to know all the potential providers, which might
         be difficult in some applications where every user can send
         messages in any group that he might be interested in.

      *  Compared to Long-lived Interest v.1, the consumers in this
         solution have to keep multiple pending queries per group per
         provider.  It might consume even more states in the network,
         which makes the solution less scalable.

      *  When a provider has more than one device (e.g., laptop and
         smartphone) that can publish contents under a same name
         /GROUP_ID/PROVIDER_ID, the solution would have the same
         synchronization issue as Long-lived Interest v.1.  If the
         solution mandates each device to have a separate provider ID,
         it will end up with even more PIT entries (states) in the
         network, and the solution becomes less "information-centric".

   Polling v.2:  To reduce the states and the control overhead in Long-
      lived Interest v.2, the solution Polling v.2 allows the provider
      process the requests in the application layer.  Periodically, the
      consumer would query each provider "if there is any update after
      Nx" (Nx is name of the last content the consumer has received).
      The query would be in the format: /GROUP_ID/PROVIDER_ID/Nx/NONCE.



Ravindran, et al.       Expires January 17, 2018                [Page 6]

Internet-Draft            Interest Notification                July 2017


      The provider would reply aggregated results in one response (with
      different segments, but under the same name), and an indication of
      "no update" if there is no publication after Nx.  Since a same
      query for /GROUP_ID/PROVIDER_ID/Nx can get different responses
      ("no update", or aggregated publications), a NONCE has to be added
      in the name to prevent possible cache hits in the network.  This
      solution can be effective in games since the publication rate
      (actions of the provider in the game) is much higher than the
      polling rate (refresh rate on the consumer).  However, it still
      has some issues (inefficiencies):

      *  There is a tradeoff between timeliness vs. in-network traffic
         when choosing the polling frequency.  The solution can be
         inefficient when the polling is too frequent: most of the
         polling will get "no update" responses.  This can consume a
         large amount of traffic in the network and extra computation on
         both the providers and the consumers.  The timeliness can be
         impaired when the polling is infrequent since the publication
         can only reach the consumer when the consumer queries.  The
         average delivery time of a publication in such solution is half
         of the polling period.

      *  In-network cache cannot be used since the response to a same
         query (without nonce) can be different according to the time
         (and maybe the consumer).

      *  Consumers still have to know all the potential providers
         similar to Long-lived Interest v.2.

   Polling with A Server:  To relieve the consumers from knowing all
      potential providers in Polling v.2, solution Polling with A Server
      introduces a server (or broker) as the delegate of all the
      providers.  The providers would publish data into the server and
      the consumers would poll for the updates from the server (similar
      to Twitter and Facebook in IP network).  In this solution, the
      consumers do not have to poll each provider for the updates, which
      reduces the overhead in the network.  With the aggregated response
      on the server, the network traffic is further reduced.  However,
      it still has several issues:

      *  Similar to all the server-based solutions like Facebook and
         Twitter, the server has to deal with all the polls.  This can
         cause single point of failure.

      *  It is not easy for the providers "publish contents to the
         server".  This becomes another notification problem and has to
         be solved by the other solutions mentioned in this section.




Ravindran, et al.       Expires January 17, 2018                [Page 7]

Internet-Draft            Interest Notification                July 2017


      *  Cache is not used in this solution similar to Polling v.2.

      *  This solution is not really "information-centric" as the
         consumers have to get the location of the content rather than
         the content itself.

   Interest Overloading:  Since all the aforementioned query/response
      solutions have issues with efficiency, scalability and/or
      timeliness, Interest Overloading tries to modify the communication
      pattern by using Interest packets to deliver publications
      directly.  The consumers in this solution propagate FIB entry of
      /GROUP_ID to all potential providers (or simply flood the
      network).  When a provider sends a publication, he would send an
      Interest with name /GROUP_ID/NONCE/<Payload> and the lifetime set
      to zero.  Since the traditional Interest packets do not have
      payload, the solution has to embed (e.g., URL encode [1]) the
      payload in the name of the Interest.  NONCE is used to prevent PIT
      aggregation since providers may publish contents with same payload
      (e.g., sensor readings).  This solution can address the timeliness
      and scalability issues with the Polling and Long-lived Interest
      solutions, yet there are still some issues:

      *  This solution creates ambiguity in the meaning of Interest
         packets (and the corresponding forwarding behaviors on the
         routers).  For a normal Interest packet, the forwarding engines
         should perform an anycast (send it to only one of the
         providers) according to FIB.  However, in this solution, the
         forwarding engines should use multicast logic for prefix
         /GROUP_ID (and avoid PIT storage).  Solution in [8] specifies
         some multicast prefixes so that the forwarding engines can
         distinguish the publications from the normal requests.  Yet,
         this places higher overhead on both the forwarding engines and
         the network management.  It also prevents providers to create
         contents under the /GROUP_ID prefix (since the query will be
         forwarded using multicast, and not kept in the PIT).

      *  The routing is also a concern in this solution.  When the
         consumers propagate FIB, it should reach all potential
         providers (in most of the time it will flood the network since
         all the users can be potential providers).  Naturally, in a
         multi-provider, multi-consumer scenario, the FIB entries would
         form a mesh in the network.  It is less scalable compared to
         the tree-based routing in IP multicast (PIM-SM).  The network
         has to specify another routing policy specifically for these
         prefixes, which places even higher overhead on network
         management.





Ravindran, et al.       Expires January 17, 2018                [Page 8]

Internet-Draft            Interest Notification                July 2017


      *  As is mentioned in [9], it is not efficient to embed large
         amount of data into the name of the Interest packets.  It adds
         more computation and storage overhead in the forwarding engines
         (PITs).

   Interest Trigger:  Similar to Interest Overloading, Interest Trigger
      uses an Interest packet as notification.  To eliminate the
      overhead of embedding the content in the Interest, this solution
      places the name of the publication in the name of the notification
      (Interest) packet.  On receiving the notification, the consumers
      can extract the content name and send another query (Interest) for
      the real content.  While this solution reduces the overhead of
      embedding the payload, it still has the ambiguity and routing
      issues similar to Interest Overloading solution.  It also incurs
      additional round trip delay before the produced data arrives at
      the listening consumer.

   To summarize CCN and NDN operates on PULL primitive optimized for
   content distribution applications.  Emulating PUSH operation over
   PULL has the following issues:

   o  It is a mismatch between an application's intent to PUSH data and
      the PULL APIs currently available.

   o  Unless Interests are marked distinctly, overloading Interests with
      notification data will undergo PIT/CS processing and are also
      subjected to similar routing and forwarding policies as regular
      Interests which is inefficient.

   o  Another concern in treating PUSH as PULL is with respect to the
      effect of local strategy layer routing policies, where the intent
      to experiment with multiple faces to fetch content is not required
      for notification messages.

   This motivates the need for treating notifications as a separate
   class of traffic which would allow a forwarder to apply the
   appropriate routing and forwarding processing in the network.

4.  Proposed Notification Primitive in CCN

   Notification is a new type of packet hence can be subjected to
   different processing logic by a forwarder.  By definition, a
   notification message is a PUSH primitive, hence is not subjected to
   PIT/CS processing.  This primitive can also be used by any other
   transactional or content distribution application towards service
   authentication or exchanging contextual information between end
   points and the service.




Ravindran, et al.       Expires January 17, 2018                [Page 9]

Internet-Draft            Interest Notification                July 2017


5.  Notification Message Encoding

   The wire packet format for a Notification is shown in Fig. 1 and Fig.
   2.  Fig. 1 shows the Notification fixed header considering the
   CCNx1.0 encoding, and Fig. 2 shows the format for the CCN
   Notification message, which is used to transport the notification
   data.  We next discuss these two packet segments of the Notification
   message.


                       1                   2                   3
   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
   +---------------+---------------+---------------+--------------+
   |    Version    |  PacketType=  |         PacketLength         |
   |               | Notification  |                              |
   +---------------+---------------+---------------+--------------+
   |   HopLimit    |   Reserved    |     Flags     | HeaderLength |
   +---------------+---------------+---------------+--------------+
   /                Optional Hop-by-hop header TLVs               /
   +---------------+---------------+---------------+--------------+
   /            Content Object as Notification Message            /
   +---------------+---------------+---------------+--------------+

                Figure 1: CCN Notification fixed header



                       1                   2                   3
   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
   +---------------+---------------+---------------+--------------+
   | MessageType = Content Object  |         MessageLength        |
   +---------------+---------------+---------------+--------------+
   |                            Name TLV                          |
   +---------------+---------------+---------------+--------------+
   |                    Optional MetaData TLVs                    |
   +---------------+---------------+---------------+--------------+
   |    Message Payload Type       |      Message Type Length     |
   +---------------+---------------+---------------+--------------+
   |              Payload or Optional Content Object              |
   +---------------+---------------+---------------+--------------+
   /             Optional CCNx ValidationAlgorithm TLV            /
   +---------------+---------------+---------------+--------------+
   / Optional CCNx ValidationPayload TLV (ValidationAlg required) /
   +---------------+---------------+---------------+--------------+

                  Figure 2: CCN Notification Message





Ravindran, et al.       Expires January 17, 2018               [Page 10]

Internet-Draft            Interest Notification                July 2017


   Notification Fixed Header: The fields in the fixed header that have
   new meaning in the context of notifications are discussed next, while
   the other fields follow the definition in [1].

   o  Packet Type: This new type code identifies that the packet is of
      type Notification [TBD].

   o  Optional Hop-by-hop header TLVs : Encodes any new hop-by-hop
      headers relevant to notifications [TBD].

   CCN Notification message: The CCN Notification message is a Content
   Object as in [1].  Notifications are always routed on the top level
   Content Object (outer CO) name.  Notification itself can be encoded
   in two forms depending on the application requirement:

   o  Notification with single name: In this case the notification
      contains a single content object.  Here the producer generates
      notification using the same name used by consumers on which they
      listen on.

   o  Notification with two names: In this case the notification
      contains a top level Content Object (outer CO), that encapsulates
      another Content Object (inner CO).  With an encapsulated Content
      Object, the meaning is that notification producers and consumers
      operate on different name-spaces requiring separate name-data
      security binding.  A good application of the encapsulation format
      is a PUB/SUB service, where the consumer learns about the
      notification service name offline, and the producer who is
      decoupled from the consumer generates a new Content Object using
      its own name and pushes the notification to the consumer.

   The interpretation of the fields shown in Fig. 2 are as follows:

   o  MessageType : The CCN message type is of type Content Object.

   o  Name TLV : Name TLV in the Content Object is used to route the
      Notification.

   o  Optional Metadata TLV: These TLVs carry metadata used to describe
      the Notification payload.

   o  Message Payload Type: This is of type T_PAYLOADTYPE defined in
      CCNx.1.0 or a new encapsulation type (T_ENCAP) that indicates the
      presence of another encapsulated Content Object [TBD].

   o  Optional Encapsulated Content Object: This is an optional
      encapsulated Content Object newly defined for the Notification
      primitive.  The name in the encapsulated Content Object



Ravindran, et al.       Expires January 17, 2018               [Page 11]

Internet-Draft            Interest Notification                July 2017


      corresponds to the producer's name-space, or anything else based
      on the application logic.  The rational for an encapsulated
      Content Object was discussed earlier.

   o  Optional Security Validation data: The Content Object optionally
      carries security validation payload as per CCNx1.0.

6.  Notification Processing

   The following steps are followed by a CCN forwarder to process the
   Notification packet.

   o  Notification packet type is identified in the fixed header of a
      CCN packet with a new type code.  The Notification carries a
      Content Object, whose name is used for routing.  This name is
      matched against the FIB entries to determine the next hop(s).
      Novel strategy layer routing techniques catering to the
      notification traffic can be applied here.

   o  CCN forwarder also processes the optional metadata associated with
      the Notification meant for the network to help with the forwarding
      strategy, for e.g., mission critical notifications can be given
      priority over all other traffic.

   o  As mentioned earlier, CCN forwarder MUST NOT cache the Content
      Objects in the notifications.

7.  Security Considerations

   The proposed processing logic of Notifications that bypass the
   processing of PIT/CS has the following security implications:

   Flow Balance : PIT state maintains the per-hop flow balance over all
   the available faces by enforcing a simple rule, that is, one Content
   Object is send over a face for a single Interest.  Bypassing PIT
   processing compromises this flow balancing property.  For scenarios
   where the notification traffic volume is not high such as for IoT
   applications, the impact may not be significant.  However, this may
   not be the case considering the plethora of social networking and
   emerging IoT applications in a general Internet scenario.  This flow
   balance tradeoff has to be understood considering an application's
   intent to PUSH data and the latency introduced by processing such
   traffic if a PULL primitive is used.  Also PIT offers a natural
   defense mechanism by throttling traffic at the network edge,
   considering the provisioned PIT size, and bypassing it could
   exacerbate DDOS attacks on producing end points.





Ravindran, et al.       Expires January 17, 2018               [Page 12]

Internet-Draft            Interest Notification                July 2017


   Cache Poisoning: This draft doesn't recommend the caching of the
   Content Object in the Notification payload, though doing so might
   help in increasing the availability of notification information in
   the network.  A possible exception would be if the inner CO is a
   nameless object [10]. as those can only be fetched from CS by hash We
   leave this possibility of applying policy-based caching of
   Notification Content Objects for future exploration.  The
   recommendation for not caching these Content objects is that, in a
   regular Interest/Content Object exchange, content arrives at the
   forwarder and is cached as a result of per-hop active Interest
   expression.  Unsolicited Content Objects, as in the case of the
   Notification, violates this rule, which could be exploited by
   malicious producers to generate DDOS attack against the cache
   resource of a CCN infrastructure.

8.  Annex

8.1.  Flow and Congestion Control

8.1.1.  Issues with Basic Notifications

   As mentioned in the previous sections, one of the main issues with
   notification is the flow and congestion control.  One naive way to
   solve this issue is the routers drop the packets from aggressive
   flows.  Flow-based fair queueing (and its variation stochastic
   fairness queueing) maintain queues for flows (or the hash of flows)
   and try to give a fair share to each flow (or a hash).  Flows can be
   classified by the prefixes in the ICN case.  However, according to
   [11], the overall network throughput will be affected when there are
   multiple bottlenecks in the network.  Therefore, [11] promotes an
   end-to-end solution for congestion control.  Flow balance is a key
   requirement to an end-to-end (or end-driven) flow and congestion
   control.  In the case of CCN query/response, flow balance entails
   that an Interest pulls at most one Data object from upstream.  The
   data consumer can therefore control the amount of traffic coming from
   the data source(s) either it is a data provider or a cache in the
   network.  However, the basic notification does not follow the rule of
   flow balance (each Subscription can result in more than one
   Notifications disseminated in the network).  In the absence of a
   proper feedback mechanism to notify the data sender or the network
   the available bandwidth and local resource the consumer has, the
   sender can easily congest the bottleneck link of the receivers
   (causing congestion collapse) and/or overflow the buffer on the
   receiver side.  In the later sections, we will describe the possible
   congestion control mechanisms in ICN and how to deal with packet loss
   when both congestion control and reliability are required.





Ravindran, et al.       Expires January 17, 2018               [Page 13]

Internet-Draft            Interest Notification                July 2017


   However, the basic notification does not follow the rule of flow
   balance (each Subscription can result in more than one Notifications
   disseminated in the network).  There is no way a receiver can notify
   the data sender or the network the available bandwidth and local
   resource it has.  As a result, the sender can easily congest the
   bottleneck link of the receivers (causing congestion collapse) and/or
   overflow the buffer on the receiver side.

8.1.2.  Flow and Congestion Control Mechanims

   Here we discuss broad approaches towards achieving flow and
   congestion control in CCN as applied to Notification traffic.  Since
   the forwarding logic of the Notification packets are quite similar to
   that of IP multicast, existing multicast congestion control solutions
   can be candidates to solve the flow/congestion control issue with
   Notification.  In addition we also summarize recent ICN research to
   address this issue.

8.1.2.1.  End-to-End Approaches

   In the multicast communication, it is not scalable to have direct
   receiver-to-sender feedback loop similar to TCP since this would
   result in each receiver sending ACKs (or NACKs) to the data sender
   and cause ACK (NACK) implosion.  To address the ACK implosion issue,
   two types of solutions have been proposed in multicast congestion
   control, namely, sender-driven approaches and receiver-driven
   approaches.

8.1.2.1.1.  Sender-driven Multicast

   In the first category, the sender controls the sending rate and to
   ensure the network friendliness, the sender usually align the sending
   rate to the slowest receiver.

   To avoid the ACK implosion issue, TCP-Friendly Multicast Congestion
   Control (TFMCC [12]) uses rate based solution.  This solution uses
   TCP-Friendly Rate Control (TFRC) to get a proper sending rate based
   on the RTT between sender and each receiver.  The sender only needs
   to collect the RTTs periodically instead of per-packet ACKs.
   Similarly, in ICN, the sender can create another channel (namespace)
   to collect the RTT measurement from the receivers.  However, due to
   the dynamics on each path, it is difficult to calculate the proper
   sending rate.

   To address the rate calculation issue, pgmcc [13], a window-based
   solution is proposed.  It uses NACKs to detect the slowest receiver
   (the ACKer).  The ACKer sends an ACK back to the sender on receiving
   each multicast packet.  A feedback loop similar to TCP is formed



Ravindran, et al.       Expires January 17, 2018               [Page 14]

Internet-Draft            Interest Notification                July 2017


   between the sender and the ACKer to control the sending rate.  Since
   the ACKer is the slowest receiver, the sender adapts its sending rate
   to the available bandwidth of the slowest receiver, the solution can
   therefore ensure the network friendliness.  In the ICN case, the
   receivers can send NACKs in the form of Notification packets through
   another namespace, and the ACKer can also use the same mechanism to
   send ACKs.

   However, since the sender is always aligning the sending rate to the
   slowest receiver to ensure the network friendliness, the performance
   of the solutions can be dramatically affected by a very slow
   receiver.

8.1.2.1.2.  Receiver-driven Multicast

   Unlike the sender-driven solutions, the receiver-driven solutions
   [14] choose to use layered-multicast to satisfy heterogeneous
   receivers.  The sender first initiates several multicast groups
   (namespaces in the case of ICN) with different sending rates.  Each
   receiver would choose to join a multicast group with the highest
   sending rate that it can afford.  The sender can also adapt the
   sending rate of each multicast group according to the receiver
   status.

   These solutions can support applications like video streaming (with
   layered codecs) efficiently.  However, they also have some issues: 1)
   they complicate the sender and receiver logic, especially for simple
   applications like file transfer; and 2) the receivers are limited by
   the sending rates initiated by the provider and would therefore
   under-utilize the available bandwidth.

8.1.2.2.  Hybrid Approaches

   In this approach, flow balance of Notification is achieved by the
   receivers notifying the network (rather than the sender or other
   receivers) about the capacity it can receive.  Here, we take
   advantage of operating the Notification service through a receiver-
   driven approach and get support from the network.

   A solution based on this approach is proposed in [15], which we
   summarize next.

   To retain flow balance, the consumers in this solution send out one
   subscription for only one next Notification instead of the original
   logic (that receives all the Notifications).  Similar to the flow and
   congestion control in query/response, the receivers can now maintain
   a congestion window to control the amount of traffic coming from
   upstream.



Ravindran, et al.       Expires January 17, 2018               [Page 15]

Internet-Draft            Interest Notification                July 2017


   Here, instead of maintaining a (name, outgoing face) pair in FIB (or
   subscription table), the routers now adds a third field --
   accumulated count -- for each entry.  The accumulated count is
   increased by 1 on receiving such a subscription and decreased by 1 on
   sending a Notification to that face.  The routers should also
   propagate the maximum accumulated count upstream till the 1st hop
   router of the provider (or the rendezvous point in the network).  The
   subscribers sends a subscription for every successfully received
   notification.  Here we also assume that, the subscribers operate
   based on the AIMD scheme.

   If the dissemination of Notification follows a tree topology in the
   network, we define the branching point of a receiver R (BP_R) as the
   router closest to R which has another outgoing face that can receive
   data faster than R.  For receivers that has bandwidth/resources to
   receive all the data from the provider, BP_R is the 1st hop router of
   the provider (or the rendezvous point).

   In this solution, we can prove that there is a feedback loop between
   each receiver and its branching point.  Therefore, when a receiver
   maintains its congestion window size using AIMD, the traffic between
   the branching point and the receiver is similar to TCP.  It can get a
   fair share at the bottleneck on the path, even if the bottleneck is
   not directly under the branching point.  In the multicast tree, the
   solution can ensure the fairness with other (TCP-like) flows on each
   branch.

   The solution can thus allow the sender to send at an application-
   efficient rate rather than being affected by the slowest receiver
   like pgmcc [13].

   It is true that the solution requires more packets and more states in
   the network compared to the basic notification solution, but the cost
   is similar to (and smaller than) that of query/response.  Since we
   are using one notification per subscription pattern, the amount of
   traffic overhead is the same as query/response.  As for the states
   stored in the router, the solution only requires 1 entry per prefix
   per face, which is smaller than the query/response which requires 1
   entry per packet per face.  Therefore, the overhead of the solution
   is acceptable in CCN.

8.1.2.2.1.  Other Challenges

   o  Sender Rate Control: The sender in the solution does not have to
      limit the sending rate to the slowest receiver to maintain network
      friendliness.  Therefore, the choice of sending rate is a tradeoff
      between network traffic and session completion time.  In the case
      where the application does not require a certain sending rate



Ravindran, et al.       Expires January 17, 2018               [Page 16]

Internet-Draft            Interest Notification                July 2017


      (like file transfer), the sender can align the sending rate to the
      slowest receiver (similar to pgmcc) to minimize the repair
      traffic, but at the cost of longer session completion time.  He
      can also send at the rate of the fastest receiver and try to get
      peer repair in the network.  This allows faster receivers finish
      the session earlier but causing higher network traffic due to the
      repair.  An ACKer-based solution similar to pgmcc can be adopted
      to allow the sender align the rate at a proportion of users (e.g.,
      top 30%).  The sender can collect feedback (throughput, latency,
      etc.) from all the receivers periodically and pick an ACKer
      according to the proportion it desires.  On receiving a
      Notification packet, the ACKer would send an ACK just like TCP.
      The sender can maintain a congestion window also like TCP.  The
      feedback loop between the sender and the ACKer can align the
      sending rate at the ACKers's available bandwidth.

   o  Receiver Window Control: Slightly different from one-sender one-
      receiver window control in TCP, the sending rate in the hybrid
      approach is not controlled by any of the receivers.  Receiving
      intermittent packets can indicate both congestion (similar to TCP)
      and not enough window size (since the sending rate is higher).  In
      the first case, the receiver should reduce the window size while
      in the second case, the receiver should increase the window size.
      An indication of congestion (e.g., Random Early Detection, RED)
      should be provided directly from the network.The receivers with
      available bandwidth higher than the sending rate would have too
      large window size since it does not see any packet loss.  Please
      refer to [15] for a detailed solution on this issue.

8.1.3.  Receiver Reliability

   The receiver would miss packets when the available bandwidth/resource
   of the receiver is lower than the sending rate of the Notification
   provider.  Some applications (like gaming and video conferencing) can
   tolerant such kind of packet loss while the others (like file
   transfer) cannot.  Therefore, another module that ensures the
   reliability is needed.  However, reliability should be separated from
   the flow and congestion control since it is not a universal
   requirement.

   With the solution described in the receiver-driver or the hybrid
   approach, the slower consumers would receive intermittent packets
   since the sending rate can be faster than their fair share.  The
   applications that require reliable transfer can query the missing
   packets similar to the normal query/response.  This also requires
   that each content in the Notifications should have a unique Content
   Name (or hash in the nameless scenario).  The clients should also be
   able to detect the missing packets either based on the sequence



Ravindran, et al.       Expires January 17, 2018               [Page 17]

Internet-Draft            Interest Notification                July 2017


   number or based on a pre-acquired meta-file.  Caching in CCN can be
   leveraged to achieve availability and reliability.

   The network can forward the requests (Interests) of the missing
   packets towards the data provider, the other consumers and/or the in-
   network cache to optimize the overall throughput of the consumers.
   This solution is similar to Scalable Reliable Multicast (SRM [16]).
   However, as mentioned in [17], solutions like SRM requires the
   consumers communicate directly with each other and therefore lose the
   privacy and trust.  CCN can ensure the privacy since the providers
   cannot get the information of the identity of the consumers.  Trust
   (data integrity) is also maintained with the signature in the Data
   packets.

8.2.  Routing Notifications

   Appropriate routing policies should be employed to ensure reliable
   forwarding of a notification to its one or many intended receivers.
   The name in the notification identifies a host or a multicast service
   being listened to by the multiple intended receivers.  Two types of
   routing strategies can be adopted to handle notifications, depending
   on whether or not an explicit pub/sub state is maintained in the
   forwarder.

   o  Stateless forwarding: In this case the notification only relies on
      the CCN FIB state to route the notification.  The FIB entries are
      populated through a routing control plane, which distinguishes the
      FIB states for the notification service from the content fetching
      FIB entries.  Through this logical separation, Notifications can
      be routed by matching its name with the matching FIB policy in the
      CCN forwarder, hence processed as notification multicast.

   o  Stateful forwarding: In this case, specific subscription state is
      managed in the forwarder to aid notification delivery.  This is
      required to scale notifications at the same time apply
      notification policies, such as filter notifications or to improve
      notification reliability and efficiency to subscribing users [18].

8.3.  Notification reliability

   This proposal doesn't provide any form of reliability.  Reliability
   can be realized by the specific application using the proposed
   notification primitive, for instance using the following potential
   approaches:

   Caching: This proposal doesn't propose any form of caching.  But
   caching feature can be explored to improve notification reliability,
   and this is a subject of future study.  For instance, consumers,



Ravindran, et al.       Expires January 17, 2018               [Page 18]

Internet-Draft            Interest Notification                July 2017


   which expect notifications and use external means (such as periodic
   updates or by receiving manifests) to track notifications, can
   recover the lost notifications using the PULL feature of CCN.

   Notification Acknowledgment: If the producer maintains per-receiver
   state, then the consumer can send back notification ACK or NACK to
   the producer of having received or not received them.

8.4.  Use Case Scenarios

   Here we provide the discussions related to the use of Notification in
   different scenarios.

8.4.1.  Realizing PUB/SUB System

   A PUB/SUB system provides a service infrastructure for subscribers to
   request update on a set of topics of interest, and with multicast
   publishers publishing content on those topics.  A PUB/SUB system maps
   the subscribers' interests to published contents and pushes them as
   Notifications to the subscribers.  A PUB/SUB system has many
   requirements as discussed in [19] which include low latency,
   reliability, fast recovery, scalability, security, minimizing false
   (positive/negative) notifications.

   Current IP based PUB/SUB systems suffer from interoperability
   challenges because of application-defined naming approach and lack of
   support of multicast in the data plane.  The proposed Notification
   primitive can be used to realize large scale PUB/SUB system, as it
   unifies naming in the network layer and support for name-based
   multicasting.

   Depending on the routing strategy discussed earlier, two kind of PUB/
   SUB approaches can be realized : 1) Rendezvous style approach ; 2)
   Distributed approach.  Each of these approaches can use the
   Notification primitive to implement their PUSH service.

   In the Rendezvous style approach, a logically centralized service
   maps subscriber's topic interest with the publisher's content and
   pushes it as notifications.  If stateless forwarding is used, the
   routing entries contain specific application-ID's requesting a given
   notification, to handle scalability, a group of these application can
   share a multicast-ID reducing the state in the FIB.

   In the Distributed approach, the CCN/NDN protocol is further enhanced
   with new subscription primitive for the subscription interested
   consumers.  When a consumer explicitly susbcribes to a multicast
   topic, its subscription request is forwarded to the upstream
   forwarder which manages this state mapping between subscription names



Ravindran, et al.       Expires January 17, 2018               [Page 19]

Internet-Draft            Interest Notification                July 2017


   to the downstream faces which has expressed interest for
   Notifications being pushed under that prefix.  An example of the
   network layer based approach is the COPSS notification proposal [19].
   Here a PUB/SUB multi-cast state state, called the subscribers
   interest table, is managed in the forwarders.  When a Notification
   arrives at a forwarder, the content descriptor in the notification is
   matched to the PUB/SUB state in the forwarder to decide the faces
   over which the Notification has to be forwarded.

9.  Informative References

   [1]        CCN Wire format, CCNX1., "http://www.ietf.org/id/
              draft-mosko-icnrg-ccnxmessages-00.txt.", 2013.

   [2]        Osseiran, A., "Scenarios for 5G Mobile and Wireless
              Communications: The Vision of the METIS Project.", IEEE
              Communication Magazine , 2014.

   [3]        NSF FIA project, MobilityFirst.,
              "http://www.nets-fia.net/", 2010.

   [4]        NSF FIA project, XIA., "https://www.cs.cmu.edu/~xia/",
              2010.

   [5]        Observing Resources in CoAp, observe.,
              "https://tools.ietf.org/html/draft-ietf-core-observe-16.",
              2015.

   [6]        Amadeo, M., Campolo, C., and A. Molinaro, "Internet of
              Things via Named Data Networking: The Support of Push
              Traffic", Network of the Future (NOF), 2014 International
              Conference and Workshop on the , 2014.

   [7]        Shang, W., Bannis, A., Liang, T., and Z. Wang, "Named Data
              Networking of Things.", IEEE IoTDI 2016, 2016.

   [8]        Zhu, Z. and A. Afanasyev, "Let's chronosync: Decentralized
              dataset state synchronization in named data networking",
              The 21st IEEE International Conference on Network
              Protocols ICNP, 2013.

   [9]        Moiseenko, I. and O. Oran, "TCP/ICN: Carrying TCP over
              Content Centric and Named Data Networks", Proceedings of
              the 3rd ACM Conference on Information-Centric
              Networking ICN, 2016.

   [10]       Mosko, M., "Nameless Objects.", IETF/ICNRG, Paris
              Interim 2016, 2016.



Ravindran, et al.       Expires January 17, 2018               [Page 20]

Internet-Draft            Interest Notification                July 2017


   [11]       Floyd, S. and F. Kevin, "Promoting The Use of End-to-End
              Congestion Control in The Internet.", IEEE ToN vol. 7(4),
              pp. 458-472, 1999.

   [12]       Widmer, J. and M. Handley, "TCP-Friendly Multicast
              Congestion Control (TFMCC): Protocol Specification.", IETF
              RFC 4654, 2006.

   [13]       Rizzo, L., "pgmcc: A TCP-Friendly Single-Rate Multicast
              Congestion Control Scheme.", SIGCOMM CCR vol. 30.4, pp.
              17-28, 2000, 2000.

   [14]       McCanne, S., Jacobson, V., and M. Vetterli, "Receiver-
              driven Layered Multicast.", SIGCOMM CCR pp. 117-130, 1996.

   [15]       Chen, J., Arumaithurai, M., Fu, X., and KK. Ramakrishnan,
              "SAID: A Control Protocol for Scalable and Adaptive
              Information Dissemination in ICN.", arXiv vol. 1510.08530,
              2015.

   [16]       Floyd, S., Jacobson, V., Liu, C., McCanne, S., and L.
              Zhang, "A Reliable Multicast Framework for Light-Weight
              Sessions and Application Level Framing.", IEEE TON vol.
              5(6), pp. 784-803, 1997.

   [17]       Floyd, N., Grossglauser, M., and KK. Ramakrishnan,
              "Distrust and Privacy: Axioms for Multicast Congestion
              Control.", Distrust and Privacy: Axioms for Multicast
              Congestion Control NOSSDAV, 1999.

   [18]       Francois et al, J., "CCN Traffic Optimization for IoT",
              Proc. of NoF , 2013.

   [19]       Chen, J., Arumaithurai, M., Jiao, L., Fu, X., and K.
              Ramakrishnan, "COPSS: An Efficient Content Oriented
              Publish/Subscribe System.", ACM/IEEE Symposium on
              Architectures for Networking and Communications Systems
              (ANCS 2011) , 2011.

   [20]       DNS Security Introduction and Requirements, DNS-SEC.,
              "http://www.ietf.org/rfc/rfc4033.txt.", 2005.

   [21]       Cisco System Inc., CISCO., "Cisco visual networking index:
              Global mobile data traffic forecast update.", 2009-2014.

   [22]       CCNx Label Forwarding, CCNLF., "http://www.ccnx.org/pubs/
              ccnx-mosko-labelforwarding-01.txt.", 2013.




Ravindran, et al.       Expires January 17, 2018               [Page 21]

Internet-Draft            Interest Notification                July 2017


Authors' Addresses

   Ravishankar Ravindran
   Huawei Technologies
   2330 Central Expressway
   Santa Clara, CA  95050
   USA

   Email: ravi.ravindran@huawei.com


   Asit Chakraborti
   Huawei Technologies
   2330 Central Expressway
   Santa Clara, CA  95050
   USA

   Email: asit.chakraborti@huawei.com


   Syed Obaid Amin
   Huawei Technologies
   2330 Central Expressway
   Santa Clara, CA  95050
   USA

   Email: obaid.amin@huawei.com


   Jiachen Chen
   Winlab, Rutgers University
   671, U.S 1
   North Brunswick, NJ  08902
   USA

   Email: jiachen@winlab.rutgers.edu















Ravindran, et al.       Expires January 17, 2018               [Page 22]