QUIC                                                              Y. Liu
Internet-Draft                                                     Y. Ma
Intended status: Standards Track                            Alibaba Inc.
Expires: 9 September 2021                                     C. Huitema
                                                    Private Octopus Inc.
                                                                   Q. An
                                                            Alibaba Inc.
                                                                   Z. Li
                                                            8 March 2021

                      Multipath Extension for QUIC


   This document specifies multipath extension for the QUIC protocol to
   enable the simultaneous usage of multiple paths for a single
   connection.  The extension is compliant with the single-path QUIC
   design.  The design principle is to support multipath by adding
   limited extension to [QUIC-TRANSPORT].

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 https://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 9 September 2021.

Copyright Notice

   Copyright (c) 2021 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 (https://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.  Conventions and Definitions
   3.  Enable Multipath QUIC - Handshake
   4.  Path Management
     4.1.  Path Identifier and Connection ID
     4.2.  Path Packet Number Spaces
     4.3.  Path Initiation
     4.4.  Path State Management
     4.5.  Path Close
       4.5.1.  Use PATH_STATUS frame to close a path
       4.5.2.  Effect of RETIRE_CONNECTION_ID frame
       4.5.3.  Idle timeout
   5.  Using TLS to Secure QUIC Multipath
     5.1.  Packet protection for QUIC Multipath
     5.2.  Key Update for QUIC Multipath
   6.  Using Multipath QUIC with load balancers
   7.  Packet scheduling
     7.1.  Basic Scheduling
     7.2.  Scheduling with QoE Feedback
     7.3.  Per-stream Policy
   8.  Congestion control and loss detection
     8.1.  Congestion control
     8.2.  Packet number space and acknowledgements
     8.3.  Flow control
   9.  New frames
     9.1.  PATH_STATUS frame
     9.2.  ACK_MP frame
     9.3.  QOE_CONTROL_SIGNALS frame
   10. Implementation Considerations
     10.1.  Handling of 0-RTT packets
   11. Security Considerations
   12. IANA Considerations
   13. Changelog
   14. Appendix.A Scenarios related to migration
   15. Appendix.B Considerations on RTT estimate and loss detection
   16. Appendix.C Difference from past proposals
   17. References
     17.1.  Normative References
     17.2.  Informative References
   Authors' Addresses

1.  Introduction

   In this document, we propose an extension to the current QUIC design
   to enable the simultaneous usage of multiple paths for a single

   This proposal is based on several basic design points:

   *  Re-use as much as possible mechanisms of QUIC-v1, which has
      supported connection migration and path validation.

   *  To avoid the risk of packets being dropped by middleboxes (which
      may only support QUIC-v1), use the same packet header formats as
      QUIC V1.

   *  Endpoints need a Path Identifier for each different path which is
      used to track states of packets.  As we want to keep the packet
      header formats unchanged [QUIC-TRANSPORT], Connection IDs (and the
      sequence number of Connection IDs) would be a good choice of Path

   *  For the convenience of packet loss detection and recovery,
      endpoints use a different packet number space for each Path

   *  Congestion Control, RTT measurements and PMTU discovery should be
      per-path (following [QUIC-TRANSPORT])

   This document is organized as follows.  It first provides definitions
   of multipath quic in Section 2.  It then specifies how to enable
   multipath quic during handshake in Section 3, and path management in
   Section 4.  It discusses packet scheduling in Section 7, and
   congestion control in Section 8.  The new frames are defined in
   Section 9.

2.  Conventions and Definitions

   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.

   We assume that the reader is familiar with the terminology used in
   [QUIC-TRANSPORT].  In addition, we define the following terms:

   *  Path Identifier: An identifier that is used to identify a path in
      a QUIC connection at an endpoint.  It is defined as the sequence
      number of the destination Connection ID used for sending packets
      on that particular path.

   *  Each node maintains a list of "Received Packets" for each of the
      CID that it provided to the peer, which is used for acknowledging
      packets received with that CID.

3.  Enable Multipath QUIC - Handshake

   This extension defines a new transport parameter, used to negotiate
   the use of the multipath extension during the connection handshake,
   as specified in [QUIC-TRANSPORT].  The new transport parameter is
   defined as follow:

   *  name: enable_multipath (TBD - experiments use 0xbaba)

   *  value: 0 (default) for disabled, 1 for enabled

   If the peer does not carry the enable_multipath(TBD - experiments use
   0xbaba) transport parameter, which means the peer does NOT support
   multipath, endpoint MUST fallback to [QUIC-TRANSPORT] with single
   path and MUST NOT send any MP frames in the following packets, also
   MUST NOT use the multipath specific AEAD algorithm defined in
   Section 5.1.

   Notice that transport parameter "active_connection_id_limit"
   [QUIC-TRANSPORT] limits the number of usable Connection IDs, and also
   limits the number of concurrent paths.

4.  Path Management

   After endpoints have negotiated in handshake flow that both endpoints
   enable multipath feature, endpoints can start using multiple paths.

   This proposal add one frame for path management:

   *  PATH_STATUS frame for the receiver side to claim the path state
      and preference

   All the new MP frames are sent in 1-RTT packets [QUIC-TRANSPORT].

4.1.  Path Identifier and Connection ID

   Endpoints need a Path Identifier for each different path which is
   used to track states of packets.  Endpoints use Connection IDs in
   1-RTT packet header as Path Identifier in each directions, and use
   the sequence number of Connection IDs in MP frames to identify the
   path referred.

   Following [QUIC-TRANSPORT], Each endpoint uses NEW_CONNECTION_ID
   frames to claim usable connections IDs for itself.  Before an
   endpoint add a new path, it SHOULD check whether there is at least
   one unused available Connection ID for each side.

   Endpoints can find which path a received packet belongs to according
   to the Destination Connection ID of the 1-RTT packet.  Endpoints can
   find the context of a path by its' Connection ID or the Sequence
   number of Connection ID.

4.2.  Path Packet Number Spaces

   For the convenience of packet loss detection and recovery, endpoints
   use a different packet number space for each Path Identifier
   (Connection ID).  ACK_MP frame includes the sequence number of the
   Destination Connection ID of the acknowledged packets as the Path

4.3.  Path Initiation

   Figure 1 illustrates an example of new path establishment.

      Client                                                  Server

      (Exchanges start on default path)
      1-RTT[]: NEW_CONNECTION_ID[C1, Seq=1] -->
                            <-- 1-RTT[]: NEW_CONNECTION_ID[S1, Seq=1]
                            <-- 1-RTT[]: NEW_CONNECTION_ID[S2, Seq=2]
      (starts new path)
      1-RTT[0]: DCID=S2, PATH_CHALLENGE[X] -->
                        Checks AEAD using nonce(CID sequence 2, PN 0)
      Checks AEAD using nonce(CID sequence 1, PN 0)
      1-RTT[1]: DCID=S2, PATH_RESPONSE[Y],
                ACK_MP[Seq=1, PN=0], ... -->

                Figure 1: Example of new path establishment

   As shown in Figure 1, client provides one unused available Connection
   ID (C1 with sequence number 1), and server provides two available
   Connection IDs (S1 with sequence number 1, and S2 with sequence
   number 2).  When client wants to start a new path, it checks whether
   there is unused available Connection IDs for each side, and choose an
   available Connection ID S2 as the Destination Connection ID in the
   new path.

   Endpoints need to exchange unused available Connection IDs with the
   NEW_CONNECTION_ID frame before an endpoint starts a new path.  For
   example, if the goal is to maintain 2 paths, each endpoint should
   provide at least 3 CID to its peer: 2 in use, and one spare.  If the
   client has used all the allocated CID, it is supposed to retire those
   that are not used anymore, and the server is supposed to provide
   replacements, as specified in [QUIC-TRANSPORT].

   If the transport parameter "active_connection_id_limit" is negotiated
   as N, and the server has provided N Connection IDs and the client has
   started N paths, the limit is reached.  If the client wants to start
   a new path, it has to retire one of the established paths.

   Path validation uses the PATH_CHALLENGE and PATH_RESPONSE frame
   defined in QUIC-Transport [QUIC-TRANSPORT].

4.4.  Path State Management

   An endpoint uses PATH_STATUS frames to inform that the peer should
   send packets in the preference expressed by these frames.  An
   endpoint uses the sequence number of the CID used by the peer for
   PATH_STATUS frames (describing the sender's path identifier).

   In the example Figure 1, if the client wants to send a PATH_STATUS
   frame to tell the server that it prefers the path with CID sequence
   number 1 (of the server's side), the client should use the identifier
   of the server (sequence 1) in PATH_STATUS frame.

   PATH_STATUS frame describes 4 kinds of path states:

   *  Abandon a path, and release the corresponding resource.

   *  Mark a path as "available", i.e., allow the peer to use its own
      logic to split traffic among available paths.

   *  Mark a path as "standby", i.e., suggest that no traffic should be
      sent on that path if another path is available.

   *  Mark the priority of a path, i.e, path 1 is weight 8, path 2 is
      weight 2, suggest that path 1 has higher priority than path 2, and
      peer should try to send more data in path 1.

   PATH_STATUS frame can be sent via a different path, instead of the
   path identified by the Path Identifier field.

4.5.  Path Close

   An endpoint that want to delete a path SHOULD NOT rely on implicit
   signals like idle time or packet losses, but instead SHOULD use
   explicit ask to abandon path by sending the PATH_STATUS frame.

4.5.1.  Use PATH_STATUS frame to close a path

   Both client and server can close a path, by sending PATH_STATUS frame
   which abandons the path with a corresponding Path Identifier.  Once a
   path is marked as "abandon", it means that the resources related to
   the path can be released.

   Figure 2 illustrates an example of path closing.  In this case, we
   are going to close the first path.  For the first path, the server's
   1-RTT packets use DCID C1, which has a sequence number of 1; the
   client's 1-RTT packets use DCID S2, which has a sequence number of 2.
   For the second path, the server's 1-RTT packets use DCID C2, which
   has a sequence number of 2; the client's 1-RTT packets use CID S3,
   which has a sequence number of 3.  Note that two paths use different
   packet number space.  (For the convience of distinguishing the CID
   sequence number and PATH_STATUS sequence number, we call the
   "PATH_STATUS sequence number" as "PSSN".)

  Client                                                          Server

  (client tells server to abandon a path)
  1-RTT[X]: DCID=S2 PATH_STATUS[id=1, PSSN1, status=abandon, pri.=0] ->
                                 (server tells client to abandon a path)
  <- 1-RTT[Y]: DCID=C1 PATH_STATUS[id=2, PSSN2, status=abandon, pri.=0],
                                                     ACK_MP[Seq=2, PN=X]
  (client abandons the path that it is using)
                             (server abandons the path that it is using)
       <- 1-RTT[V]: DCID=C2 RETIRE_CONNECTION_ID[1], ACK_MP[Seq=3, PN=U]

                 Figure 2: Example of closing a path

   In scenarios such as client detects the network environment change
   (client's 4G/Wi-Fi is turned off, Wi-Fi signal is fading to a
   threshold), or endpoints detect that the quality of RTT or loss rate
   is becoming worse, client or server can terminate a path immediately.

4.5.2.  Effect of RETIRE_CONNECTION_ID frame

   Receiving a RETIRE_CONNECTION_ID frame causes the endpoint to discard
   the resources associated with that connection ID.  If the connection
   ID was used by the peer to identify a path from the peer to this
   endpoint, the resources include the list of received packets used to
   send acknowledgements.  The peer MAY decide to keep sending data
   using the same IP addresses and UDP ports previously associated with
   the connection ID, but MUST use a different connection ID when doing

4.5.3.  Idle timeout

   [QUIC-TRANSPORT] allows for closing of connections if they stay idle
   for too long.  The connection idle timeout in multipath QUIC is
   defined as "no packet received on any path for the duration of the
   idle timeout".  It means that if all paths remain idle for the idle
   timeout, the connection is implicitly closed.

5.  Using TLS to Secure QUIC Multipath

   In order to facilitate loss detection and recovery when sending data
   over multiple paths, this specification defines how packets sent over
   multiple paths use different packet number spaces.  This requires
   changes in the way AEAD is applied for packet protection, as
   explained in Section 5.1, and tighter constrainst for key updates, as
   explained in Section 5.2.

5.1.  Packet protection for QUIC Multipath

   Packet protection for QUIC V1 is specified is section 5 of
   [QUIC-TLS].  The general principles of packet protection are not
   changed for QUIC Multipath.  No changes are needed for setting packet
   protection keys, initial secrets, header protection, use of 0-RTT
   keys, receiving out-of-order protected packets, receiving protected
   packets, or retry packet integrity.  However, the use of multiple
   number spaces for 1-RTT packets requires changes in AEAD usage.

   Section 5.3 of [QUIC-TLS] specifies AEAD usage, and in particular the
   use of a nonce, N, formed by combining the packet protection IV with
   the packet number.  QUIC multipath uses multiple packet number
   spaces, and thus the packet number alone would not guarantee the
   uniqueness of the nonce.  In order to guarantee this uniqueness, we
   construct the nonce N by combining the packet protection IV with the
   packet number and with the identifier of the path, which for 1-RTT
   packets is the Sequence Number of the Destination Connection ID
   present in the packet header, as defined in Section 5.1.1 of
   [QUIC-TRANSPORT], or zero if the Connection ID is zero-length.
   Section 19 of [QUIC-TRANSPORT] encode this Connection ID Sequence
   Number as a A variable-length integer, allowing values up to 2^62-1;
   for QUIC multipath, we require that a range of no more than 2^32-1
   values be used without updating the packet protection key.

   For QUIC multipath, the construction of the nonce starts with the
   construction of a 96 bit path-and-packet-number, composed of the 32
   bit Connection ID Sequence Number in byte order, two zero bits, and
   the 62 bits of the reconstructed QUIC packet number in network byte
   order.  If the IV is larger than 96 bits, path-and-packet-number is
   left-padded with zeros to the size of the IV.  The exclusive OR of
   the padded packet number and the IV forms the AEAD nonce.

   For example, assuming the IV value is "6b26114b9cba2b63a9e8dd4f", the
   connection ID sequence number is "3", and the packet number is
   "aead", the nonce will be set to "6b2611489cba2b63a9a873e2".

5.2.  Key Update for QUIC Multipath

   The Key Phase bit update process for QUIC V1 is specified in
   Section 6 of [QUIC-TLS].  The general principles of key update are
   not changed for Multipath QUIC.  Following QUIC V1, the Key Phase bit
   is used to indicate which packet protection keys are used to protect
   the packet.  The Key Phase bit is toggled to signal each subsequent
   key update.  Because of network delays, packets protected with the
   older key might arrive later than the packets protected with the new
   key.  Therefore, the endpoint needs to retain old packet keys to
   allow these delayed packets to be processed and it must distinguish
   between the new key and the old key.  In QUIC V1, this is done using
   packet numbers so that the rule is made simple: Use the older key if
   packet number is lower than any packet number frome the current key

   In QUIC multipath, some care is needed in the initiating Key Update
   process.  Because different paths use different packet number spaces
   but share a single key, when a key update is initiated on one path,
   packets sent to the other path needs to know when transition is
   complete.  Otherwise, it is possible that the other paths send
   packets with the old keys, but skip sending any packets in the
   current key phase and directly jump to sending packet in the next key
   phase.  When that happens, as the endpoint can only retain two sets
   of packet protection keys with the 1-bit Key Phase bit, the other
   paths cannot distinguish which key should be used to decode received
   packets, which results in a key rotation synchronization problem.

   To address such a synchronization issue, in QUIC multipath, if key
   update is initilized on one path, the sender should send at least one
   packet with the new key on all active paths.  Regarding the
   responding to Key Update process, the endpoint MUST NOT initiate a
   subsequent key update until a packet with the current key has been
   acknowledged on each path.

   Following the Section 5.4. of [QUIC-TLS], the Key Phase bit is
   protected, so sending multiple packets with Key Phase bit flipping at
   the same time should not cause linkability issue.

6.  Using Multipath QUIC with load balancers

   This specification follows the Connection ID negotiation defined in
   [QUIC-TRANSPORT].  For stateless or low-state load balancers
   supporting Multipath QUIC, implementations SHOULD use the
   specification of Connection ID generation and Load balancer routing
   defined in [QUIC-LB], guarantee that packets with Connection IDs
   belonging to the same connection, can be routed to same server.

7.  Packet scheduling

7.1.  Basic Scheduling

   For an outgoing packet, the packet scheduler decides which path the
   packet shall be transmitted.  A basic static scheduling strategy
   consists of four major components:

   1.  Path state: A scheduler may want to decide which path shall be
       activated to transmit data.  For instance, a scheduler can choose
       to use only one of the two paths and completely ignore the other
       one.  A scheduler marks the selected paths to be in the
       "available" state and the un-selected ones in the "standby"

   2.  Path priority: Due to the fact that costs of transmitting data
       over different paths are not always equal.  For example, the
       energy (battery) cost over a 5G path and a wifi path are very
       different.  In another example, transmissions over a wifi path
       and a cellular path may incur different charges per packet.  Note
       that a user's preference may change over time.  For instance,
       certain mobile carriers offer unlimited free data for a
       particular streaming app.  Therefore, the path priority should be
       made available in the scheduler.

   3.  Path selection algorithm: A selection algorithm splits packets
       across different paths and determines the order of paths to be
       selected.  The selection algorithm takes congestion controller
       states as inputs, such as smoothed RTTs (sRTTs), estimated
       bandwidths (eBWs) and congestion window sizes (CWNDs) as well as
       application-defined information such as path priorities and path
       states.  The outputs of the algorithm is an ordered list of paths
       to put a packet on.  To name a few, some of the commonly used
       algorithms are: - Round-Robin: There is no priority. it selects
       paths one by one in order to transmit data. - Lowest-RTT: It
       first chooses the path with the lowest RTT and feeds packets to
       it until that path's congestion window is full.  Then it chooses
       the path with the second lowest RTT. - Highest-Sending-Rate: It
       first chooses the path with the highest bandwidth and feeds
       packets to it until that path's congestion window is full.  Then
       it chooses path with the second largest bandwidth.

   4.  Packet redundancy: One major challenge in multi-path transmission
       is that a packet loss on the slow path might block the overall
       transmission when packets are split across fast-changing paths.
       As the path selection algorithm takes inputs from congestion
       controllers on predictions of the network which may not be
       accurate enough for fast-changing wireless channels, such an
       imprecise estimation could lead to network overuse/underuse.  A
       solution to this problem is to implement packet redundancy
       strategy.  A redundancy strategy can be applied to only ACK
       packets(partial redundancy) or all data packets (full
       redundancy).  It is up to the application to determine whether,
       when, and on which packets to activate redundancy.

   The path state and path priority are managed by PATH_STATUS frame.
   The path selection algorithm and packet redundancy are application
   related and should be controlled by the applicaiton.

7.2.  Scheduling with QoE Feedback

   Applications may have completely different QoE requirements---the
   interactive applications are delay sensitive, while the video
   streaming applications are more throughput sensitive.  There is thus
   a trend of cross-layer design that takes applications' demands into
   account when managing paths or scheduling packets.  The QoE feedback
   is used to fully support application-awareness in multipath
   scheduling and is carried in the QOE_CONTROL_SIGNALS frames Figure 6.
   The QOE_CONTROL_SIGNALS frames can include general application-level
   information that is needed by the schedulers.  The frequency of such
   feedback should be controlled to limit the amount of extra packets.
   The QoE control signal allows a synchronization of viewpoints between
   two endhosts.  It is up to the application to determine the
   interpretation of QoE control signals.

7.3.  Per-stream Policy

   As QUIC supports stream multiplexing, streams are allowed to
   associate stream priorities to express applications intent.  For
   instance, objects in a web page may be dependent on others and thus
   have different priorities multipath quic scheduler.  A stream
   priority-aware packet scheduling algorithm will improve the
   performance notably.

       High priority  /\  +---------+
                      ||  |         |
                      ||  +---------+
                      ||  +---------+
                      ||  |         |
                      ||  +---------+
                      ||     ...          User-defined stream priority
                      ||  +---------+
       Low priority   ||  |         |
                      ||  +---------+
       High priority  /\  +---------+
                      ||  |         |
                      ||  +---------+
                      ||  +---------+
                      ||  |         |
                      ||  +---------+
                      ||     ...          Default stream priority
                      ||  +---------+
       Low priority   ||  |         |
                      ||  +---------+

                         Figure 3: Stream priority

   The priority management scheme composes two separated priority
   ranges.  The user-defined priority range includes those streams that
   the applications explicitly designate priorities, while the default
   priority range includes the streams with no priorities set by the
   applications.  Only when the streams in the user-defined ranges have
   no data to send, the streams in the default priority range can send.
   In the same range, one can use the weighted-round robin for
   scheduling---the higher-priority streams get more quota for data to
   send in each round.  One can also dynamically set/change the
   priorities of the streams in the default priority ranges to enable
   short stream first if needed.

8.  Congestion control and loss detection

8.1.  Congestion control

   Implementations MAY support coupled congestion controllers such as
   LIA [MPTCP-LIA], OLIA [MPTCP-OLIA], and etc., or support decoupled
   congestion controllers in environments using disjoint network paths.

   In decoupled congestion control, each path runs its own congestion
   controller without interacting with the congestion controllers of
   other paths.  That is to say, in the aspect of congestion control, a
   path behaves exactly the same as a normal QUIC connection over the
   same network path.

   Each path MAY choose congestion control algorithm independently.

8.2.  Packet number space and acknowledgements

   Each path has it's own packet number space for transmitting 1-RTT

   Acknowledgements of Initial and Handshake packets MUST be carried
   using ACK frames, as specified in [QUIC-TRANSPORT].  The ACK frames,
   as defined in [QUIC-TRANSPORT], do not carry path identifiers.  If
   for some reason ACK frames are received in 1RTT packets while the
   state of multipath negotiation is ambiguous, they MUST be interpreted
   as acknowledging packets sent on path number 0.  After endpoints
   successfully negotiate multipath support, they SHOULD use ACK_MP
   frames instead of ACK frames to signal acknowledgement of 1-RTT
   packets, and also 0-RTT packets as specified in Section 10.1.

   ACK_MP frame Section 9.2 can be returned via either a different path,
   or the same path identified by the Path Identifier, based on
   different strategies of sending ACK_MP frames.

8.3.  Flow control


9.  New frames

   All the new frames MUST be sent in 1-RTT packet, and MUST NOT use
   other encryption levels.

   If an endpoint receives MP frames from packets of other encryption
   levels, it MUST return MP_PROTOCOL_VIOLATION as a connection error
   and close the connection.

9.1.  PATH_STATUS frame

   PATH_STATUS Frame are used by endpoints to inform the peer of the
   current status of one path, and the peer should send packets
   according to the preference expressed in these frames.  Endpoint use
   the sequence number of the CID used by the peer for PATH_STATUS
   frames (describing the sender's path identifier).  PATH_STATUS frames
   are formatted as shown in Figure 4.

     PATH_STATUS Frame {
       Type (i) = TBD-03 (experiments use 0xbaba03),
       Path Identifier (i),
       Path Status sequence number (i),
       Path Status (i),
       Path Priority (i),

                     Figure 4: PATH_STATUS Frame Format

   PATH_STATUS Frames contain the following fields:

   Path Identifier: A variable-length integer specifying the path

   Path Status sequence number: A variable-length integer specifying the
   sequence number assigned for this PATH_STATUS frame.  There is a
   different path status sequence number space for each path.

   Available values of Path Status field are:

   *  0: Abandon

   *  1: Standby

   *  2: Available

   If the value of Path Status field is 2-available, the receiver side
   can use the Path Priority field to express the priority weight of a
   path for the peer.

   Frames may be received out of order.  A peer MUST ignore an incoming
   PATH_STATUS frame if it previously received another PATH_STATUS frame
   for the same Path Identifier with a sequence number equal to or
   higher than the sequence number of the incoming frame.

   PATH_STATUS frames SHOULD be acknowledged.  If a packet containing a
   PATH_STATUS frame is considered lost, the peer should only repeat it
   if it was the last status sent for that path -- as indicated by the
   sequence number.

9.2.  ACK_MP frame

   ACK_MP frame allows for acknowledgements on different paths.  ACK_MP
   frame is formatted by adding a Path Identifier field to
   [QUIC-TRANSPORT] ACK frame.  ACK_MP frame is formatted as shown in
   Figure 5.

     ACK_MP Frame {
       Type (i) = TBD-00..TBD-01 (experiments use 0xbaba00..0xbaba01),
       Path Identifier (i),
       Largest Acknowledged (i),
       ACK Delay (i),
       ACK Range Count (i),
       First ACK Range (i),
       ACK Range (..) ...,
       [ECN Counts (..)],

                       Figure 5: ACK_MP Frame Format

   Type(i) = TBD-00 (experiments use 0xbaba00) , with no ECN Counts
   Type(i) = TBD-01 (experiments use 0xbaba01) , with ECN Counts


   QOE_CONTROL_SIGNALS frame is used to carry quality of experience
   (QoE) information.  A typical use of such information is to provide
   feedback to help application-aware scheduling.  Note that different
   applications may have very different needs, the interpretation of the
   QoE control signal can be up to the users.  QOE_CONTROL_SIGNALS
   frames are formatted as shown in Figure 6.

       Type (i) = TBD-02 (experiments use 0xbaba02),
       Path Identifier (i),
       QoE Control Signals Length(8),
       QoE Control Signals (..)

                 Figure 6: QOE_CONTROL_SIGNALS Frame Format

   QOE_CONTROL_SIGNALS frames may be received out of order, peers SHOULD
   pass them to the application as they arrive.  Although
   QOE_CONTROL_SIGNALS frames are not retransmitted upon loss detection,
   they are ack-eliciting [QUIC-RECOVERY].

10.  Implementation Considerations

   ## Management of acknowledgements delay If implementation uses
   ACK_FREQUENCY Frame in [QUIC-DELAYED-ACK] to let senders control the
   frequency of acknowledgements, the same mechanism can be used in
   multi-path QUIC.  There are two parameters in the ACK_FREQUENCY
   Frame, "Packet Tolerance" and "Update Max Ack Delay".

   Those two parameters are typically computed in real time based on
   observed performance:

   *  "Packet Tolerance" is set to a fraction of the congestion window

   *  "Update Max Ack Delay" is set to a fraction of the RTT -- but not
      smaller than the specified min delay

   In multi-path QUIC, there are multiple paths with different RTT and
   different congestion windows.  In this draft, it is suggested that
   implementations can use the smallest RTT of the available paths to
   compute the delay, and use the sum of congestion windows of all
   available(not including standby/abandon state) paths.

10.1.  Handling of 0-RTT packets

   The draft specifies a packet number space for each path.  Because
   multi-path is enabled after the handshake negotiation complete, there
   will be a separate context for each Connection ID after multi-path is
   negotiated. 0-RTT packets are sent before these per path contexts are
   established.  To avoid confusion, this draft provides a way for
   implementations to deal with 0-RTT packets that is both easy to
   implement and compatible with [QUIC-TRANSPORT]:

   *  All 0-RTT packet are initially tracked in the "global" application

   *  On the client side, 0-RTT packets are initially sent in the
      "global" application context.  The handshake concludes before any
      1-RTT packet can be sent or received.  When the handshake
      completes, if multipath is negotiated, the tracking of 0-RTT
      packets moves from the "global" application context to the "path
      0" application context.  That means the sequence number of the
      first 1-RTT packets sent by the client will follow the sequence
      number of the last 0-RTT packet.

   *  On the server side, the negotiation completes after the client
      first flight is received and the the server first flight is sent.
      0-RTT packets are received after that.  If multipath is
      negotiated, they are considered received on "path 0".

   In conclusion, 0-RTT packets are tracked and processed with path
   identifier 0.

11.  Security Considerations


12.  IANA Considerations

   This document defines a new transport parameter for the negotiation
   of enable multiple paths for QUIC, and three new frame types.  The
   draft defines provisional values for experiments, but we expect IANA
   to allocate short values if the draft is approved.

   The following entry in Table 1 should be added to the "QUIC Transport
   Parameters" registry under the "QUIC Protocol" heading.

    | Value                        | Parameter Name.  | Specification |
    | TBD (experiments use 0xbaba) | enable_multipath | Section 3     |

           Table 1: Addition to QUIC Transport Parameters Entries

   The following frame types defined in Table 2 should be added to the
   "QUIC Frame Types" registry under the "QUIC Protocol" heading.

       | Value              | Frame Name          | Specification |
       | TBD-00 - TBD-01    | ACK_MP              | Section 9.2   |
       | (experiments use   |                     |               |
       | 0xbaba00-0xbaba01) |                     |               |
       | TBD-02             | QOE_CONTROL_SIGNALS | Section 9.3   |
       | (experiments use   |                     |               |
       | 0xbaba02)          |                     |               |
       | TBD-03             | PATH_STATUS         | Section 9.1   |
       | (experiments use   |                     |               |
       | 0xbaba03)          |                     |               |

              Table 2: Addition to QUIC Frame Types Entries

13.  Changelog

14.  Appendix.A Scenarios related to migration

   In QUIC V1, there are four scenarios related to migration: CID
   renewal, NAT Rebinding, controlled migration, and migration to server
   preferred address.  It would be useful to explain exactly how these
   four scenarios are supported or changed with Multipath QUIC.  For V1,
   these scenarios are described as follow:

   *  CID Renewal happens when the client starts using a new CID for
      1-RTT packet, while still using the same four-tuple.  This is
      typically done for privacy, for example after a long period of
      silence.  The expected result is that the server will also use a
      new CID for its next packets.  In that scenario, RTT and
      congestion control parameters remain the same before and after

   *  NAT Rebinding happens when a NAT on the path changes its mappings.
      The server receives packets that bear the same CID as previously,
      but arrive on a different four tuple.  The complication is that
      this could be an attack in which the attacker captures a packet
      from the client and resends it from a different address.  The
      server is expected to perform continuity tests for both the old
      and the new path, typically using a different CID for the new
      path.  If the continuity test on the new path succeeds before the
      old path, the server migrates to the new path, otherwise it
      continues using the old path and ignores the new path.

   *  Controlled migration happens when a client tests a new path.  The
      server receives packets that bear a new CID and arrive on a new
      four tuple.  The server responds to the path challenge, perform
      its own continuity test on the new path.  If the client sends non-
      path-validation packets on the new path, the server switches to
      sending on the new path and discards the old path.

   *  Preferred address migration happens when the server sends the
      preferred address TP during the exchange.  The client performs a
      controlled migration to the new path, and if that is successful
      discards the old path.

   We could sum up these scenarios in the following table:

        | CID | 4-tuple | preferred address | result             |
        | Old | Old     | -                 | Not a migration.   |
        | Old | New     | -                 | NAT Rebinding.     |
        | New | Old     | -                 | CID Renewal.       |
        | New | New     | matches PFA       | Migration to       |
        |     |         |                   | Preferred Address. |
        | New | New     | other             | Controlled         |
        |     |         |                   | Migration.         |

                 Table 3: Scenarios related to migration

   The expectation in those scenarios is:

       | Scenario     | Expectation                                |
       | Not a        | Continue using existing path               |
       | migration    |                                            |
       | NAT          | After validation, use new path and discard |
       | Rebinding    | previous path.                             |
       | CID Renewal  | Create new path with new CIDs, discard old |
       |              | path.  Reuse RTT and CC parameter.         |
       | Controlled   | Create new path with new CIDs.  Server     |
       | Migration    | creates a new path,ready to use both       |
       |              | paths.  Client may later discard old path. |
       | Migration to | Same as Controlled Migration, but the      |
       | Preferred    | client is expected to abandon the old path |
       | Address      |                                            |

           Table 4: Expectation in scenarios related to migration

   In multipath quic, client / server create a new path and abandon the
   old path to do exactly the same thing as connection migration in the
   previous scenarios.

15.  Appendix.B Considerations on RTT estimate and loss detection

   QUIC implementations use RTT estimates in many ways:

   *  For loss detection, RTT estimates are used to evaluate how long to
      wait for an acknowledgement before a packet is declared lost.

   *  Several congestion control algorithm (e.g.  LEDBAT, VEGAS,
      HYSTART) use variations of the RTT above the minimum value to
      detect the beginning of congestion.

   *  BBR uses the minimal RTT to compute the minimal size of the
      congestion window for a target data rate.

   *  ACK delays are often set as a fraction of the RTT.

   In a multipath environment, the RTT can be estimated each time a new
   packet is acknolwedged.  However, the observed RTT will vary not only
   based on the state of the send path, but also based on the choice of
   the return path used for acknowledgements.  Each RTT measurement will
   the sum of the one-way delay on the send path and the one-way delay
   on the return path.  This has a number of implications for the
   different ways of using the RTT presented above:

   *  If the goal is to detect possible losses, it is probably
      sufficient to consider all RTT measurements for a given path.
      Classic formulas like adding smoothed RTT and a number of
      deviations aim at estimating a reasonable upper bound of the
      acknowledgement delays.  Statistics on observed acknowledgement
      delays will provide a valid estimate, regardless of the selection
      of the return path by the peer.

   *  If the goal is to detect the onset of collision and tune a
      congestion algorithm, the variations of delays due to the choices
      of return paths will be a source of errors.  Implementations will
      need to pick a strategy, such as for example only considering
      acknowledgements received through the "fastest" return path, or
      maybe those received through the matching four tuple for the
      sending path.  An alternative would be to use time stamps to
      directly estimate variations of the one way delays.
      [QUIC-Timestamp] provides good support for such one-way-delay

   *  If BBR is in use and ACKs are returned on different paths, it may
      cause an ambiguity issue with the computation of bandwidth and
      delay product (BDP).  In BBR, BDP is used to limit the number of
      inflight packets.  One may choose to use the smallest RTT measured
      to compute BDP.  However, if the majority of ACKs are returned
      from a high-latency path, the cwnd = cwnd_gain * bandwidth *
      min_rtt may be lower than what is needed to achieve good
      performance.  One possible solution is to transmit a new packet
      and its ACK on the same path.  Other possible solutions may
      include transmitting ACKs on the shortest path with relative
      increase of cwnd_gain.  For the time being, we think there is a
      research problem and it is up to the implementers to pick the best

16.  Appendix.C Difference from past proposals

   This proposal differs from past proposals
   [I-D.deconinck-quic-multipath] in two fundamental perspectives:

   *  The multi-path QUIC is built on top of the concept of the
      bidirectional paths, which readily fits into the nature of both
      cellular and wifi links that cover the majority of multi-path
      applications in QUIC while keeping the design simple and easy to
      implement.  In doing so, we are able to re-use most of the current
      QUIC transport design with the sole addition of three new frames.

   *  The multi-path QUIC design enables feedback-based dynamic
      scheduling strategy.  As the major goal of multi-path QUIC is to
      enhance performance in mobile applications, where the sender and
      receiver may have different viewpoints about the fast-changing
      wireless connectivity, especially in high-mobility scenarios, the
      proposed design allows the sender and receiver to synchronize
      their viewpoints via message exchange in ACK packet in order to
      maximize performance.

17.  References

17.1.  Normative References

              Iyengar, J., Ed. and I. Swett, Ed., "Sender Control of
              Acknowledgement Delays in QUIC", Work in Progress,
              Internet-Draft, draft-iyengar-quic-delayed-ack-02,

   [QUIC-LB]  Duke, M., Ed. and N. Banks, Ed., "QUIC-LB: Generating
              Routable QUIC Connection IDs", Work in Progress, Internet-
              Draft, draft-ietf-quic-load-balancers,

              Iyengar, J., Ed. and I. Swett, Ed., "QUIC Loss Detection
              and Congestion Control", Work in Progress, Internet-Draft,

   [QUIC-TLS] Thomson, M., Ed. and S. Turner, Ed., "Using TLS to Secure
              QUIC", Work in Progress, Internet-Draft, draft-ietf-quic-
              tls, <https://tools.ietf.org/html/draft-ietf-quic-tls>.

              Iyengar, J., Ed. and M. Thomson, Ed., "QUIC: A UDP-Based
              Multiplexed and Secure Transport", Work in Progress,
              Internet-Draft, draft-ietf-quic-transport,

   [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, <https://www.rfc-editor.org/info/rfc8174>.

17.2.  Informative References

              Coninck, Q. and O. Bonaventure, "Multipath Extensions for
              QUIC (MP-QUIC)", Work in Progress, Internet-Draft, draft-
              deconinck-quic-multipath-06, 2 November 2020,

              Raiciu, C., Handly, M., and D. Wischik, "Coupled
              Congestion Control for Multipath Transport Protocols",
              October 2011, <https://tools.ietf.org/html/rfc6356>.

              Khalili, R., Gast, N., and J. Boudec, "Opportunistic
              Linked-Increases Congestion Control Algorithm for MPTCP",
              July 2014, <https://datatracker.ietf.org/doc/html/draft-

              Huitema, C., "Quic Timestamps For Measuring One-Way
              Delays", August 2020,

Authors' Addresses

   Yanmei Liu
   Alibaba Inc.

   Email: miaoji.lym@alibaba-inc.com

   Yunfei Ma
   Alibaba Inc.

   Email: yunfei.ma@alibaba-inc.com

   Christian Huitema
   Private Octopus Inc.

   Email: huitema@huitema.net

   Qing An
   Alibaba Inc.

   Email: anqing.aq@alibaba-inc.com

   Zhenyu Li

   Email: zyli@ict.ac.cn