Multipath sequence maintenance

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 6 May 2021.¶

Copyright Notice

Copyright (c) 2020 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

Mobile end user devices nowadays are mostly equipped with multiple network interfaces allowing to connect to more than one network at a time and thus increase data throughput, reliability, coverage and so on. Ideally the user data stream originating from the application at the device is split between the available (here: N) paths at the sender side and re-assembled at an intermediate aggregation node before transmitted to the corresponding host in the network as depicted in Figure 1.¶

                   ------------
                  /            \
       +---------| Access Net 1 |--+
       |         \             /   |
       |          -------------    |
       |          ------------     |
       |         /            \    |
       | +------| Access Net 2 |-+ |
       | |       \            /  | |
       | |        ------------   | |
       | |                       | |
       | |                       | |
       | |               +-------+-+---+
    +--+-+-+             |             |           +------+
    |End   |             | Aggregation +----/.../--| Host |
    |User  |             |    Node     |           +------+
    |Device|             |             |
    +--+-+-+             +-------+-+---+
       | |                       | |
       | |     --------------    | |
       | |    /              \   | |
       | +---| Access Net N-1 |--+ |
       |      \              /     |
       |       --------------      |
       |                           |
       |          ------------     |
       |         /            \    |
       +--------| Access Net N |---+
                 \            /
                  ------------

However, when several paths are utilized concurrently to transmit user data between the sender and the receiver, different characteristics of the paths in terms of bandwidth, delay, or error proneness can impact the overall performance due to delayed packet arrival and need for re-transmit in case of lost packets. Without further arrangements the original order of packets at the sending UE side is no longer maintained at the receiving host and a re-ordering or re-arrangement has to occur before delivery to the application at the far end site. This can be performed at earliest at the aggregation node with a minimum additional delay due to re-transmission requests or at latest either by the application on the host itself or the transmission protocol.¶

It is a goal of the present document to collect and describe mechanisms to maintain the sequence of split traffic over multiple paths. These mechanisms are generic and not dedicated to a specific multipath network protocol, but give clear guidance on requirements and benefits to maintainers of multipath network protocols.¶

2. State of the Art

Regular TCP protocol [RFC0793] offers such mechanism with queues for in-order and out-of order (including damaged, lost, duplicated) arrival of packets.¶

This is also provided by MPTCP [RFC6824] as the first and successful Multipath protocol which however also requires new methods as sequence numbers both on (whole) data (stream) and subflow level to ensure in-order delivery to the application layer on the receiver side [RFC8684]. Moreover, careful design of buffer sizes and interpretation of sequence numbers to distinguish between (delayed) out-of-order packets and completely lost ones has to be considered.¶

[I-D.bonaventure-iccrg-schedulers] already reflects on proper packet scheduling schemes (at the sender side) to reduce the effort for re-assembly or even make such (time consuming) treatment unnecessary.¶

MP-QUIC [I-D.deconinck-quic-multipath] introduces the concept of uniflows with own IDs claiming to get rid of additional sequence numbers for re-ordering as required in Multipath TCP [RFC6824]. Although [] admits that statistical performance information should help a host in deciding on optimum packet scheduling and flow control a dedicated packet scheduling policy is out of scope of that document. A further improvement versus MPTCP can be achieved by decoupling paths used for data transmission from those for sending acknowledgments (ACKs) or claiming for re-transmission by NACKs to not introduce further latency.¶

[I-D.ietf-quic-recovery] specifies algorithms for QUIC Loss Detection and Congestion Control by using measurement of Round Trip Time (RTT) to determine when packets should be retransmitted. Draft [I-D.huitema-quic-ts] proposes to enable one way delay (1WD) measurements in QUIC by defining a TIME_STAMP frame to carry the time at which a packet is sent and combine the ACKs sent with a timestamp field and thus allow for more precise estimation of the (one-way) delay of each uniflow, assisting proper scheduling decisions.¶

Also other protocols as Multi-Access Management Services (MAMS) [RFC8743] consider the need for re-ordering on User Plane level which may be done at network and client level by introducing a new Multi-Access (MX) Convergence Layer. [I-D.zhu-intarea-mams-user-protocol] introduces accordingly Traffic Splitting Update (TSU) messages and Packet Loss Report (PLR) messages including beside others Traffic Splitting Parameters and an expected next (in-order) sequence number, respectively.¶

[I-D.zhu-intarea-gma] on Generic Multi-Access (GMA) Convergence Encapsulation Protocols introduces a trailer-based encapsulation which carries one or multiple IP packets or fragments thereof in a Protocol Data Unit (PDU). At receiver side PDUs with identical Sequence Numbers (in the trailer) are to be placed in the relative order indicated by a so-called Fragment Offset.¶

3. Problem Statement

Assuming for simplicity the minimum multipath scenario with two separate paths for transmission of a flow of packets with sequence numbers (SN) SN1 ... SM. In case the scheduling of packets is done equally to both paths and path 2 exhibits a delay of the duration of transmission time required for e.g. two packets (assuming fixed packet size and same constant data for both paths) for an exemplary App-originated sequence of packets as SN1 SN2 SN3 SN4 SN5 SN6 SN7 SN8 ... the resulting sequence of packets could look as depicted in Figure 2 which of course depends on the queue processing and buffering at the Aggregation Proxy.¶

APP              UE             Aggregation Node                 Host
 |  SN1 ... SN8  |                         |                       |
 |-------------->|path 1 SN1 SN3 SN5 SN7...|                       |
 |               |------------------------>|                       |
 |               |path 2 SN2 SN4 SN6 SN8...|                       |
 |               |------------------------>|                       |
 |               |                         |SN1 SN3 SN2 SN5 SN4 SN7|
 |               |                         |======================>|
 |               |                         |                       |

In such a case re-ordering at the Aggregation Node would be simple and straight forward. It even could be avoided if the scheduling would already take the expected different delays into account (e.g. by pre-delaying the traffic on path 1 thus of course not leveraging the lower delay). Different from this simplistic scenario in general the data rate on both paths will vary in time and be not equal, also different and variable latency (jitter) per path will be introduced and in addition loss of packets as well as potential duplication may occur making the situation much more complicated. In case of loss detection after a threshold waiting time a retransmission could be initiated by the Host or if possible already by the Aggregation Node. Alternatively the UE could send redundant packets in advance coded in such a way that it allows for derivation of e.g. one lost packet per M correctly received ones or by a (real-time) application able to survive singular lost packets.¶

Holding multiple queues and a large enough buffer both at UE and at the Aggregation Node would be required to apply proper scheduling at UE and reordering during re-assembly at Aggregation Node to mitigate the sketched impact of multiple paths variable characteristics in terms of transmission performance.¶

...¶

4. Scheduling mechanisms

Scheduling mechanisms decide on sender side how traffic is distributed over the paths of a multipath-setup. [I-D.bonaventure-iccrg-schedulers] gives an overview of possible distribution schemes. For this document it is assumed, that schedulers are used, which simultaneously distribute traffic over more than one path, whereas path characteristics differ between those multiple paths (e.g. a latency difference exists). While on the one hand, the traffic scheduling causes out-of-order multipath delivery when simultaneously utilize heterogeneous paths, it can also be used to mitigate this problem. Pre-delaying data on a fast path, according to the latency difference of the slowest path is aimed e.g. by OTIAS [OTIAS], DAPS [DAPS] and BLEST [BLEST]. However, the success is much dependent on the accuracy of path information like path latency, throughput and packet loss rate. In heterogeneous and volatile environments most often such information have to be estimated, e.g. using congestion control. That means, it takes at least one RTT to gain first indications and probably several RTTs to converge to a worthwhile accuracy. Changes of path characteristics in sub-RTT time frames puts such a system to test. Dependent on the demand on in-order delivery and/or the accuracy of the relevant path information, scheduling might be an exclusive alternative or can be applied in conjunction with other discussed mechanisms in this document.¶

Scheduling will not help to overcome any degree of out-of-order delivery, when the scheduling goal is different to this. For example a strict cost prioritization of Wi-Fi over cellular access in a mobile phone might be assumed counterproductive.¶

5. Resequencing mechanisms

Resequencing mechanism are responsible to modify the sequence of received data split over multiple paths according to a sequencing scheme. The degree of resequencing can reach from no measure up to re-generating the exact order.¶

Typically at least one sequencing scheme, describing the order of how data was generated on sender side is prerequisite. This is referred to as "connection sequencing". Under certain circumstances an additional sequencing scheme per path of the multi-path setup can be leveraged, to optimize packet loss detection and is further elaborated in Section 5.6. For most multipath protocols both sequencing schemes are already available. Packet loss detection becomes important when multipath protocols are applied which do not guarantee successful transmission as TCP achieves by acknowledgement of successful reception. For example [I-D.amend-tsvwg-multipath-dccp] or the combination of [I-D.deconinck-quic-multipath] and [I-D.ietf-quic-datagram] are unreliable protocols in that sense.¶

For simplicity all the mechanism described in the following are explained based on two paths but in principle would work with an unlimited number though.¶

6. Recovery mechanisms

Recovering packets, in particular lost packets or assumed lost packets on receiver side avoids re-transmission and potentially mitigates the resequencing process in respect to detecting packet loss. Shorter latencies will be an expected outcome. Discussing the complexity, computation overhead and reachable benefit is subject of this section.¶

7. Retransmission mechanisms

Re-transmission becomes interesting when it can help to reduce the time spent on waiting for outstanding packets for re-sequencing. In particular scenarios when for example a known path RTT (Round Trip Time) lets expect a shorter time to re-transmit than wait for packet loss detection, a likely scenario in e.g. Figure 1. It could also avoid a potential late triggering of re-transmission by the end-to-end service.¶

Authors' Addresses