Forward Erasure Correction for QUIC loss recovery

Introduction The QUIC protocol relies on retransmissions to ensure the reliable delivery of stream data. Retransmitting the lost information requires the loss recovery mechanism to identify lost packets which may take up to several hundreds of milliseconds . Depending on their delay-sensitivity, some applications using QUIC could not afford such a waiting time to ensure a good quality of experience to their users. Works has already been done to consider the use of Forward Erasure Correction (FEC) for the QUIC protocol to ensure timely data delivery for delay-sensitive applications . This document defines additions to the QUIC protocol to extend its loss recovery mechanism and make it able to recover from packet losses prior to retransmission using FEC.

Conventions and Definitions The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

FEC-related definitions Source symbol: piece of information exchanged by two endpoints. This document considers QUIC packets payloads as source symbols. Repair symbol: redundant information constructed from the combination of several source symbols. Erasure: loss of one or more symbols Erasure correction code: algorithm generating repair symbols and reconstructing missing source symbols from a set of source and repair symbols. Forward Erasure Correction: process of recovering erased symbols before the detection of their erasure. FEC scheme: the conjunction of an erasure correction code and the specific protocol elements required to use it with the design described in this document. Encoder: entity producing repair symbols using an erasure correction code. The encoder can be a library used by the protocol implementation or a program running in a separate process or machine. Decoder: entity reconstructing missing source symbols using an erasure correction code. The decoder can be a library used by the protocol implementation or a program running in a separate process or machine. Coding window: window of source symbols that are required to decode a repair symbol.

The network channel and the coding channel QUIC endpoints exchange information over a network channel. Adding Forward Erasure Correction to QUIC enables an endpoint to receive information over a coding channel. A coding channel can be seen as a communication channel between a QUIC receiver and a FEC decoder. The decoder often runs on the same machine as the QUIC receiver and can even be part of the protocol implementation itself. A source symbol is received through the coding channel when it is recovered by the FEC decoder instead of being explicitly received through the network. Only the data sent on the network channel is really transmitted on the wire between the two QUIC endpoints. illustrates how symbols can be received from both channels.

Receiving symbols through the network channel | SOURCE_SYMBOL(1) | APP_DATA(1) | | |------------------->|------------->| | PACKET(2)[SOURCE_SYMBOL(2)] | | | |--------------x | | | | | | | | PACKET(3)[REPAIR_SYMBOL] | | | |----------------------------->| REPAIR_SYMBOL | | | |------------------->| | | | |(recomputes) | | | |(the source) | | | |(symbol ) | | | | | | | | APP_DATA(2) | | | |------------->| | | | | | | | | ]]> In this illustration, the sender sends three QUIC packets through the network channel. Packets 1 and 2 carry one source symbol each and packet 3 carries one repair symbol protecting the two source symbols. The source symbols each carry application data (APP_DATA), e.g. through STREAM frames Packet 2 is lost due to network imperfection preventing SOURCE_SYMBOL(2) from being received through the network channel. The FEC decoder reconstructs SOURCE_SYMBOL(2) by combining SOURCE_SYMBOL(1) and REPAIR_SYMBOL. SOURCE_SYMBOL(2) is thus received through the coding channel. Note that SOURCE_SYMBOL(2) is not received as a packet since QUIC packets are only exchanged through the network channel. On the other hand, source symbols are carried by QUIC packets through the network channel and sent directly to the decoder through the coding channel..

FEC and the loss recovery mechanism The FEC mechanism described in this document is an enhancement of the classical QUIC loss recovery mechanism . It does not replace it by any means. A QUIC endpoint MAY ignore every received repair symbol and MAY not perform any symbol recovery at all. The FEC mechanism is only intended to allow a receiver recovering faster from packet losses on the network channel if it values the timeliness of data delivery.

Protocol requirements for protecting information through FEC In this section, we list the points that must be defined by the protocol for allowing QUIC endpoints to protect information using FEC.

Defining the FEC-protected parts of a QUIC payload There is no need to protect every piece information sent on the wire by QUIC. Some pieces of information are already sent redundantly (e.g. ACKs) and some data are not delay sensitive and can be retransmitted later with no harm (e.g. background download on a separate stream). Endpoints need to agree on which parts of the packets are part of the protected source symbols and how to compose a source symbol from what is sent on the wire. Versions 01 and 02 of only protect streams payload. The idea is simple when using a single stream but becomes complicated and requires more signaling in a multi-stream scenario. It also cannot protect DATAGRAM frames. In this document, we propose to consider whole frames as part of the source symbols. Source symbols are thus the counterpart to QUIC packets for the coding channel. Packets carry frames through the network channel and source symbols allow receiving frames from the coding channel. In order to reduce signalling between the peers, a single source symbol MUST NOT contain the frames of several QUIC packets at the same time.

Identifying the source symbols from QUIC packets Upon reception of a QUIC packet, a receiver needs to identify the source symbols contained in the packet to forward to the FEC decoder. The decoder also needs to know which source symbols were lost. Since QUIC does not enforce sending contiguously increasing packet numbers, it is not possible for a receiver to distinguish a lost packet from a packet that has never been sent. Furthermore, a QUIC sender may not want to protect the payload of some packets if they do not carry latency-sensitive information. Source symbols are thus attributed a Symbol ID (SID). The SID of the first source symbol MUST be zero and the SIDs are contiguously increasing. When a FEC decoder notes gaps in the received SIDs, the missing SIDs correspond to lost source symbols that can be recovered using FEC. As the QUIC packet number cannot be used to carry the SID, it must be transmitted using either a dedicated QUIC frame or a dedicated header field. The second solution being incompatible with , it is not discussed in this document. The source symbol payload can either be put inside a dedicated frame () or infered when handling a specific frame (). Both alternatives lead to the exact same packet wire format and outcome.

Alternative 1: sending the source symbol inside a frame This alternative is compatible with . It defines a new SOURCE_SYMBOL frame as shown in .

SOURCE_SYMBOL frame format The frame explicitly represents a source symbol. The FEC Protected Payload field is analogous to the payload of a QUIC packet: it contains a sequence of frames that are protected by FEC. The SOURCE_SYMBOL frame is idempotent and explicit: it exactly describes the frames inside the source symbol. The main drawback is that existing QUIC implementations are not used to write frames inside other frames which may increase the implementation cost of the approach. An example of the use of the SOURCE_SYMBOL frame is shown in .

SID Alternative 1 | | | | Pkt(6)[STREAM(2, "xyz"), | | SOURCE_SYMBOL(1, { STREAM(8, "def"), | | DATAGRAM("msg") }] | |------------------------------------------------------------>| | | | Pkt(1)[ACK[..]] | |<------------------------------------------------------------| | | ]]> The two SOURCE_SYMBOL frames contain the source symbols with SID 0 and 1. The first source symbol contains a frame for stream 4 while the second one contains a frame for stream 8 and a DATAGRAM frame. The ACK frame of packet 5 and the STREAM frame for stream 2 of packet 6 are not part of the source symbol and are thus not protected by FEC. In this scenario, every source symbol is correctly received and their reception can be deduced from the acknowledgements of packets 5 and 6.

Alternative 2: only sending the SID inside a frame This alternative is compatible with . It defines a new SID frame as shown in .

SID frame format A QUIC packet carrying an SID frame means that the frames following the SID frames in the packet payload are part of a FEC source symbol. The SID of this symbol is represented by the only field of the SID frame. A packet MUST NOT contain more than one SID frame. A packet whose payload is FEC-protected MUST contain a SID frame whose SID field is the SID of the related source symbol. This alternative has one drawback: the SID frame is not idempotent since it is related to its containing packet. An example of the use of the SID frame is shown in .

SID Alternative 2 | | | | Pkt(6)[STREAM(2, "xyz"), | | SID(1), STREAM(8, "def"), | | DATAGRAM("msg")] | |------------------------------------------->| | | | Pkt(1)[ACK[..]] | |<-------------------------------------------| | | ]]> The source symbols carried by the packets in this example are the same as for and the outcome and wire format are the same.

Sending the repair symbols The repair symbols and the metadata attached to them are transferred using the REPAIR frame shown in .

REPAIR frame format The payload of the REPAIR frame is specific to the underlying FEC scheme. In addition to the repair symbol itself, it may contain any metadata needed by the erasure-correcting code (e.g. identifying the source symbols protected by the repair symbol carried by the frame). Depending on the FEC scheme, the REPAIR frame MAY contain only a part of a repair symbol.

Announcing the coding window size The receiver needs to store the received symbols in order to recover the lost source symbols. The FEC_WINDOW frame is sent by the receiver to announce the number of symbols that can be stored simultaneously by the receiver at a given point of time. The format of the FEC_WINDOW frame is described in .

FEC_WINDOW frame format The FEC Window Epoch field is a unique identifier for the announced window. The first epoch is set to 0. Each time a new FEC_WINDOW frame is sent, the FEC Window Epoch field is increased by exactly one. The Window Size field indicates the number of symbols that can be stored simultaneously by the receiver. The Window Size value overrides the window sizes received for smaller window epochs.

Announcing the recovered symbols The FEC receiver MAY advertise the source symbols that have been received either through the network or coding channels to avoid the sender retransmitting the data of the recovered source symbols. This can be done using the SYMBOL_ACK frame as shown in .

SYMBOL_ACK frame format The frame has a similar format as the ACK frame, announcing the reception of SIDs instead of packet numbers. In addition to symbols recovered by FEC, this frame MAY also announce symbols received regularly through the network to avoid gaps in the ACK ranges and reduce the frame size. There is no obligation for a FEC receiver to send SYMBOL_ACK frames. The FEC receiver MAY decide to only advertize a subset of the received source symbols. The SYMBOL_ACK frame MUST NOT be used to infer any congestion state on the network (see ).

Coding channel and congestion control The coding and network channels being unrelated, receiving symbols through the coding channel MUST NOT be used to infer any congestion state on the network channel. More specifically, receiving a symbol through the coding channel MUST NOT be used to hide the network loss event of the corresponding packet to the congestion control, applying Recommandation 1 of .

Negociating the FEC extension using transport parameters This section defines the new transport parameters used to negociate and parametrize the FEC extension described in this document.

enable_fec The use of the FEC extension is negociated using the enable_fec transport parameter defined in : Values for enable_fec

Option	Definition
0x0	don't support FEC
0x1	supports FEC as defined in this document

When the enable_fec value is 0 or is not advertized by the peer, the QUIC endpoint MUST NOT use any frame or mechanism described in this document.

decoder_fec_scheme Each QUIC endpoint uses the decoder_fec_scheme transport parameter to define the FEC scheme used to decode the received repair symbols. The QUIC sender MUST use the specified FEC scheme to generate repair symbols. The decoder_fec_scheme parameter is an integer value representing the identifier of the desired FEC scheme. For instance, a FEC scheme using Reed Solomon could be identified by the ID 0x0 and a FEC scheme using LDPC could be identified by 0x1. This document does not specify nor identify any FEC scheme yet. Several FEC schemes have been proposed . The next version of this document will detail how some of these schemes can be directly integrated in QUIC. Future versions of this document will provide ways to format FEC scheme-specific payload for REPAIR frames. When the decoder_fec_scheme parameter is not advertized by the peer, the QUIC sender MUST NOT send any repair symbol.

initial_coding_window Each QUIC endpoint uses the initial_coding_window transport parameter to define the initial coding window size it uses to store source and repair symbols (see ). When the initial_coding_window parameter is not advertized by the peer, the QUIC sender MUST consider a default value of 0 and MUST NOT send any repair symbol.

Security Considerations The FEC mechanism for QUIC only runs under 0-RTT and 1-RTT encryption levels and only operates inside the encrypted payload.

DoS due to difficult symbols recoveries An attacker could try to cause a DoS of a receiver by selectively sending source and repair symbols to trigger intensive erasure correction operations on the receiver. A QUIC receiver is never forced to perform any erasure correction and may ignore any received repair symbol if it has doubts in its capabilities to decode it in a reasonable amount of time.

IANA Considerations Disclaimer: the IDs defined in this section are present for experimental purposes only. They are not requested codepoints and are subject to change in the next versions of this document. This document defines three new transport parameters and five new frames. The SID and SOURCE_SYMBOL frames serve the same purpose. One of them will be removed in next versions of this document. The values present in the tables are used for experiments.

New transport parameters New transport parameters

Parameter ID	Parameter name	Specification
0x238ffeceXX	enable_fec
0x238ffecd	decoder_fec_scheme
0x238ffecc	initial_coding_window

The XX in 0x238ffeceXX are to be replaced by the version of this document that is implemented by the QUIC endpoint (e.g. the parameter ID for the version 00 of this document is 0x238ffece00).

New frames New frames

Frame ID	Frame name	Specification
0x32a80fec	REPAIR
0x32a80fec55	SOURCE_SYMBOL
0x32a80fec1d	SID
0x32a80fecac	SYMBOL_ACK
0x32a80fecc0	FEC_WINDOW

Acknowledgments Maxime Piraux, Olivier Bonaventure and all the authors of .