|
| 1 | +# gossipsub v1.4: Message preamble to limit duplicate transmissions of large messages |
| 2 | + |
| 3 | +| Lifecycle Stage | Maturity | Status | Latest Revision | |
| 4 | +|-----------------|---------------------------|--------|-----------------| |
| 5 | +| 1A | Working Draft | Active | r0, 2024-12-17 | |
| 6 | + |
| 7 | +Authors: [@ufarooqstatus], [@kaiserd] |
| 8 | + |
| 9 | +Interest Group: TBD |
| 10 | + |
| 11 | +[@kaiserd]: https://github.com/kaiserd |
| 12 | +[@ufarooqstatus]: https://github.com/ufarooqstatus |
| 13 | + |
| 14 | +See the [lifecycle document][lifecycle-spec] for context about maturity level and spec status. |
| 15 | + |
| 16 | +[lifecycle-spec]: https://github.com/libp2p/specs/blob/master/00-framework-01-spec-lifecycle.md |
| 17 | + |
| 18 | +# Overview |
| 19 | + |
| 20 | +This document outlines small extensions to the [gossipsub |
| 21 | +v1.2](https://github.com/libp2p/specs/blob/master/pubsub/gossipsub/gossipsub-v1.2.md) |
| 22 | +protocol to improve performance when handling large message transmissions. |
| 23 | +The extensions are optional, fully compatible with v1.2 of the protocol, |
| 24 | +and involve minor modifications to peer behavior when receiving or forwarding large messages. |
| 25 | + |
| 26 | +The proposed modifications address the issue that the number of IWANT requests and duplicates increases significantly with the message size. |
| 27 | +This happens because sending a large message can take considerable time, |
| 28 | +and during this time, the receiver is unaware of the IDs of the messages it is receiving. |
| 29 | + |
| 30 | +Under the current arrangement, if IHAVE announcements are received for a message that is already being received, |
| 31 | +the receiver may generate multiple IWANT requests, triggering unnecessary retransmissions of the same message. |
| 32 | +Higher message reception time also increases the probability of simultaneously receiving the same message from many peers. |
| 33 | + |
| 34 | +Prepending a preamble while sending a large message can allow receivers |
| 35 | +to immediately gain insights into the IDs and lengths of the messages they are receiving. |
| 36 | +This can allow receivers to defer IWANT requests for messages they are currently receiving, |
| 37 | +thereby reducing the number of unnecessary IWANT requests. |
| 38 | + |
| 39 | +At the same time, receivers may inform their mesh members about ongoing message reception. |
| 40 | +For this purpose, a new control message (IMReceiving) is introduced, |
| 41 | +indicating that the peer sending IMReceiving message is currently receiving messages identified by their message IDs, |
| 42 | +and sending duplicates might be unnecessary. |
| 43 | + |
| 44 | +On receiving an IMReceiving message, a peer should defer sending of messages identified by the message IDs in IMReceiving message. |
| 45 | +This can lead to a significant reduction in bandwidth utilization and message latency. |
| 46 | + |
| 47 | +# Specification |
| 48 | + |
| 49 | +## Protocol Id |
| 50 | + |
| 51 | +Nodes that support this Gossipsub extension should additionally advertise the |
| 52 | +version number `1.4.0`. Gossipsub nodes can advertise their own protocol-id |
| 53 | +prefix, by default this is `meshsub` giving the default protocol id: |
| 54 | +- `/meshsub/1.4.0` |
| 55 | + |
| 56 | +## Parameters |
| 57 | + |
| 58 | +This section lists the configuration parameters that clients need to agree on to avoid peer penalizations. |
| 59 | + |
| 60 | +| Parameter | Description | Reasonable Default | |
| 61 | +|--------------------------|------------------------------------------------------------------|--------------| |
| 62 | +| `peer_preamble_announcements` | The maximum number of preamble announcements for unfinished transfers per peer | 1??? | |
| 63 | +| `mesh_preamble_announcements` | The maximum number of preamble announcements to accept for unfinished transfers per topic per heartbeat interval | 3??? | |
| 64 | +| `max_iwant_requests` | The maximum number of simultaneous IWANT requests for a message | 1??? | |
| 65 | +| `preamble_threshold` | The smallest message size to use message preamble | 200KB??? | |
| 66 | + |
| 67 | + |
| 68 | + |
| 69 | +## Message Preamble |
| 70 | + |
| 71 | +### Basic scenario |
| 72 | + |
| 73 | +When a peer starts relaying a message that exceeds the preamble_threshold size, |
| 74 | +it may include a control message (called message preamble) at the start of the message. |
| 75 | + |
| 76 | +The purpose of the preamble is to allow receivers to instantly learn about the incoming message. |
| 77 | +The preamble must include the message ID and length, |
| 78 | +providing receivers with immediate access to critical information about the incoming message. |
| 79 | +The receiver must immediately process a message preamble without waiting to download the entire message. |
| 80 | + |
| 81 | +On receiving a preamble that advertises a message ID not present in the seen cache, |
| 82 | +a receiver should add that message ID to the ongoing_receives list. |
| 83 | +The ongoing_receives list is crucial in limiting IWANT requests and simultaneous reception of the same message from mesh peers. |
| 84 | + |
| 85 | +The receiver may use message length to leniently estimate message download time. |
| 86 | +If the message is successfully downloaded before the estimated download time has elapsed, |
| 87 | +the message ID is removed from the ongoing_receives list and added to the seen cache. |
| 88 | + |
| 89 | +If the download takes longer than the estimated download time, |
| 90 | +the sender may be penalized through P4 to discourage peers from intentionally delaying message transmission. |
| 91 | + |
| 92 | +The message preamble is considered _optional_ for both the sender and the receiver. |
| 93 | +This means that the sender may choose not to prepend the preamble, |
| 94 | +and the receiver may also opt to ignore it. |
| 95 | +Adding a message preamble may increase control overhead for small messages. |
| 96 | +Therefore, it is preferable to use it only for messages that exceed the preamble_threshold. |
| 97 | + |
| 98 | + |
| 99 | +### Limiting IWANT Requests |
| 100 | + |
| 101 | +When a peer receives an IHAVE announcement for a message ID not present in the seen cache, |
| 102 | +the peer must also check the ongoing_receives list before making an IWANT request. |
| 103 | + |
| 104 | +If the message ID is found in the ongoing_receives list, |
| 105 | +the peer should postpone sending the IWANT request for a defer_interval. |
| 106 | +The defer_interval may be based on the message download time. |
| 107 | + |
| 108 | +If the message download completes before the defer_interval expires, |
| 109 | +the IWANT request for that message will not be generated. |
| 110 | +However, if the defer_interval elapses and the download has not completed, |
| 111 | +the peer can proceed to make the IWANT request for the missing message. |
| 112 | + |
| 113 | +The total number of outstanding IWANT requests for a single message should not exceed max_iwant_requests. |
| 114 | +Every peer must respond to the received IWANT requests. |
| 115 | +Failing to do so may result in behavioural penalty through P7. |
| 116 | +This will discourage peers from intentionally not replying to IWANT requests. |
| 117 | + |
| 118 | +### IMReceiving Message |
| 119 | + |
| 120 | +The IMReceiving message serves a distinct purpose compared to the IDONTWANT message. |
| 121 | +An IDONTWANT can only be transmitted after receiving the entire message. |
| 122 | +In contrast, an IMReceiving should be transmitted immediately after receiving a preamble indicating a large message transfer. |
| 123 | +The IMReceiving message requests peers to refrain from resending a large message that is already being received. |
| 124 | + |
| 125 | +When a peer receives a message preamble indicating a message ID that is not present in the seen cache, |
| 126 | +it should send an IMReceiving message to the nodes in its full message mesh. |
| 127 | +On receiving an IMReceiving from a peer, |
| 128 | +a node should postpone sending the messages indicated by the IMReceiving to that peer. |
| 129 | +The defer_interval can be leniently estimated based on the message length. |
| 130 | + |
| 131 | +An IDONTWANT from the peer will indicate successful reception of the message. |
| 132 | +If an IDONTWANT for the deferred message is not received and the defer_interval has elapsed, |
| 133 | +the node may proceed to transmit the message. |
| 134 | + |
| 135 | + |
| 136 | +## Protobuf Extension |
| 137 | + |
| 138 | +The protobuf messages are identical to those specified in the [gossipsub v1.2 |
| 139 | +specification](https://github.com/libp2p/specs/blob/master/pubsub/gossipsub/gossipsub-v1.2.md) |
| 140 | +with the following control message modifications: |
| 141 | + |
| 142 | +```protobuf |
| 143 | +message RPC { |
| 144 | + // ... see definition in the gossipsub specification |
| 145 | +} |
| 146 | +
|
| 147 | +message ControlMessage { |
| 148 | + // ... see definition in the gossipsub specification |
| 149 | + repeated ControlPreamble preamble = 6; |
| 150 | + repeated ControlIMReceiving imreceiving = 7; |
| 151 | +} |
| 152 | +
|
| 153 | +message ControlPreamble { |
| 154 | + optional bytes messageID = 1; |
| 155 | + optional int32 messageLength = 2; |
| 156 | +} |
| 157 | +
|
| 158 | +message ControlIMReceiving { |
| 159 | + optional bytes messageID = 1; |
| 160 | + optional int32 messageLength = 2; |
| 161 | +} |
| 162 | +
|
| 163 | +``` |
| 164 | + |
0 commit comments