]> git.bitcoin.ninja Git - rust-lightning/commitdiff
Avoid needless MPP on multiple channels to the same first-hop 2022-03-pref-first-hop-chans
authorMatt Corallo <git@bluematt.me>
Fri, 18 Mar 2022 23:19:51 +0000 (23:19 +0000)
committerMatt Corallo <git@bluematt.me>
Wed, 23 Mar 2022 21:01:49 +0000 (21:01 +0000)
When we have many channels to the same first-hop, many of which do
not have sufficient balance to make the requested payment, but when
some do, instead of simply using the available channel balance we
may switch to MPP, potentially with many, many paths.

Instead, we should seek to use the smallest channel which can
easily handle the requested payment, which we do here by sorting
the first_hops in our router before beginning the graph search.

Note that the "real" fix for this should be to instead decide which
channel to use at HTLC-send time, as most other nodes do during
relay, but this provides a minimal fix without needing to do the
rather-large work of refactoring our HTLC send+relay pipelines.

Issues with overly-aggressive MPP on many channels were reported by
Cash App.

lightning/src/routing/router.rs

index 1b52af8815624e3c4e3e3d9a400577875cfa8869..f899f76199999bee3987b2773f7661a37bdef0f8 100644 (file)
@@ -811,6 +811,29 @@ where L::Target: Logger {
        // - when we want to stop looking for new paths.
        let mut already_collected_value_msat = 0;
 
+       for (_, channels) in first_hop_targets.iter_mut() {
+               // Sort the first_hops channels to the same node(s) in priority order of which channel we'd
+               // most like to use.
+               //
+               // First, if channels are below `recommended_value_msat`, sort them in descending order,
+               // preferring larger channels to avoid splitting the payment into more MPP parts than is
+               // required.
+               //
+               // Second, because simply always sorting in descending order would always use our largest
+               // available outbound capacity, needlessly fragmenting our available channel capacities,
+               // sort channels above `recommended_value_msat` in ascending order, preferring channels
+               // which have enough, but not too much, capacity for the payment.
+               channels.sort_unstable_by(|chan_a, chan_b| {
+                       if chan_b.outbound_capacity_msat < recommended_value_msat || chan_a.outbound_capacity_msat < recommended_value_msat {
+                               // Sort in descending order
+                               chan_b.outbound_capacity_msat.cmp(&chan_a.outbound_capacity_msat)
+                       } else {
+                               // Sort in ascending order
+                               chan_a.outbound_capacity_msat.cmp(&chan_b.outbound_capacity_msat)
+                       }
+               });
+       }
+
        log_trace!(logger, "Building path from {} (payee) to {} (us/payer) for value {} msat.", payment_params.payee_pubkey, our_node_pubkey, final_value_msat);
 
        macro_rules! add_entry {
@@ -4894,6 +4917,32 @@ mod tests {
                        assert_eq!(route.paths[1][0].short_channel_id, 2);
                        assert_eq!(route.paths[1][0].fee_msat, 50_000);
                }
+
+               {
+                       // If we have a bunch of outbound channels to the same node, where most are not
+                       // sufficient to pay the full payment, but one is, we should default to just using the
+                       // one single channel that has sufficient balance, avoiding MPP.
+                       //
+                       // If we have several options above the 3xpayment value threshold, we should pick the
+                       // smallest of them, avoiding further fragmenting our available outbound balance to
+                       // this node.
+                       let route = get_route(&our_id, &payment_params, &network_graph.read_only(), Some(&[
+                               &get_channel_details(Some(2), nodes[0], InitFeatures::known(), 50_000),
+                               &get_channel_details(Some(3), nodes[0], InitFeatures::known(), 50_000),
+                               &get_channel_details(Some(5), nodes[0], InitFeatures::known(), 50_000),
+                               &get_channel_details(Some(6), nodes[0], InitFeatures::known(), 300_000),
+                               &get_channel_details(Some(7), nodes[0], InitFeatures::known(), 50_000),
+                               &get_channel_details(Some(8), nodes[0], InitFeatures::known(), 50_000),
+                               &get_channel_details(Some(9), nodes[0], InitFeatures::known(), 50_000),
+                               &get_channel_details(Some(4), nodes[0], InitFeatures::known(), 1_000_000),
+                       ]), 100_000, 42, Arc::clone(&logger), &scorer, &random_seed_bytes).unwrap();
+                       assert_eq!(route.paths.len(), 1);
+                       assert_eq!(route.paths[0].len(), 1);
+
+                       assert_eq!(route.paths[0][0].pubkey, nodes[0]);
+                       assert_eq!(route.paths[0][0].short_channel_id, 6);
+                       assert_eq!(route.paths[0][0].fee_msat, 100_000);
+               }
        }
 
        #[test]