Merge pull request #2797 from halseth/autopilot-prefattach-small-chan-penalize

[autopilot] penalize small channels in preferantial attachment heuristic

autopilot/graph.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"math/big"
 	"net"
+	"sort"
 	"sync/atomic"
 	"time"
 
@@ -501,3 +502,22 @@ func (m memNode) ForEachChannel(cb func(ChannelEdge) error) error {
 
 	return nil
 }
+
+// Median returns the median value in the slice of Amounts.
+func Median(vals []btcutil.Amount) btcutil.Amount {
+	sort.Slice(vals, func(i, j int) bool {
+		return vals[i] < vals[j]
+	})
+
+	num := len(vals)
+	switch {
+	case num == 0:
+		return 0
+
+	case num%2 == 0:
+		return (vals[num/2-1] + vals[num/2]) / 2
+
+	default:
+		return vals[num/2]
+	}
+}

autopilot/graph_test.go

@@ -0,0 +1,50 @@
+package autopilot_test
+
+import (
+	"testing"
+
+	"github.com/btcsuite/btcutil"
+	"github.com/lightningnetwork/lnd/autopilot"
+)
+
+// TestMedian tests the Median method.
+func TestMedian(t *testing.T) {
+	t.Parallel()
+
+	testCases := []struct {
+		values []btcutil.Amount
+		median btcutil.Amount
+	}{
+		{
+			values: []btcutil.Amount{},
+			median: 0,
+		},
+		{
+			values: []btcutil.Amount{10},
+			median: 10,
+		},
+		{
+			values: []btcutil.Amount{10, 20},
+			median: 15,
+		},
+		{
+			values: []btcutil.Amount{10, 20, 30},
+			median: 20,
+		},
+		{
+			values: []btcutil.Amount{30, 10, 20},
+			median: 20,
+		},
+		{
+			values: []btcutil.Amount{10, 10, 10, 10, 5000000},
+			median: 10,
+		},
+	}
+
+	for _, test := range testCases {
+		res := autopilot.Median(test.values)
+		if res != test.median {
+			t.Fatalf("expected median %v, got %v", test.median, res)
+		}
+	}
+}

autopilot/prefattach.go

@@ -8,6 +8,12 @@ import (
 	"github.com/btcsuite/btcutil"
 )
 
+// minMedianChanSizeFraction determines the minimum size a channel must have to
+// count positively when calculating the scores using preferential attachment.
+// The minimum channel size is calculated as median/minMedianChanSizeFraction,
+// where median is the median channel size of the entire graph.
+const minMedianChanSizeFraction = 4
+
 // PrefAttachment is an implementation of the AttachmentHeuristic interface
 // that implement a non-linear preferential attachment heuristic. This means
 // that given a threshold to allocate to automatic channel establishment, the
@@ -64,6 +70,10 @@ func (p *PrefAttachment) Name() string {
 // implemented globally for each new participant, this results in a channel
 // graph that is scale-free and follows a power law distribution with k=-3.
 //
+// To avoid assigning a high score to nodes with a large number of small
+// channels, we only count channels at least as large as a given fraction of
+// the graph's median channel size.
+//
 // The returned scores will be in the range [0.0, 1.0], where higher scores are
 // given to nodes already having high connectivity in the graph.
 //
@@ -72,12 +82,50 @@ func (p *PrefAttachment) NodeScores(g ChannelGraph, chans []Channel,
 	chanSize btcutil.Amount, nodes map[NodeID]struct{}) (
 	map[NodeID]*NodeScore, error) {
 
-	// Count the number of channels for each particular node in the graph.
+	// We first run though the graph once in order to find the median
+	// channel size.
+	var (
+		allChans  []btcutil.Amount
+		seenChans = make(map[uint64]struct{})
+	)
+	if err := g.ForEachNode(func(n Node) error {
+		err := n.ForEachChannel(func(e ChannelEdge) error {
+			if _, ok := seenChans[e.ChanID.ToUint64()]; ok {
+				return nil
+			}
+			seenChans[e.ChanID.ToUint64()] = struct{}{}
+			allChans = append(allChans, e.Capacity)
+			return nil
+		})
+		if err != nil {
+			return err
+		}
+
+		return nil
+	}); err != nil {
+		return nil, err
+	}
+
+	medianChanSize := Median(allChans)
+
+	// Count the number of large-ish channels for each particular node in
+	// the graph.
 	var maxChans int
 	nodeChanNum := make(map[NodeID]int)
 	if err := g.ForEachNode(func(n Node) error {
 		var nodeChans int
-		err := n.ForEachChannel(func(_ ChannelEdge) error {
+		err := n.ForEachChannel(func(e ChannelEdge) error {
+			// Since connecting to nodes with a lot of small
+			// channels actually worsens our connectivity in the
+			// graph (we will potentially waste time trying to use
+			// these useless channels in path finding), we decrease
+			// the counter for such channels.
+			if e.Capacity < medianChanSize/minMedianChanSizeFraction {
+				nodeChans--
+				return nil
+			}
+
+			// Larger channels we count.
 			nodeChans++
 			return nil
 		})
@@ -132,9 +180,9 @@ func (p *PrefAttachment) NodeScores(g ChannelGraph, chans []Channel,
 		case ok:
 			continue
 
-		// If the node had no channels, we skip it, since it would have
-		// gotten a zero score anyway.
-		case nodeChans == 0:
+		// If the node had no large channels, we skip it, since it
+		// would have gotten a zero score anyway.
+		case nodeChans <= 0:
 			continue
 		}
 

lnrpc/rpc.proto

@@ -1738,6 +1738,7 @@ message NetworkInfo {
     double avg_channel_size = 7 [json_name = "avg_channel_size"];
     int64 min_channel_size = 8 [json_name = "min_channel_size"];
     int64 max_channel_size = 9 [json_name = "max_channel_size"];
+    int64 median_channel_size_sat = 10 [json_name = "median_channel_size_sat"];
 
     // TODO(roasbeef): fee rate info, expiry
     //  * also additional RPC for tracking fee info once in

lnrpc/rpc.swagger.json

@@ -2400,6 +2400,10 @@
         "max_channel_size": {
           "type": "string",
           "format": "int64"
+        },
+        "median_channel_size_sat": {
+          "type": "string",
+          "format": "int64"
         }
       }
     },

rpcserver.go

@@ -3954,6 +3954,7 @@ func (r *rpcServer) GetNetworkInfo(ctx context.Context,
 		totalNetworkCapacity btcutil.Amount
 		minChannelSize       btcutil.Amount = math.MaxInt64
 		maxChannelSize       btcutil.Amount
+		medianChanSize       btcutil.Amount
 	)
 
 	// We'll use this map to de-duplicate channels during our traversal.
@@ -3961,6 +3962,10 @@ func (r *rpcServer) GetNetworkInfo(ctx context.Context,
 	// edges for each channel within the graph.
 	seenChans := make(map[uint64]struct{})
 
+	// We also keep a list of all encountered capacities, in order to
+	// calculate the median channel size.
+	var allChans []btcutil.Amount
+
 	// We'll run through all the known nodes in the within our view of the
 	// network, tallying up the total number of nodes, and also gathering
 	// each node so we can measure the graph diameter and degree stats
@@ -4007,6 +4012,7 @@ func (r *rpcServer) GetNetworkInfo(ctx context.Context,
 			numChannels++
 
 			seenChans[edge.ChannelID] = struct{}{}
+			allChans = append(allChans, edge.Capacity)
 			return nil
 		}); err != nil {
 			return err
@@ -4023,6 +4029,9 @@ func (r *rpcServer) GetNetworkInfo(ctx context.Context,
 		return nil, err
 	}
 
+	// Find the median.
+	medianChanSize = autopilot.Median(allChans)
+
 	// If we don't have any channels, then reset the minChannelSize to zero
 	// to avoid outputting NaN in encoded JSON.
 	if numChannels == 0 {
@@ -4032,7 +4041,6 @@ func (r *rpcServer) GetNetworkInfo(ctx context.Context,
 	// TODO(roasbeef): graph diameter
 
 	// TODO(roasbeef): also add oldest channel?
-	//  * also add median channel size
 	netInfo := &lnrpc.NetworkInfo{
 		MaxOutDegree:         maxChanOut,
 		AvgOutDegree:         float64(numChannels) / float64(numNodes),
@@ -4041,8 +4049,9 @@ func (r *rpcServer) GetNetworkInfo(ctx context.Context,
 		TotalNetworkCapacity: int64(totalNetworkCapacity),
 		AvgChannelSize:       float64(totalNetworkCapacity) / float64(numChannels),
 
-		MinChannelSize: int64(minChannelSize),
-		MaxChannelSize: int64(maxChannelSize),
+		MinChannelSize:       int64(minChannelSize),
+		MaxChannelSize:       int64(maxChannelSize),
+		MedianChannelSizeSat: int64(medianChanSize),
 	}
 
 	// Similarly, if we don't have any channels, then we'll also set the