Raise bucket weights to the power four in the historical model

Utilizing the results of probes sent once a minute to a random node
in the network for a random amount (within a reasonable range), we
were able to analyze the accuracy of our resulting success
probability estimation with various PDFs across the historical and
live-bounds models.

For each candidate PDF (as well as other parameters, including the
histogram bucket weight), we used the
`min_zero_implies_no_successes` fudge factor in
`success_probability` as well as a total probability multiple fudge
factor to get both the historical success model and the a priori
model to be neither too optimistic nor too pessimistic (as measured
by the relative log-loss between succeeding and failing hops in our
sample data).

We then compared the resulting log-loss for the historical success
model and selected the candidate PDF with the lowest log-loss,
skipping a few candidates with similar resulting log-loss but with
more extreme constants (such as a power of 11 with a higher
`min_zero_implies_no_successes` penalty).

Somewhat surprisingly (to me at least), the (fairly strongly)
preferred model was one where the bucket weights in the historical
histograms are exponentiated. In the current design, the weights
are effectively squared as we multiply the minimum- and maximum-
histogram buckets together before adding the weight*probabilities
together.

Here we multiply the weights yet again before addition. While the
simulation runs seemed to prefer a slightly stronger weight than
the 4th power we do here, the difference wasn't substantial
(log-loss 0.5058 to 0.4941), so we do the simpler single extra
multiply here.

Note that if we did this naively we'd run out of bits in our
arithmetic operations - we have 16-bit buckets, which when raised
to the 4th can fully fill a 64-bit int. Additionally, when looking
at the 0th min-bucket we occasionally add up to 32 weights together
before multiplying by the probability, requiring an additional five
bits.

Instead, we move to using floats during our histogram walks, which
further avoids some float -> int conversions because it allows for
retaining the floats we're already using to calculate probability.

Across the last handful of commits, the increased pessimism more
than makes up for the increased runtime complexity, leading to a
40-45% pathfinding speedup on a Xeon Silver 4116 and a 25-45%
speedup on a Xeon E5-2687W v3.

Author: TheBlueMatt

lightning/src/routing/scoring.rs

@@ -1218,14 +1218,33 @@ fn nonlinear_success_probability(
 /// Given liquidity bounds, calculates the success probability (in the form of a numerator and
 /// denominator) of an HTLC. This is a key assumption in our scoring models.
 ///
-/// Must not return a numerator or denominator greater than 2^31 for arguments less than 2^31.
-///
 /// `total_inflight_amount_msat` includes the amount of the HTLC and any HTLCs in flight over the
 /// channel.
 ///
 /// min_zero_implies_no_successes signals that a `min_liquidity_msat` of 0 means we've not
 /// (recently) seen an HTLC successfully complete over this channel.
 #[inline(always)]
+fn success_probability_float(
+	total_inflight_amount_msat: u64, min_liquidity_msat: u64, max_liquidity_msat: u64,
+	capacity_msat: u64, params: &ProbabilisticScoringFeeParameters,
+	min_zero_implies_no_successes: bool,
+) -> (f64, f64) {
+	debug_assert!(min_liquidity_msat <= total_inflight_amount_msat);
+	debug_assert!(total_inflight_amount_msat < max_liquidity_msat);
+	debug_assert!(max_liquidity_msat <= capacity_msat);
+
+	if params.linear_success_probability {
+		let (numerator, denominator) = linear_success_probability(total_inflight_amount_msat, min_liquidity_msat, max_liquidity_msat, min_zero_implies_no_successes);
+		(numerator as f64, denominator as f64)
+	} else {
+		nonlinear_success_probability(total_inflight_amount_msat, min_liquidity_msat, max_liquidity_msat, capacity_msat, min_zero_implies_no_successes)
+	}
+}
+
+#[inline(always)]
+/// Identical to [`success_probability_float`] but returns integer numerator and denominators.
+///
+/// Must not return a numerator or denominator greater than 2^31 for arguments less than 2^31.
 fn success_probability(
 	total_inflight_amount_msat: u64, min_liquidity_msat: u64, max_liquidity_msat: u64,
 	capacity_msat: u64, params: &ProbabilisticScoringFeeParameters,
@@ -1798,7 +1817,7 @@ mod bucketed_history {
 		// Because the first thing we do is check if `total_valid_points` is sufficient to consider
 		// the data here at all, and can return early if it is not, we want this to go first to
 		// avoid hitting a second cache line load entirely in that case.
-		total_valid_points_tracked: u64,
+		total_valid_points_tracked: f64,
 		min_liquidity_offset_history: HistoricalBucketRangeTracker,
 		max_liquidity_offset_history: HistoricalBucketRangeTracker,
 	}
@@ -1808,7 +1827,7 @@ mod bucketed_history {
 			HistoricalLiquidityTracker {
 				min_liquidity_offset_history: HistoricalBucketRangeTracker::new(),
 				max_liquidity_offset_history: HistoricalBucketRangeTracker::new(),
-				total_valid_points_tracked: 0,
+				total_valid_points_tracked: 0.0,
 			}
 		}
 
@@ -1819,7 +1838,7 @@ mod bucketed_history {
 			let mut res = HistoricalLiquidityTracker {
 				min_liquidity_offset_history,
 				max_liquidity_offset_history,
-				total_valid_points_tracked: 0,
+				total_valid_points_tracked: 0.0,
 			};
 			res.recalculate_valid_point_count();
 			res
@@ -1842,12 +1861,15 @@ mod bucketed_history {
 		}
 
 		fn recalculate_valid_point_count(&mut self) {
-			self.total_valid_points_tracked = 0;
+			let mut total_valid_points_tracked = 0;
 			for (min_idx, min_bucket) in self.min_liquidity_offset_history.buckets.iter().enumerate() {
 				for max_bucket in self.max_liquidity_offset_history.buckets.iter().take(32 - min_idx) {
-					self.total_valid_points_tracked += (*min_bucket as u64) * (*max_bucket as u64);
+					let mut bucket_weight = (*min_bucket as u64) * (*max_bucket as u64);
+					bucket_weight *= bucket_weight;
+					total_valid_points_tracked += bucket_weight;
 				}
 			}
+			self.total_valid_points_tracked = total_valid_points_tracked as f64;
 		}
 
 		pub(super) fn writeable_min_offset_history(&self) -> &HistoricalBucketRangeTracker {
@@ -1933,20 +1955,23 @@ mod bucketed_history {
 				let mut actual_valid_points_tracked = 0;
 				for (min_idx, min_bucket) in min_liquidity_offset_history_buckets.iter().enumerate() {
 					for max_bucket in max_liquidity_offset_history_buckets.iter().take(32 - min_idx) {
-						actual_valid_points_tracked += (*min_bucket as u64) * (*max_bucket as u64);
+						let mut bucket_weight = (*min_bucket as u64) * (*max_bucket as u64);
+						bucket_weight *= bucket_weight;
+						actual_valid_points_tracked += bucket_weight;
 					}
 				}
-				assert_eq!(total_valid_points_tracked, actual_valid_points_tracked);
+				assert_eq!(total_valid_points_tracked, actual_valid_points_tracked as f64);
 			}
 
 			// If the total valid points is smaller than 1.0 (i.e. 32 in our fixed-point scheme),
 			// treat it as if we were fully decayed.
-			const FULLY_DECAYED: u16 = BUCKET_FIXED_POINT_ONE * BUCKET_FIXED_POINT_ONE;
+			const FULLY_DECAYED: f64 = BUCKET_FIXED_POINT_ONE as f64 * BUCKET_FIXED_POINT_ONE as f64 *
+				BUCKET_FIXED_POINT_ONE as f64 * BUCKET_FIXED_POINT_ONE as f64;
 			if total_valid_points_tracked < FULLY_DECAYED.into() {
 				return None;
 			}
 
-			let mut cumulative_success_prob_times_billion = 0;
+			let mut cumulative_success_prob = 0.0f64;
 			// Special-case the 0th min bucket - it generally means we failed a payment, so only
 			// consider the highest (i.e. largest-offset-from-max-capacity) max bucket for all
 			// points against the 0th min bucket. This avoids the case where we fail to route
@@ -1959,16 +1984,18 @@ mod bucketed_history {
 				// max-bucket with at least BUCKET_FIXED_POINT_ONE.
 				let mut highest_max_bucket_with_points = 0;
 				let mut highest_max_bucket_with_full_points = None;
-				let mut total_max_points = 0; // Total points in max-buckets to consider
+				let mut total_weight = 0;
 				for (max_idx, max_bucket) in max_liquidity_offset_history_buckets.iter().enumerate() {
 					if *max_bucket >= BUCKET_FIXED_POINT_ONE {
 						highest_max_bucket_with_full_points = Some(cmp::max(highest_max_bucket_with_full_points.unwrap_or(0), max_idx));
 					}
 					if *max_bucket != 0 {
 						highest_max_bucket_with_points = cmp::max(highest_max_bucket_with_points, max_idx);
 					}
-					total_max_points += *max_bucket as u64;
+					total_weight += (*max_bucket as u64) * (*max_bucket as u64)
+						* (min_liquidity_offset_history_buckets[0] as u64) * (min_liquidity_offset_history_buckets[0] as u64);
 				}
+				debug_assert!(total_weight as f64 <= total_valid_points_tracked);
 				// Use the highest max-bucket with at least BUCKET_FIXED_POINT_ONE, but if none is
 				// available use the highest max-bucket with any non-zero value. This ensures that
 				// if we have substantially decayed data we don't end up thinking the highest
@@ -1977,40 +2004,39 @@ mod bucketed_history {
 				let selected_max = highest_max_bucket_with_full_points.unwrap_or(highest_max_bucket_with_points);
 				let max_bucket_end_pos = BUCKET_START_POS[32 - selected_max] - 1;
 				if payment_pos < max_bucket_end_pos {
-					let (numerator, denominator) = success_probability(payment_pos as u64, 0,
+					let (numerator, denominator) = success_probability_float(payment_pos as u64, 0,
 						max_bucket_end_pos as u64, POSITION_TICKS as u64 - 1, params, true);
-					let bucket_prob_times_billion =
-						(min_liquidity_offset_history_buckets[0] as u64) * total_max_points
-							* 1024 * 1024 * 1024 / total_valid_points_tracked;
-					cumulative_success_prob_times_billion += bucket_prob_times_billion *
-						numerator / denominator;
+					let bucket_prob = total_weight as f64 / total_valid_points_tracked;
+					cumulative_success_prob += bucket_prob * numerator / denominator;
 				}
 			}
 
 			for (min_idx, min_bucket) in min_liquidity_offset_history_buckets.iter().enumerate().skip(1) {
 				let min_bucket_start_pos = BUCKET_START_POS[min_idx];
 				for (max_idx, max_bucket) in max_liquidity_offset_history_buckets.iter().enumerate().take(32 - min_idx) {
 					let max_bucket_end_pos = BUCKET_START_POS[32 - max_idx] - 1;
-					// Note that this multiply can only barely not overflow - two 16 bit ints plus
-					// 30 bits is 62 bits.
-					let bucket_prob_times_billion = (*min_bucket as u64) * (*max_bucket as u64)
-						* 1024 * 1024 * 1024 / total_valid_points_tracked;
+					let mut bucket_weight = (*min_bucket as u64) * (*max_bucket as u64);
+					bucket_weight *= bucket_weight;
+					debug_assert!(bucket_weight as f64 <= total_valid_points_tracked);
+
 					if payment_pos >= max_bucket_end_pos {
 						// Success probability 0, the payment amount may be above the max liquidity
 						break;
-					} else if payment_pos < min_bucket_start_pos {
-						cumulative_success_prob_times_billion += bucket_prob_times_billion;
+					}
+
+					let bucket_prob = bucket_weight as f64 / total_valid_points_tracked;
+					if payment_pos < min_bucket_start_pos {
+						cumulative_success_prob += bucket_prob;
 					} else {
-						let (numerator, denominator) = success_probability(payment_pos as u64,
+						let (numerator, denominator) = success_probability_float(payment_pos as u64,
 							min_bucket_start_pos as u64, max_bucket_end_pos as u64,
 							POSITION_TICKS as u64 - 1, params, true);
-						cumulative_success_prob_times_billion += bucket_prob_times_billion *
-							numerator / denominator;
+						cumulative_success_prob += bucket_prob * numerator / denominator;
 					}
 				}
 			}
 
-			Some(cumulative_success_prob_times_billion)
+			Some((cumulative_success_prob * (1024.0 * 1024.0 * 1024.0)) as u64)
 		}
 	}
 }