From 77fa29320c5270005f1bc8b2a213e303da21047b Mon Sep 17 00:00:00 2001
From: William Chargin <wchargin@gmail.com>
Date: Tue, 24 Jul 2018 17:51:40 -0700
Subject: [PATCH] Reduce memory pressure by double-buffered PageRank (#520)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Summary:
This commit switches to a double-buffered PageRank implementation. When
benchmarked on `ipfs/js-ipfs`, the critical section improves from
3059 ms to 2433 ms (79.5% of original), and peak heap usage drops from
342 MB to 207 MB. (Tested non-rigorously in Chrome 67.)

Test Plan:
Existing unit tests for `sparseMarkovChainAction`,
`findStationaryDistribution`, and `pagerank` are sufficient.

wchargin-branch: pagerank-dbuf
---
 src/core/attribution/markovChain.js | 33 ++++++++++++++++++-----------
 1 file changed, 21 insertions(+), 12 deletions(-)

diff --git a/src/core/attribution/markovChain.js b/src/core/attribution/markovChain.js
index 779bca4..b02dae8 100644
--- a/src/core/attribution/markovChain.js
+++ b/src/core/attribution/markovChain.js
@@ -73,20 +73,28 @@ export function uniformDistribution(n: number): Distribution {
   return new Float64Array(n).fill(1 / n);
 }
 
-export function sparseMarkovChainAction(
+function sparseMarkovChainActionInto(
   chain: SparseMarkovChain,
-  pi: Distribution
-): Distribution {
-  const result = new Float64Array(pi.length);
+  input: Distribution,
+  output: Distribution
+): void {
   chain.forEach(({neighbor, weight}, dst) => {
     const inDegree = neighbor.length; // (also `weight.length`)
     let probability = 0;
     for (let i = 0; i < inDegree; i++) {
       const src = neighbor[i];
-      probability += pi[src] * weight[i];
+      probability += input[src] * weight[i];
     }
-    result[dst] = probability;
+    output[dst] = probability;
   });
+}
+
+export function sparseMarkovChainAction(
+  chain: SparseMarkovChain,
+  pi: Distribution
+): Distribution {
+  const result = new Float64Array(pi.length);
+  sparseMarkovChainActionInto(chain, pi, result);
   return result;
 }
 
@@ -98,7 +106,8 @@ function* findStationaryDistributionGenerator(
     +maxIterations: number,
   |}
 ): Generator<void, Distribution, void> {
-  let r0 = uniformDistribution(chain.length);
+  let pi = uniformDistribution(chain.length);
+  let scratch = new Float64Array(pi.length);
   function computeDelta(pi0, pi1) {
     let maxDelta = -Infinity;
     // Here, we assume that `pi0.nodeOrder` and `pi1.nodeOrder` are the
@@ -115,12 +124,12 @@ function* findStationaryDistributionGenerator(
       if (options.verbose) {
         console.log(`[${iteration}] FAILED to converge`);
       }
-      return r0;
+      return pi;
     }
     iteration++;
-    const r1 = sparseMarkovChainAction(chain, r0);
-    const delta = computeDelta(r0, r1);
-    r0 = r1;
+    sparseMarkovChainActionInto(chain, pi, scratch);
+    const delta = computeDelta(pi, scratch);
+    [scratch, pi] = [pi, scratch];
     if (options.verbose) {
       console.log(`[${iteration}] delta = ${delta}`);
     }
@@ -128,7 +137,7 @@ function* findStationaryDistributionGenerator(
       if (options.verbose) {
         console.log(`[${iteration}] CONVERGED`);
       }
-      return r0;
+      return pi;
     }
     yield;
   }