Cache buckets to speed up BytesRefHash#sort (#12784)

lucene/CHANGES.txt

@@ -133,6 +133,8 @@ Optimizations
 
 * GITHUB#12381: Skip docs with DocValues in NumericLeafComparator. (Lu Xugang, Adrien Grand)
 
+* GITHUB#12784: Cache buckets to speed up BytesRefHash#sort. (Guo Feng)
+
 Changes in runtime behavior
 ---------------------
 

lucene/core/src/java/org/apache/lucene/index/BufferedUpdates.java

@@ -263,11 +263,10 @@ <E extends Exception> void forEachOrdered(DeletedTermConsumer<E> consumer) throw
         scratch.field = deleteFieldEntry.getKey();
         BufferedUpdates.BytesRefIntMap terms = deleteFieldEntry.getValue();
         int[] indices = terms.bytesRefHash.sort();
-        for (int index : indices) {
-          if (index != -1) {
-            terms.bytesRefHash.get(index, scratch.bytes);
-            consumer.accept(scratch, terms.values[index]);
-          }
+        for (int i = 0; i < terms.bytesRefHash.size(); i++) {
+          int index = indices[i];
+          terms.bytesRefHash.get(index, scratch.bytes);
+          consumer.accept(scratch, terms.values[index]);
         }
       }
     }

lucene/core/src/java/org/apache/lucene/util/BytesRefHash.java

@@ -145,8 +145,63 @@ public int[] compact() {
    */
   public int[] sort() {
     final int[] compact = compact();
+    assert count * 2 <= compact.length : "We need load factor <= 0.5f to speed up this sort";
+    final int tmpOffset = count;
     new StringSorter(BytesRefComparator.NATURAL) {
 
+      @Override
+      protected Sorter radixSorter(BytesRefComparator cmp) {
+        return new MSBStringRadixSorter(cmp) {
+
+          private int k;
+
+          @Override
+          protected void buildHistogram(
+              int prefixCommonBucket,
+              int prefixCommonLen,
+              int from,
+              int to,
+              int k,
+              int[] histogram) {
+            this.k = k;
+            histogram[prefixCommonBucket] = prefixCommonLen;
+            Arrays.fill(
+                compact, tmpOffset + from - prefixCommonLen, tmpOffset + from, prefixCommonBucket);
+            for (int i = from; i < to; ++i) {
+              int b = getBucket(i, k);
+              compact[tmpOffset + i] = b;
+              histogram[b]++;
+            }
+          }
+
+          @Override
+          protected boolean shouldFallback(int from, int to, int l) {
+            // We lower the fallback threshold because the bucket cache speeds up the reorder
+            return to - from <= LENGTH_THRESHOLD / 2 || l >= LEVEL_THRESHOLD;
+          }
+
+          private void swapBucketCache(int i, int j) {
+            swap(i, j);
+            int tmp = compact[tmpOffset + i];
+            compact[tmpOffset + i] = compact[tmpOffset + j];
+            compact[tmpOffset + j] = tmp;
+          }
+
+          @Override
+          protected void reorder(int from, int to, int[] startOffsets, int[] endOffsets, int k) {
+            assert this.k == k;
+            for (int i = 0; i < HISTOGRAM_SIZE; ++i) {
+              final int limit = endOffsets[i];
+              for (int h1 = startOffsets[i]; h1 < limit; h1 = startOffsets[i]) {
+                final int b = compact[tmpOffset + from + h1];
+                final int h2 = startOffsets[b]++;
+                swapBucketCache(from + h1, from + h2);
+              }
+            }
+          }
+        };
+      }
+
       @Override
       protected void swap(int i, int j) {
         int tmp = compact[i];

lucene/core/src/java/org/apache/lucene/util/MSBRadixSorter.java

@@ -32,11 +32,11 @@ public abstract class MSBRadixSorter extends Sorter {
   // this is used as a protection against the fact that radix sort performs
   // worse when there are long common prefixes (probably because of cache
   // locality)
-  private static final int LEVEL_THRESHOLD = 8;
+  protected static final int LEVEL_THRESHOLD = 8;
   // size of histograms: 256 + 1 to indicate that the string is finished
   protected static final int HISTOGRAM_SIZE = 257;
-  // buckets below this size will be sorted with introsort
-  private static final int LENGTH_THRESHOLD = 100;
+  // buckets below this size will be sorted with fallback sorter
+  protected static final int LENGTH_THRESHOLD = 100;
 
   // we store one histogram per recursion level
   private final int[][] histograms = new int[LEVEL_THRESHOLD][];
@@ -130,15 +130,15 @@ public void sort(int from, int to) {
   }
 
   protected void sort(int from, int to, int k, int l) {
-    if (to - from <= LENGTH_THRESHOLD || l >= LEVEL_THRESHOLD) {
-      introSort(from, to, k);
+    if (shouldFallback(from, to, l)) {
+      getFallbackSorter(k).sort(from, to);
     } else {
       radixSort(from, to, k, l);
     }
   }
 
-  private void introSort(int from, int to, int k) {
-    getFallbackSorter(k).sort(from, to);
+  protected boolean shouldFallback(int from, int to, int l) {
+    return to - from <= LENGTH_THRESHOLD || l >= LEVEL_THRESHOLD;
   }
 
   /**
@@ -233,8 +233,6 @@ private int computeCommonPrefixLengthAndBuildHistogram(int from, int to, int k,
         if (b != commonPrefix[j]) {
           commonPrefixLength = j;
           if (commonPrefixLength == 0) { // we have no common prefix
-            histogram[commonPrefix[0] + 1] = i - from;
-            histogram[b + 1] = 1;
             break outer;
           }
           break;
@@ -245,7 +243,7 @@ private int computeCommonPrefixLengthAndBuildHistogram(int from, int to, int k,
     if (i < to) {
       // the loop got broken because there is no common prefix
       assert commonPrefixLength == 0;
-      buildHistogram(i + 1, to, k, histogram);
+      buildHistogram(commonPrefix[0] + 1, i - from, i, to, k, histogram);
     } else {
       assert commonPrefixLength > 0;
       histogram[commonPrefix[0] + 1] = to - from;
@@ -258,7 +256,9 @@ private int computeCommonPrefixLengthAndBuildHistogram(int from, int to, int k,
    * Build an histogram of the k-th characters of values occurring between offsets {@code from} and
    * {@code to}, using {@link #getBucket}.
    */
-  private void buildHistogram(int from, int to, int k, int[] histogram) {
+  protected void buildHistogram(
+      int prefixCommonBucket, int prefixCommonLen, int from, int to, int k, int[] histogram) {
+    histogram[prefixCommonBucket] = prefixCommonLen;
     for (int i = from; i < to; ++i) {
       histogram[getBucket(i, k)]++;
     }

lucene/core/src/java/org/apache/lucene/util/StableStringSorter.java

@@ -78,7 +78,9 @@ protected void restore(int i, int j) {
 
       @Override
       protected int compare(int i, int j) {
-        return StableStringSorter.this.compare(i, j);
+        get(scratch1, scratchBytes1, i);
+        get(scratch2, scratchBytes2, j);
+        return cmp.compare(scratchBytes1, scratchBytes2);
       }
 
       @Override

lucene/core/src/java/org/apache/lucene/util/StringSorter.java

@@ -58,24 +58,35 @@ public void sort(int from, int to) {
     }
   }
 
-  protected Sorter radixSorter(BytesRefComparator cmp) {
-    return new MSBRadixSorter(cmp.comparedBytesCount) {
-      @Override
-      protected void swap(int i, int j) {
-        StringSorter.this.swap(i, j);
-      }
+  /** A radix sorter for {@link BytesRef} */
+  protected class MSBStringRadixSorter extends MSBRadixSorter {
 
-      @Override
-      protected int byteAt(int i, int k) {
-        get(scratch1, scratchBytes1, i);
-        return cmp.byteAt(scratchBytes1, k);
-      }
+    private final BytesRefComparator cmp;
 
-      @Override
-      protected Sorter getFallbackSorter(int k) {
-        return fallbackSorter((o1, o2) -> cmp.compare(o1, o2, k));
-      }
-    };
+    protected MSBStringRadixSorter(BytesRefComparator cmp) {
+      super(cmp.comparedBytesCount);
+      this.cmp = cmp;
+    }
+
+    @Override
+    protected void swap(int i, int j) {
+      StringSorter.this.swap(i, j);
+    }
+
+    @Override
+    protected int byteAt(int i, int k) {
+      get(scratch1, scratchBytes1, i);
+      return cmp.byteAt(scratchBytes1, k);
+    }
+
+    @Override
+    protected Sorter getFallbackSorter(int k) {
+      return fallbackSorter((o1, o2) -> cmp.compare(o1, o2, k));
+    }
+  }
+
+  protected Sorter radixSorter(BytesRefComparator cmp) {
+    return new MSBStringRadixSorter(cmp);
   }
 
   protected Sorter fallbackSorter(Comparator<BytesRef> cmp) {
@@ -87,7 +98,9 @@ protected void swap(int i, int j) {
 
       @Override
       protected int compare(int i, int j) {
-        return StringSorter.this.compare(i, j);
+        get(scratch1, scratchBytes1, i);
+        get(scratch2, scratchBytes2, j);
+        return cmp.compare(scratchBytes1, scratchBytes2);
       }
 
       @Override