You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@lucene.apache.org by "mayya-sharipova (via GitHub)" <gi...@apache.org> on 2025/03/05 18:29:14 UTC
[PR] Speedup merging of HNSW graphs [lucene]
mayya-sharipova opened a new pull request, #14331:
URL: https://github.com/apache/lucene/pull/14331
Currently when doing merging of HNSW graphs incrementally,
we first initialize a graph from the biggest segment, and for other segments,
we rebuild the graphs completely by going through a segment's vector values
one by one, searching for it in the new graph to find best neighbours to connect with.
This PR proposes more efficient merging based on the idea if we know where
we want to insert a node, we have a good idea of where we want to insert its neighbours.
Similarly to the current approach, we initialize a new graph from a graph from the
biggest segments. For all other segments, we find a smaller set of nodes
that "covers" their graph, and we insert that set as usual. For other nodes,
outside of J set, we do lighter searches with calculated eps.
This allows substantial speedups.
The algorithm is based on the following steps:
1. Get all graphs that don't have deletions and sort them by size (descending).
2. Copy the largest graph to the new graph (`gL`).
3. For each remaining small graph (`gS`):
- Find the nodes that best cover `gS` (join set `j`). These nodes will be inserted into `gL`
as usual: by searching `gL` to find the best candidates (`w`) to which connect the nodes.
- For each remaining node in `gS`:
- We provide `eps` to search in `gL`. We form `eps` by the union of the node's
neighbors in `gS` and the node's neighbors' neighbors in `gL`. We also limit
`beamWidth` (`efConstruction` to `M * 2`).
Algorithm designed by Thomas Veasey
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "benwtrent (via GitHub)" <gi...@apache.org>.
benwtrent commented on PR #14331:
URL: https://github.com/apache/lucene/pull/14331#issuecomment-2737726576
> Experiment 3 new QSQ format:
...
These improvements make sense to me. The overall bottleneck of vector ops is way lower here, so simply doing fewer ops isn't going to have as big as an impact, though its nice to see that the impact is measurable and still nice :D
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "mayya-sharipova (via GitHub)" <gi...@apache.org>.
mayya-sharipova merged PR #14331:
URL: https://github.com/apache/lucene/pull/14331
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "mayya-sharipova (via GitHub)" <gi...@apache.org>.
mayya-sharipova commented on code in PR #14331:
URL: https://github.com/apache/lucene/pull/14331#discussion_r2005416489
##########
lucene/core/src/java/org/apache/lucene/util/hnsw/ConcurrentHnswMerger.java:
##########
@@ -51,19 +57,85 @@ protected HnswBuilder createBuilder(KnnVectorValues mergedVectorValues, int maxO
OnHeapHnswGraph graph;
BitSet initializedNodes = null;
- if (initReader == null) {
+ if (graphReaders.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
+ graphReaders.sort(Comparator.comparingInt(GraphReader::graphSize).reversed());
+ GraphReader initGraphReader = graphReaders.get(0);
+ KnnVectorsReader initReader = initGraphReader.reader();
+ MergeState.DocMap initDocMap = initGraphReader.initDocMap();
+ int initGraphSize = initGraphReader.graphSize();
HnswGraph initializerGraph = ((HnswGraphProvider) initReader).getGraph(fieldInfo.name);
+
if (initializerGraph.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
initializedNodes = new FixedBitSet(maxOrd);
- int[] oldToNewOrdinalMap = getNewOrdMapping(mergedVectorValues, initializedNodes);
+ int[] oldToNewOrdinalMap =
+ getNewOrdMapping(
+ fieldInfo,
+ initReader,
+ initDocMap,
+ initGraphSize,
+ mergedVectorValues,
+ initializedNodes);
graph = InitializedHnswGraphBuilder.initGraph(initializerGraph, oldToNewOrdinalMap, maxOrd);
}
}
return new HnswConcurrentMergeBuilder(
taskExecutor, numWorker, scorerSupplier, beamWidth, graph, initializedNodes);
}
+
+ /**
+ * Creates a new mapping from old ordinals to new ordinals and returns the total number of vectors
+ * in the newly merged segment.
+ *
+ * @param mergedVectorValues vector values in the merged segment
+ * @param initializedNodes track what nodes have been initialized
+ * @return the mapping from old ordinals to new ordinals
+ * @throws IOException If an error occurs while reading from the merge state
+ */
+ private static final int[] getNewOrdMapping(
+ FieldInfo fieldInfo,
+ KnnVectorsReader initReader,
+ MergeState.DocMap initDocMap,
+ int initGraphSize,
+ KnnVectorValues mergedVectorValues,
+ BitSet initializedNodes)
+ throws IOException {
+ KnnVectorValues.DocIndexIterator initializerIterator = null;
+
+ switch (fieldInfo.getVectorEncoding()) {
+ case BYTE -> initializerIterator = initReader.getByteVectorValues(fieldInfo.name).iterator();
+ case FLOAT32 ->
+ initializerIterator = initReader.getFloatVectorValues(fieldInfo.name).iterator();
+ }
+
+ IntIntHashMap newIdToOldOrdinal = new IntIntHashMap(initGraphSize);
+ int maxNewDocID = -1;
+ for (int docId = initializerIterator.nextDoc();
+ docId != NO_MORE_DOCS;
+ docId = initializerIterator.nextDoc()) {
+ int newId = initDocMap.get(docId);
+ maxNewDocID = Math.max(newId, maxNewDocID);
+ newIdToOldOrdinal.put(newId, initializerIterator.index());
+ }
+
+ if (maxNewDocID == -1) {
+ return new int[0];
+ }
+ final int[] oldToNewOrdinalMap = new int[initGraphSize];
+ KnnVectorValues.DocIndexIterator mergedVectorIterator = mergedVectorValues.iterator();
+ for (int newDocId = mergedVectorIterator.nextDoc();
+ newDocId <= maxNewDocID;
+ newDocId = mergedVectorIterator.nextDoc()) {
+ int hashDocIndex = newIdToOldOrdinal.indexOf(newDocId);
+ if (newIdToOldOrdinal.indexExists(hashDocIndex)) {
Review Comment:
@benwtrent Thanks for the comment. This piece is moved from [IncrementalHNSWGraphMerger](https://github.com/apache/lucene/blob/main/lucene/core/src/java/org/apache/lucene/util/hnsw/IncrementalHnswGraphMerger.java#L154) without modifications.
> Is this stuff around indexOf indexExists, etc. just performance improvements over a simple newIdToOldOrdinal.get(...)
Agree about `IntIntHashMap` and may be this piece can be optimized. But as it is not a part of this PR, I will leave TODO item for that for follow up.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "msokolov (via GitHub)" <gi...@apache.org>.
msokolov commented on PR #14331:
URL: https://github.com/apache/lucene/pull/14331#issuecomment-2737095119
yes, looks good, I think this is the right tradeoff. We even seem to get improved query performance in some cases. +1 to merge this
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "benwtrent (via GitHub)" <gi...@apache.org>.
benwtrent commented on code in PR #14331:
URL: https://github.com/apache/lucene/pull/14331#discussion_r2003974286
##########
lucene/core/src/java/org/apache/lucene/util/hnsw/ConcurrentHnswMerger.java:
##########
@@ -51,19 +57,85 @@ protected HnswBuilder createBuilder(KnnVectorValues mergedVectorValues, int maxO
OnHeapHnswGraph graph;
BitSet initializedNodes = null;
- if (initReader == null) {
+ if (graphReaders.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
+ graphReaders.sort(Comparator.comparingInt(GraphReader::graphSize).reversed());
+ GraphReader initGraphReader = graphReaders.get(0);
+ KnnVectorsReader initReader = initGraphReader.reader();
+ MergeState.DocMap initDocMap = initGraphReader.initDocMap();
+ int initGraphSize = initGraphReader.graphSize();
HnswGraph initializerGraph = ((HnswGraphProvider) initReader).getGraph(fieldInfo.name);
+
if (initializerGraph.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
initializedNodes = new FixedBitSet(maxOrd);
- int[] oldToNewOrdinalMap = getNewOrdMapping(mergedVectorValues, initializedNodes);
+ int[] oldToNewOrdinalMap =
+ getNewOrdMapping(
+ fieldInfo,
+ initReader,
+ initDocMap,
+ initGraphSize,
+ mergedVectorValues,
+ initializedNodes);
graph = InitializedHnswGraphBuilder.initGraph(initializerGraph, oldToNewOrdinalMap, maxOrd);
}
}
return new HnswConcurrentMergeBuilder(
taskExecutor, numWorker, scorerSupplier, beamWidth, graph, initializedNodes);
}
+
+ /**
+ * Creates a new mapping from old ordinals to new ordinals and returns the total number of vectors
+ * in the newly merged segment.
+ *
+ * @param mergedVectorValues vector values in the merged segment
+ * @param initializedNodes track what nodes have been initialized
+ * @return the mapping from old ordinals to new ordinals
+ * @throws IOException If an error occurs while reading from the merge state
+ */
+ private static final int[] getNewOrdMapping(
+ FieldInfo fieldInfo,
+ KnnVectorsReader initReader,
+ MergeState.DocMap initDocMap,
+ int initGraphSize,
+ KnnVectorValues mergedVectorValues,
+ BitSet initializedNodes)
+ throws IOException {
+ KnnVectorValues.DocIndexIterator initializerIterator = null;
+
+ switch (fieldInfo.getVectorEncoding()) {
+ case BYTE -> initializerIterator = initReader.getByteVectorValues(fieldInfo.name).iterator();
+ case FLOAT32 ->
+ initializerIterator = initReader.getFloatVectorValues(fieldInfo.name).iterator();
+ }
+
+ IntIntHashMap newIdToOldOrdinal = new IntIntHashMap(initGraphSize);
+ int maxNewDocID = -1;
+ for (int docId = initializerIterator.nextDoc();
+ docId != NO_MORE_DOCS;
+ docId = initializerIterator.nextDoc()) {
+ int newId = initDocMap.get(docId);
+ maxNewDocID = Math.max(newId, maxNewDocID);
+ newIdToOldOrdinal.put(newId, initializerIterator.index());
+ }
+
+ if (maxNewDocID == -1) {
+ return new int[0];
+ }
+ final int[] oldToNewOrdinalMap = new int[initGraphSize];
+ KnnVectorValues.DocIndexIterator mergedVectorIterator = mergedVectorValues.iterator();
+ for (int newDocId = mergedVectorIterator.nextDoc();
+ newDocId <= maxNewDocID;
+ newDocId = mergedVectorIterator.nextDoc()) {
+ int hashDocIndex = newIdToOldOrdinal.indexOf(newDocId);
+ if (newIdToOldOrdinal.indexExists(hashDocIndex)) {
Review Comment:
Is this stuff around `indexOf` `indexExists`, etc. just performance improvements over a simple `newIdToOldOrdinal.get(...)`?
Looking at the `IntIntHashMap` its weird that "does not exist" may actually just be `0`, where `0` is a valid doc id :/.
##########
lucene/core/src/java/org/apache/lucene/util/hnsw/ConcurrentHnswMerger.java:
##########
@@ -51,19 +57,85 @@ protected HnswBuilder createBuilder(KnnVectorValues mergedVectorValues, int maxO
OnHeapHnswGraph graph;
BitSet initializedNodes = null;
- if (initReader == null) {
+ if (graphReaders.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
+ graphReaders.sort(Comparator.comparingInt(GraphReader::graphSize).reversed());
+ GraphReader initGraphReader = graphReaders.get(0);
+ KnnVectorsReader initReader = initGraphReader.reader();
+ MergeState.DocMap initDocMap = initGraphReader.initDocMap();
+ int initGraphSize = initGraphReader.graphSize();
HnswGraph initializerGraph = ((HnswGraphProvider) initReader).getGraph(fieldInfo.name);
+
if (initializerGraph.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
initializedNodes = new FixedBitSet(maxOrd);
- int[] oldToNewOrdinalMap = getNewOrdMapping(mergedVectorValues, initializedNodes);
+ int[] oldToNewOrdinalMap =
+ getNewOrdMapping(
+ fieldInfo,
+ initReader,
+ initDocMap,
+ initGraphSize,
+ mergedVectorValues,
+ initializedNodes);
graph = InitializedHnswGraphBuilder.initGraph(initializerGraph, oldToNewOrdinalMap, maxOrd);
}
}
return new HnswConcurrentMergeBuilder(
taskExecutor, numWorker, scorerSupplier, beamWidth, graph, initializedNodes);
}
+
+ /**
+ * Creates a new mapping from old ordinals to new ordinals and returns the total number of vectors
+ * in the newly merged segment.
+ *
+ * @param mergedVectorValues vector values in the merged segment
+ * @param initializedNodes track what nodes have been initialized
+ * @return the mapping from old ordinals to new ordinals
+ * @throws IOException If an error occurs while reading from the merge state
+ */
+ private static final int[] getNewOrdMapping(
Review Comment:
```suggestion
private static int[] getNewOrdMapping(
```
##########
lucene/core/src/java/org/apache/lucene/util/hnsw/MergingHnswGraphBuilder.java:
##########
@@ -0,0 +1,198 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.lucene.util.hnsw;
+
+import static org.apache.lucene.search.DocIdSetIterator.NO_MORE_DOCS;
+
+import java.io.IOException;
+import java.util.HashSet;
+import java.util.Set;
+import org.apache.lucene.util.BitSet;
+
+/**
+ * A graph builder that is used during segments' merging.
+ *
+ * <p>This builder uses a smart algorithm to merge multiple graphs into a single graph. The
+ * algorithm is based on the idea if we know where we want to insert a node, we have a good idea of
+ * where we want to insert its neighbours.
+ *
+ * <p>The algorithm is based on the following steps: 1. Get all graphs that don't have deletions and
+ * sort them by size. 2. Copy the largest graph to the new graph (gL). 3. For each remaining small
+ * graph (gS) - Find the nodes that best cover gS: join set `j`. These nodes will be inserted into
+ * gL as usual: by searching gL to find the best candidates: `w` to which connect the nodes. - For
+ * each remaining node in gS: - we do NOT do search in gL. Instead, we form `w` by union of the
+ * node's neighbors' in Gs and the node's neighbors' neighbors' in gL.
+ *
+ * <p>We expect the size of join set `j` to be small, 1/5-1/2 of size gS. And for the rest of the
+ * nodes of gS, we expect savings by not doing extra search in gL.
+ *
+ * @lucene.experimental
+ */
Review Comment:
class comments here twice?
##########
lucene/core/src/java/org/apache/lucene/util/hnsw/ConcurrentHnswMerger.java:
##########
@@ -51,19 +57,85 @@ protected HnswBuilder createBuilder(KnnVectorValues mergedVectorValues, int maxO
OnHeapHnswGraph graph;
BitSet initializedNodes = null;
- if (initReader == null) {
+ if (graphReaders.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
+ graphReaders.sort(Comparator.comparingInt(GraphReader::graphSize).reversed());
+ GraphReader initGraphReader = graphReaders.get(0);
+ KnnVectorsReader initReader = initGraphReader.reader();
+ MergeState.DocMap initDocMap = initGraphReader.initDocMap();
+ int initGraphSize = initGraphReader.graphSize();
HnswGraph initializerGraph = ((HnswGraphProvider) initReader).getGraph(fieldInfo.name);
+
if (initializerGraph.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
initializedNodes = new FixedBitSet(maxOrd);
- int[] oldToNewOrdinalMap = getNewOrdMapping(mergedVectorValues, initializedNodes);
+ int[] oldToNewOrdinalMap =
+ getNewOrdMapping(
+ fieldInfo,
+ initReader,
+ initDocMap,
+ initGraphSize,
+ mergedVectorValues,
+ initializedNodes);
graph = InitializedHnswGraphBuilder.initGraph(initializerGraph, oldToNewOrdinalMap, maxOrd);
}
}
return new HnswConcurrentMergeBuilder(
taskExecutor, numWorker, scorerSupplier, beamWidth, graph, initializedNodes);
}
+
+ /**
+ * Creates a new mapping from old ordinals to new ordinals and returns the total number of vectors
+ * in the newly merged segment.
+ *
+ * @param mergedVectorValues vector values in the merged segment
+ * @param initializedNodes track what nodes have been initialized
+ * @return the mapping from old ordinals to new ordinals
+ * @throws IOException If an error occurs while reading from the merge state
+ */
+ private static final int[] getNewOrdMapping(
+ FieldInfo fieldInfo,
+ KnnVectorsReader initReader,
+ MergeState.DocMap initDocMap,
+ int initGraphSize,
+ KnnVectorValues mergedVectorValues,
+ BitSet initializedNodes)
+ throws IOException {
+ KnnVectorValues.DocIndexIterator initializerIterator = null;
+
+ switch (fieldInfo.getVectorEncoding()) {
+ case BYTE -> initializerIterator = initReader.getByteVectorValues(fieldInfo.name).iterator();
+ case FLOAT32 ->
+ initializerIterator = initReader.getFloatVectorValues(fieldInfo.name).iterator();
+ }
+
+ IntIntHashMap newIdToOldOrdinal = new IntIntHashMap(initGraphSize);
+ int maxNewDocID = -1;
+ for (int docId = initializerIterator.nextDoc();
+ docId != NO_MORE_DOCS;
+ docId = initializerIterator.nextDoc()) {
+ int newId = initDocMap.get(docId);
+ maxNewDocID = Math.max(newId, maxNewDocID);
+ newIdToOldOrdinal.put(newId, initializerIterator.index());
Review Comment:
```suggestion
assert newIdToOldOrdinal.containsKey(newId) == false;
newIdToOldOrdinal.put(newId, initializerIterator.index());
```
We should assert if we put `newId` more than once.
##########
lucene/core/src/java/org/apache/lucene/util/hnsw/HnswGraphBuilder.java:
##########
@@ -208,6 +211,11 @@ private void addVectors(int maxOrd) throws IOException {
}
public void addGraphNode(int node, UpdateableRandomVectorScorer scorer) throws IOException {
+ addGraphNodeInternal(node, scorer, null);
+ }
+
+ private void addGraphNodeInternal(
+ int node, UpdateableRandomVectorScorer scorer, Set<Integer> eps0) throws IOException {
Review Comment:
`IntHashSet` instead maybe?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "mayya-sharipova (via GitHub)" <gi...@apache.org>.
mayya-sharipova commented on code in PR #14331:
URL: https://github.com/apache/lucene/pull/14331#discussion_r2005461835
##########
lucene/core/src/java/org/apache/lucene/util/hnsw/ConcurrentHnswMerger.java:
##########
@@ -51,19 +57,85 @@ protected HnswBuilder createBuilder(KnnVectorValues mergedVectorValues, int maxO
OnHeapHnswGraph graph;
BitSet initializedNodes = null;
- if (initReader == null) {
+ if (graphReaders.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
+ graphReaders.sort(Comparator.comparingInt(GraphReader::graphSize).reversed());
+ GraphReader initGraphReader = graphReaders.get(0);
+ KnnVectorsReader initReader = initGraphReader.reader();
+ MergeState.DocMap initDocMap = initGraphReader.initDocMap();
+ int initGraphSize = initGraphReader.graphSize();
HnswGraph initializerGraph = ((HnswGraphProvider) initReader).getGraph(fieldInfo.name);
+
if (initializerGraph.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
initializedNodes = new FixedBitSet(maxOrd);
- int[] oldToNewOrdinalMap = getNewOrdMapping(mergedVectorValues, initializedNodes);
+ int[] oldToNewOrdinalMap =
+ getNewOrdMapping(
+ fieldInfo,
+ initReader,
+ initDocMap,
+ initGraphSize,
+ mergedVectorValues,
+ initializedNodes);
graph = InitializedHnswGraphBuilder.initGraph(initializerGraph, oldToNewOrdinalMap, maxOrd);
}
}
return new HnswConcurrentMergeBuilder(
taskExecutor, numWorker, scorerSupplier, beamWidth, graph, initializedNodes);
}
+
+ /**
+ * Creates a new mapping from old ordinals to new ordinals and returns the total number of vectors
+ * in the newly merged segment.
+ *
+ * @param mergedVectorValues vector values in the merged segment
+ * @param initializedNodes track what nodes have been initialized
+ * @return the mapping from old ordinals to new ordinals
+ * @throws IOException If an error occurs while reading from the merge state
+ */
+ private static final int[] getNewOrdMapping(
Review Comment:
Addressed in cb852a6387a09ba43049b8a24f1e026c309b368b
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "msokolov (via GitHub)" <gi...@apache.org>.
msokolov commented on PR #14331:
URL: https://github.com/apache/lucene/pull/14331#issuecomment-2701861504
oh, this is a neat idea! Looks like we sacrifice some query performance (in some cases) for a big improvement in indexing time. I wonder if we've tried other values of `beamWidth` to see if we can recover the query performance?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "mayya-sharipova (via GitHub)" <gi...@apache.org>.
mayya-sharipova commented on PR #14331:
URL: https://github.com/apache/lucene/pull/14331#issuecomment-2737598747
I've done additional benchmarks with the new Optimized Scalar Quantization format that quantize 32x times to 1 single bit (Lucene102HnswBinaryQuantizedVectorsFormat). And here we can see less improvements but still improvements:
### Experiment 3 new QSQ format:
The average speedups from baseline to candidate are:
<mark>Index Time Speedup: **1.33x**</mark>
<mark>Force Merge Speedup: **1.34x**</mark>
Evaluation is done with Luceneutil on these datasets:
1. **quora-E5-small**; 522931 docs; 384 dims; 7 bits quantized; cosine metric
- baseline: index time: **70.71s**, force merge: **59.38s**
- candidate: index time: **58.25s**, force merge: **40.15s**
2. **cohere-wikipedia-v2**; 1M docs; 768 dims; 7 bits quantized; cosine metric
- baseline: index time: **203.08s**, force merge: **107.27s**
- candidate: index time: **142.27s**, force merge: **85.68s**
3. **gist**; 960 dims, 1M docs; 7 bits quantized; euclidean metric
- baseline: index time: **110.35s**, force merge: **323.66s**
- candidate: index time: **105.52s**, force merge: **202.20s**
4. **cohere-wikipedia-v3**; 1M docs; 1024 dims; 7 bits quantized; dot_product metric
- baseline: index time: **313.43s**, force merge: **165.98s**
- candidate: index time: **190.63s,** force merge: **159.95s**
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "mayya-sharipova (via GitHub)" <gi...@apache.org>.
mayya-sharipova commented on code in PR #14331:
URL: https://github.com/apache/lucene/pull/14331#discussion_r2005464935
##########
lucene/core/src/java/org/apache/lucene/util/hnsw/MergingHnswGraphBuilder.java:
##########
@@ -0,0 +1,198 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.lucene.util.hnsw;
+
+import static org.apache.lucene.search.DocIdSetIterator.NO_MORE_DOCS;
+
+import java.io.IOException;
+import java.util.HashSet;
+import java.util.Set;
+import org.apache.lucene.util.BitSet;
+
+/**
+ * A graph builder that is used during segments' merging.
+ *
+ * <p>This builder uses a smart algorithm to merge multiple graphs into a single graph. The
+ * algorithm is based on the idea if we know where we want to insert a node, we have a good idea of
+ * where we want to insert its neighbours.
+ *
+ * <p>The algorithm is based on the following steps: 1. Get all graphs that don't have deletions and
+ * sort them by size. 2. Copy the largest graph to the new graph (gL). 3. For each remaining small
+ * graph (gS) - Find the nodes that best cover gS: join set `j`. These nodes will be inserted into
+ * gL as usual: by searching gL to find the best candidates: `w` to which connect the nodes. - For
+ * each remaining node in gS: - we do NOT do search in gL. Instead, we form `w` by union of the
+ * node's neighbors' in Gs and the node's neighbors' neighbors' in gL.
+ *
+ * <p>We expect the size of join set `j` to be small, 1/5-1/2 of size gS. And for the rest of the
+ * nodes of gS, we expect savings by not doing extra search in gL.
+ *
+ * @lucene.experimental
+ */
Review Comment:
Addressed in cb852a6387a09ba43049b8a24f1e026c309b368b
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "mayya-sharipova (via GitHub)" <gi...@apache.org>.
mayya-sharipova commented on code in PR #14331:
URL: https://github.com/apache/lucene/pull/14331#discussion_r2005416489
##########
lucene/core/src/java/org/apache/lucene/util/hnsw/ConcurrentHnswMerger.java:
##########
@@ -51,19 +57,85 @@ protected HnswBuilder createBuilder(KnnVectorValues mergedVectorValues, int maxO
OnHeapHnswGraph graph;
BitSet initializedNodes = null;
- if (initReader == null) {
+ if (graphReaders.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
+ graphReaders.sort(Comparator.comparingInt(GraphReader::graphSize).reversed());
+ GraphReader initGraphReader = graphReaders.get(0);
+ KnnVectorsReader initReader = initGraphReader.reader();
+ MergeState.DocMap initDocMap = initGraphReader.initDocMap();
+ int initGraphSize = initGraphReader.graphSize();
HnswGraph initializerGraph = ((HnswGraphProvider) initReader).getGraph(fieldInfo.name);
+
if (initializerGraph.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
initializedNodes = new FixedBitSet(maxOrd);
- int[] oldToNewOrdinalMap = getNewOrdMapping(mergedVectorValues, initializedNodes);
+ int[] oldToNewOrdinalMap =
+ getNewOrdMapping(
+ fieldInfo,
+ initReader,
+ initDocMap,
+ initGraphSize,
+ mergedVectorValues,
+ initializedNodes);
graph = InitializedHnswGraphBuilder.initGraph(initializerGraph, oldToNewOrdinalMap, maxOrd);
}
}
return new HnswConcurrentMergeBuilder(
taskExecutor, numWorker, scorerSupplier, beamWidth, graph, initializedNodes);
}
+
+ /**
+ * Creates a new mapping from old ordinals to new ordinals and returns the total number of vectors
+ * in the newly merged segment.
+ *
+ * @param mergedVectorValues vector values in the merged segment
+ * @param initializedNodes track what nodes have been initialized
+ * @return the mapping from old ordinals to new ordinals
+ * @throws IOException If an error occurs while reading from the merge state
+ */
+ private static final int[] getNewOrdMapping(
+ FieldInfo fieldInfo,
+ KnnVectorsReader initReader,
+ MergeState.DocMap initDocMap,
+ int initGraphSize,
+ KnnVectorValues mergedVectorValues,
+ BitSet initializedNodes)
+ throws IOException {
+ KnnVectorValues.DocIndexIterator initializerIterator = null;
+
+ switch (fieldInfo.getVectorEncoding()) {
+ case BYTE -> initializerIterator = initReader.getByteVectorValues(fieldInfo.name).iterator();
+ case FLOAT32 ->
+ initializerIterator = initReader.getFloatVectorValues(fieldInfo.name).iterator();
+ }
+
+ IntIntHashMap newIdToOldOrdinal = new IntIntHashMap(initGraphSize);
+ int maxNewDocID = -1;
+ for (int docId = initializerIterator.nextDoc();
+ docId != NO_MORE_DOCS;
+ docId = initializerIterator.nextDoc()) {
+ int newId = initDocMap.get(docId);
+ maxNewDocID = Math.max(newId, maxNewDocID);
+ newIdToOldOrdinal.put(newId, initializerIterator.index());
+ }
+
+ if (maxNewDocID == -1) {
+ return new int[0];
+ }
+ final int[] oldToNewOrdinalMap = new int[initGraphSize];
+ KnnVectorValues.DocIndexIterator mergedVectorIterator = mergedVectorValues.iterator();
+ for (int newDocId = mergedVectorIterator.nextDoc();
+ newDocId <= maxNewDocID;
+ newDocId = mergedVectorIterator.nextDoc()) {
+ int hashDocIndex = newIdToOldOrdinal.indexOf(newDocId);
+ if (newIdToOldOrdinal.indexExists(hashDocIndex)) {
Review Comment:
@benwtrent Thanks for the comment. This piece is moved from [IncrementalHNSWGraphMerger](https://github.com/apache/lucene/blob/main/lucene/core/src/java/org/apache/lucene/util/hnsw/IncrementalHnswGraphMerger.java#L154) without modifications.
> Is this stuff around indexOf indexExists, etc. just performance improvements over a simple newIdToOldOrdinal.get(...)
Agree about `IntIntHashMap` and may be this piece can be optimized.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "mayya-sharipova (via GitHub)" <gi...@apache.org>.
mayya-sharipova commented on PR #14331:
URL: https://github.com/apache/lucene/pull/14331#issuecomment-2730506974
@msokolov Thanks for the comment.
I've experimented setting: beamCandidates0 to `M * 3` increasing it from the previous `M*2` when building merged graphs.
Graphs look better, but there are still significant speedups
Evaluation is done with Luceneutil on these datasets:
1. **quora-E5-small**; 522931 docs; 384 dims; 7 bits quantized; cosine metric
- baseline: index time: **112.41s**, force merge: **113.81s**
- candidate: index time: **81.55s**, force merge: **70.87s**
2. **cohere-wikipedia-v2**; 1M docs; 768 dims; 7 bits quantized; cosine metric
- baseline: index time: **158.1s**, force merge: **425.20s**
- candidate: index time: **122.95s**, force merge: **239.28s**
3. **gist**; 960 dims, 1M docs; 7 bits quantized; euclidean metric
- baseline: index time: **141.82s**, force merge: **536.07s**
- candidate: index time: **119.26s**, force merge: **279.05s**
4. **cohere-wikipedia-v3**; 1M docs; 1024 dims; 7 bits quantized; dot_product metric
- baseline: index time: **211.86s**, force merge: **654.97s**
- candidate: index time: **168.22s,** force merge: **414.12s**
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "mayya-sharipova (via GitHub)" <gi...@apache.org>.
mayya-sharipova commented on PR #14331:
URL: https://github.com/apache/lucene/pull/14331#issuecomment-2701756507
Evaluation is done with Luceneutil on these datasets:
Rebased against Lucene main branch:
1. **quora-E5-small**; 522931 docs; 384 dims; 7 bits quantized; cosine metric
- baseline: index time: **112.41s**, force merge: **113.81s**
- candidate: index time: **77.17s**, force merge: **56.01s**
2. **cohere-wikipedia-v2**; 1M docs; 768 dims; 7 bits quantized; cosine metric
- baseline: index time: **158.1s**, force merge: **425.20s**
- candidate: index time: **113.68s**, force merge: **201.52s**
3. **gist**; 960 dims, 1M docs; 7 bits quantized; euclidean metric
- baseline: index time: **141.82s**, force merge: **536.07s**
- candidate: index time: **108.77s**, force merge: **279.05s**
4. **cohere-wikipedia-v3**; 1M docs; 1024 dims; 7 bits quantized; dot_product metric
- baseline: index time: **211.86s**, force merge: **654.97s**
- candidate: index time: **161.51s,** force merge: **320.54s**
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org
Re: [PR] Speedup merging of HNSW graphs [lucene]
Posted by "mayya-sharipova (via GitHub)" <gi...@apache.org>.
mayya-sharipova commented on code in PR #14331:
URL: https://github.com/apache/lucene/pull/14331#discussion_r2005462586
##########
lucene/core/src/java/org/apache/lucene/util/hnsw/ConcurrentHnswMerger.java:
##########
@@ -51,19 +57,85 @@ protected HnswBuilder createBuilder(KnnVectorValues mergedVectorValues, int maxO
OnHeapHnswGraph graph;
BitSet initializedNodes = null;
- if (initReader == null) {
+ if (graphReaders.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
+ graphReaders.sort(Comparator.comparingInt(GraphReader::graphSize).reversed());
+ GraphReader initGraphReader = graphReaders.get(0);
+ KnnVectorsReader initReader = initGraphReader.reader();
+ MergeState.DocMap initDocMap = initGraphReader.initDocMap();
+ int initGraphSize = initGraphReader.graphSize();
HnswGraph initializerGraph = ((HnswGraphProvider) initReader).getGraph(fieldInfo.name);
+
if (initializerGraph.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
initializedNodes = new FixedBitSet(maxOrd);
- int[] oldToNewOrdinalMap = getNewOrdMapping(mergedVectorValues, initializedNodes);
+ int[] oldToNewOrdinalMap =
+ getNewOrdMapping(
+ fieldInfo,
+ initReader,
+ initDocMap,
+ initGraphSize,
+ mergedVectorValues,
+ initializedNodes);
graph = InitializedHnswGraphBuilder.initGraph(initializerGraph, oldToNewOrdinalMap, maxOrd);
}
}
return new HnswConcurrentMergeBuilder(
taskExecutor, numWorker, scorerSupplier, beamWidth, graph, initializedNodes);
}
+
+ /**
+ * Creates a new mapping from old ordinals to new ordinals and returns the total number of vectors
+ * in the newly merged segment.
+ *
+ * @param mergedVectorValues vector values in the merged segment
+ * @param initializedNodes track what nodes have been initialized
+ * @return the mapping from old ordinals to new ordinals
+ * @throws IOException If an error occurs while reading from the merge state
+ */
+ private static final int[] getNewOrdMapping(
+ FieldInfo fieldInfo,
+ KnnVectorsReader initReader,
+ MergeState.DocMap initDocMap,
+ int initGraphSize,
+ KnnVectorValues mergedVectorValues,
+ BitSet initializedNodes)
+ throws IOException {
+ KnnVectorValues.DocIndexIterator initializerIterator = null;
+
+ switch (fieldInfo.getVectorEncoding()) {
+ case BYTE -> initializerIterator = initReader.getByteVectorValues(fieldInfo.name).iterator();
+ case FLOAT32 ->
+ initializerIterator = initReader.getFloatVectorValues(fieldInfo.name).iterator();
+ }
+
+ IntIntHashMap newIdToOldOrdinal = new IntIntHashMap(initGraphSize);
+ int maxNewDocID = -1;
+ for (int docId = initializerIterator.nextDoc();
+ docId != NO_MORE_DOCS;
+ docId = initializerIterator.nextDoc()) {
+ int newId = initDocMap.get(docId);
+ maxNewDocID = Math.max(newId, maxNewDocID);
+ newIdToOldOrdinal.put(newId, initializerIterator.index());
Review Comment:
Addressed in cb852a6387a09ba43049b8a24f1e026c309b368b
##########
lucene/core/src/java/org/apache/lucene/util/hnsw/ConcurrentHnswMerger.java:
##########
@@ -51,19 +57,85 @@ protected HnswBuilder createBuilder(KnnVectorValues mergedVectorValues, int maxO
OnHeapHnswGraph graph;
BitSet initializedNodes = null;
- if (initReader == null) {
+ if (graphReaders.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
+ graphReaders.sort(Comparator.comparingInt(GraphReader::graphSize).reversed());
+ GraphReader initGraphReader = graphReaders.get(0);
+ KnnVectorsReader initReader = initGraphReader.reader();
+ MergeState.DocMap initDocMap = initGraphReader.initDocMap();
+ int initGraphSize = initGraphReader.graphSize();
HnswGraph initializerGraph = ((HnswGraphProvider) initReader).getGraph(fieldInfo.name);
+
if (initializerGraph.size() == 0) {
graph = new OnHeapHnswGraph(M, maxOrd);
} else {
initializedNodes = new FixedBitSet(maxOrd);
- int[] oldToNewOrdinalMap = getNewOrdMapping(mergedVectorValues, initializedNodes);
+ int[] oldToNewOrdinalMap =
+ getNewOrdMapping(
+ fieldInfo,
+ initReader,
+ initDocMap,
+ initGraphSize,
+ mergedVectorValues,
+ initializedNodes);
graph = InitializedHnswGraphBuilder.initGraph(initializerGraph, oldToNewOrdinalMap, maxOrd);
}
}
return new HnswConcurrentMergeBuilder(
taskExecutor, numWorker, scorerSupplier, beamWidth, graph, initializedNodes);
}
+
+ /**
+ * Creates a new mapping from old ordinals to new ordinals and returns the total number of vectors
+ * in the newly merged segment.
+ *
+ * @param mergedVectorValues vector values in the merged segment
+ * @param initializedNodes track what nodes have been initialized
+ * @return the mapping from old ordinals to new ordinals
+ * @throws IOException If an error occurs while reading from the merge state
+ */
+ private static final int[] getNewOrdMapping(
+ FieldInfo fieldInfo,
+ KnnVectorsReader initReader,
+ MergeState.DocMap initDocMap,
+ int initGraphSize,
+ KnnVectorValues mergedVectorValues,
+ BitSet initializedNodes)
+ throws IOException {
+ KnnVectorValues.DocIndexIterator initializerIterator = null;
+
+ switch (fieldInfo.getVectorEncoding()) {
+ case BYTE -> initializerIterator = initReader.getByteVectorValues(fieldInfo.name).iterator();
+ case FLOAT32 ->
+ initializerIterator = initReader.getFloatVectorValues(fieldInfo.name).iterator();
+ }
+
+ IntIntHashMap newIdToOldOrdinal = new IntIntHashMap(initGraphSize);
+ int maxNewDocID = -1;
+ for (int docId = initializerIterator.nextDoc();
+ docId != NO_MORE_DOCS;
+ docId = initializerIterator.nextDoc()) {
+ int newId = initDocMap.get(docId);
+ maxNewDocID = Math.max(newId, maxNewDocID);
+ newIdToOldOrdinal.put(newId, initializerIterator.index());
+ }
+
+ if (maxNewDocID == -1) {
+ return new int[0];
+ }
+ final int[] oldToNewOrdinalMap = new int[initGraphSize];
+ KnnVectorValues.DocIndexIterator mergedVectorIterator = mergedVectorValues.iterator();
+ for (int newDocId = mergedVectorIterator.nextDoc();
+ newDocId <= maxNewDocID;
+ newDocId = mergedVectorIterator.nextDoc()) {
+ int hashDocIndex = newIdToOldOrdinal.indexOf(newDocId);
+ if (newIdToOldOrdinal.indexExists(hashDocIndex)) {
Review Comment:
Addressed in cb852a6387a09ba43049b8a24f1e026c309b368b
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@lucene.apache.org
For additional commands, e-mail: issues-help@lucene.apache.org