Speed up IndexedDISI Sparse #AdvanceExactWithinBlock for tiny step advance #12324

gf2121 · 2023-05-22T07:17:32Z

Today Sparse#AdvanceExactWithinBlock always need to read next doc and seek back if a doc not exists. This could do harm to performance in dense hit queries.

For example, a field exists in doc 1, 5. When advanceExact 2,3,4 it always need to read next doc (5) and seek back.

I think caching the next existing doc in block can help dense hit queries without too much harm to other cases.

I ran a benchmark with MatchAllDocsQuery on some fields with different sparsity:

sparsity=n means field only exists when doc % n == 0

sparsity	baseline(ms)	candidate(ms)	diff
32	255	112	-56.08%
64	260	95	-63.46%
128	264	94	-64.39%
256	262	93	-64.50%
512	260	91	-65.00%
1024	259	90	-65.25%
2048	258	90	-65.12%
4096	253	90	-64.43%
8192	243	90	-62.96%
16384	224	90	-59.82%
32768	184	90	-51.09%

zhaih

Nice optimization! I have a few minor comments but overall looks good

lucene/core/src/java/org/apache/lucene/codecs/lucene90/IndexedDISI.java

zhaih · 2023-06-01T04:51:07Z

lucene/core/src/java/org/apache/lucene/codecs/lucene90/IndexedDISI.java

@@ -495,6 +496,7 @@ private void readBlockHeader() throws IOException {
    if (numValues <= MAX_ARRAY_LENGTH) {
      method = Method.SPARSE;
      blockEnd = slice.getFilePointer() + (numValues << 1);
+      nextExistDocInBlock = -1;
    } else if (numValues == 65536) {


Can we change this to BLOCK_SIZE just by the way?

Sorry, I'm not sure i've got your point. Do you mean change to nextExistDocInBlock = BLOCK_SIZE; ?
I think that will make SPARSE$advanceExact always return false.

Oh sorry I (perhaps github webpage as well :)) didn't make it clear: I want to change 65536 to BLOCK_SIZE.
(Actually later on I found there's several more 65536 in this file, would you like to do a batch replacement?)

zhaih · 2023-06-01T04:53:40Z

lucene/core/src/java/org/apache/lucene/codecs/lucene90/IndexedDISI.java

            return true;
          }
        }
+        disi.nextExistDocInBlock = Integer.MAX_VALUE;


nit: I actually wonder whether we really need to set it to MAX_VALUE here (and below) since the next advance within the same block would be prevented by disi.index anyway. But set it to MAX_VALUE does no harm as well.

Oh yes! disi.index should be enough :)

zhaih

Thank you for addressing the comments!

jpountz

LGTM!

…vance (apache#12324)

…vance (#12324)

…dc8ca633e8bcf`) (#20) * Add next minor version 9.7.0 * Fix SynonymQuery equals implementation (apache#12260) The term member of TermAndBoost used to be a Term instance and became a BytesRef with apache#11941, which means its equals impl won't take the field name into account. The SynonymQuery equals impl needs to be updated accordingly to take the field into account as well, otherwise synonym queries with same term and boost across different fields are equal which is a bug. * Fix MMapDirectory documentation for Java 20 (apache#12265) * Don't generate stacktrace in CollectionTerminatedException (apache#12270) CollectionTerminatedException is always caught and never exposed to users so there's no point in filling in a stack-trace for it. * add missing changelog entry for apache#12260 * Add missing author to changelog entry for apache#12220 * Make query timeout members final in ExitableDirectoryReader (apache#12274) There's a couple of places in the Exitable wrapper classes where queryTimeout is set within the constructor and never modified. This commit makes such members final. * Update javadocs for QueryTimeout (apache#12272) QueryTimeout was introduced together with ExitableDirectoryReader but is now also optionally set to the IndexSearcher to wrap the bulk scorer with a TimeLimitingBulkScorer. Its javadocs needs updating. * Make TimeExceededException members final (apache#12271) TimeExceededException has three members that are set within its constructor and never modified. They can be made final. * DOAP changes for release 9.6.0 * Add back-compat indices for 9.6.0 * `ToParentBlockJoinQuery` Explain Support Score Mode (apache#12245) (apache#12283) * `ToParentBlockJoinQuery` Explain Support Score Mode --------- Co-authored-by: Marcus <marcuseagan@gmail.com> * Simplify SliceExecutor and QueueSizeBasedExecutor (apache#12285) The only behaviour that QueueSizeBasedExecutor overrides from SliceExecutor is when to execute on the caller thread. There is no need to override the whole invokeAll method for that. Instead, this commit introduces a shouldExecuteOnCallerThread method that can be overridden. * [Backport] GITHUB-11838 Add api to allow concurrent query rewrite (apache#12197) * GITHUB-11838 Change API to allow concurrent query rewrite (apache#11840) Replace Query#rewrite(IndexReader) with Query#rewrite(IndexSearcher) Co-authored-by: Patrick Zhai <zhaih@users.noreply.github.com> Co-authored-by: Adrien Grand <jpountz@gmail.com> Backport of apache#11840 Changes from original: - Query keeps `rewrite(IndexReader)`, but it is now deprecated - VirtualMethod is used to correct delegate to the overridden methods - The changes to `RewriteMethod` type classes are reverted, this increased the backwards compatibility impact. ------------------------------ ### Description Issue: apache#11838 #### Updated Proposal * Change signature of rewrite to `rewrite(IndexSearcher)` * How did I migrate the usage: * Use Intellij to do preliminary refactoring for me * For test usage, use searcher whenever is available, otherwise create one using `newSearcher(reader)` * For very few non-test classes which doesn't have IndexSearcher available but called rewrite, create a searcher using `new IndexSearcher(reader)`, tried my best to avoid creating it recurrently (Especially in `FieldQuery`) * For queries who have implemented the rewrite and uses some part of reader's functionality, use shortcut method when possible, otherwise pull out the reader from indexSearcher. * Backport: Concurrent rewrite for KnnVectorQuery (apache#12160) (apache#12288) * Concurrent rewrite for KnnVectorQuery (apache#12160) - Reduce overhead of non-concurrent search by preserving original execution - Improve readability by factoring into separate functions --------- Co-authored-by: Kaival Parikh <kaivalp2000@gmail.com> * adjusting for backport --------- Co-authored-by: Kaival Parikh <46070017+kaivalnp@users.noreply.github.com> Co-authored-by: Kaival Parikh <kaivalp2000@gmail.com> * toposort use iterator to avoid stackoverflow (apache#12286) Co-authored-by: tangdonghai <tangdonghai@meituan.com> # Conflicts: # lucene/CHANGES.txt * Fix test to compile with Java 11 after backport of apache#12286 * Update Javadoc for topoSortStates method after apache#12286 (apache#12292) * Optimize HNSW diversity calculation (apache#12235) * Minor cleanup and improvements to DaciukMihovAutomatonBuilder (apache#12305) * GITHUB-12291: Skip blank lines from stopwords list. (apache#12299) * Wrap Query rewrite backwards layer with AccessController (apache#12308) * Make sure APIJAR reproduces with different timezone (unfortunately java encodes the date using local timezone) (apache#12315) * Add multi-thread searchability to OnHeapHnswGraph (apache#12257) * Fix backport error * [MINOR] Update javadoc in Query class (apache#12233) - add a few missing full stops - update wording in the description of Query#equals method * [Backport] Integrate the Incubating Panama Vector API apache#12311 (apache#12327) Leverage accelerated vector hardware instructions in Vector Search. Lucene already has a mechanism that enables the use of non-final JDK APIs, currently used for the Previewing Pamana Foreign API. This change expands this mechanism to include the Incubating Pamana Vector API. When the jdk.incubator.vector module is present at run time the Panamaized version of the low-level primitives used by Vector Search is enabled. If not present, the default scalar version of these low-level primitives is used (as it was previously). Currently, we're only targeting support for JDK 20. A subsequent PR should evaluate JDK 21. --------- Co-authored-by: Uwe Schindler <uschindler@apache.org> Co-authored-by: Robert Muir <rmuir@apache.org> * Parallelize knn query rewrite across slices rather than segments (apache#12325) The concurrent query rewrite for knn vectory query introduced with apache#12160 requests one thread per segment to the executor. To align this with the IndexSearcher parallel behaviour, we should rather parallelize across slices. Also, we can reuse the same slice executor instance that the index searcher already holds, in that way we are using a QueueSizeBasedExecutor when a thread pool executor is provided. * Optimize ConjunctionDISI.createConjunction (apache#12328) This method is showing up as a little hot when profiling some queries. Almost all the time spent in this method is just burnt on ceremony around stream indirections that don't inline. Moving this to iterators, simplifying the check for same doc id and also saving one iteration (for the min cost) makes this method far cheaper and easier to read. * Update changes to be correct with ARM (it is called NEON there) * GH#12321: Marked DaciukMihovAutomatonBuilder as deprecated (apache#12332) Preparing to reduce visibility of this class in a future release * add BitSet.clear() (apache#12268) # Conflicts: # lucene/CHANGES.txt * Clenaup and update changes and synchronize with 9.x * Update TestVectorUtilProviders.java (apache#12338) * Don't generate stacktrace for TimeExceededException (apache#12335) The exception is package private and never rethrown, we can avoid generating a stacktrace for it. * Introduced the Word2VecSynonymFilter (apache#12169) Co-authored-by: Alessandro Benedetti <a.benedetti@sease.io> * Word2VecSynonymFilter constructor null check (apache#12169) * Use thread-safe search version of HnswGraphSearcher (apache#12246) Addressing comment received in the PR apache#12246 * Word2VecSynonymProvider to use standard Integer max value for hnsw searches (apache#12235) We observed this change was not ported previously from main in an old cherry-pick * Fix searchafter high latency when after value is out of range for segment (apache#12334) * Make memory fence in `ByteBufferGuard` explicit (apache#12290) * Add "direct to binary" option for DaciukMihovAutomatonBuilder and use it in TermInSetQuery#visit (apache#12320) * Add updateDocuments API which accept a query (reopen) (apache#12346) * GITHUB#11350: Handle backward compatibility when merging segments with different FieldInfo This commits restores Lucene 9's ability to handle indices created with Lucene 8 where there are discrepancies in FieldInfos, such as different IndexOptions * [Tessellator] Improve the checks that validate the diagonal between two polygon nodes (apache#12353) # Conflicts: # lucene/CHANGES.txt * feat: soft delete optimize (apache#12339) * Better paging when random reads go backwards (apache#12357) When reading data from outside the buffer, BufferedIndexInput always resets its buffer to start at the new read position. If we are reading backwards (for example, using an OffHeapFSTStore for a terms dictionary) then this can have the effect of re-reading the same data over and over again. This commit changes BufferedIndexInput to use paging when reading backwards, so that if we ask for a byte that is before the current buffer, we read a block of data of bufferSize that ends at the previous buffer start. Fixes apache#12356 * Work around SecurityManager issues during initialization of vector api (JDK-8309727) (apache#12362) * Restrict GraphTokenStreamFiniteStrings#articulationPointsRecurse recursion depth (apache#12249) * Implement MMapDirectory with Java 21 Project Panama Preview API (apache#12294) Backport incl JDK21 apijar file with java.util.Objects regenerated * remove relic in apijar folder caused by vector additions * Speed up IndexedDISI Sparse #AdvanceExactWithinBlock for tiny step advance (apache#12324) * Add checks in KNNVectorField / KNNVectorQuery to only allow non-null, non-empty and finite vectors (apache#12281) --------- Co-authored-by: Uwe Schindler <uschindler@apache.org> * Implement VectorUtilProvider with Java 21 Project Panama Vector API (apache#12363) (apache#12365) This commit enables the Panama Vector API for Java 21. The version of VectorUtilPanamaProvider for Java 21 is identical to that of Java 20. As such, there is no specific 21 version - the Java 20 version will be loaded from the MRJAR. * Add CHANGES.txt for apache#12334 Honor after value for skipping documents even if queue is not full for PagingFieldCollector (apache#12368) Signed-off-by: gashutos <gashutos@amazon.com> * Move TermAndBoost back to its original location. (apache#12366) PR apache#12169 accidentally moved the `TermAndBoost` class to a different location, which would break custom sub-classes of `QueryBuilder`. This commit moves it back to its original location. * GITHUB-12252: Add function queries for computing similarity scores between knn vectors (apache#12253) Co-authored-by: Alessandro Benedetti <a.benedetti@sease.io> * hunspell (minor): reduce allocations when processing compound rules (apache#12316) (cherry picked from commit a454388) * hunspell (minor): reduce allocations when reading the dictionary's morphological data (apache#12323) there can be many entries with morph data, so we'd better avoid compiling and matching regexes and even stream allocation (cherry picked from commit 4bf1b94) * TestHunspell: reduce the flakiness probability (apache#12351) * TestHunspell: reduce the flakiness probability We need to check how the timeout interacts with custom exception-throwing checkCanceled. The default timeout seems not enough for some CI agents, so let's increase it. Co-authored-by: Dawid Weiss <dawid.weiss@gmail.com> (cherry picked from commit 5b63a18) * This allows VectorUtilProvider tests to be executed although hardware may not fully support vectorization or if C2 is not enabled (apache#12376) --------- Signed-off-by: gashutos <gashutos@amazon.com> Co-authored-by: Alan Woodward <romseygeek@apache.org> Co-authored-by: Luca Cavanna <javanna@apache.org> Co-authored-by: Uwe Schindler <uschindler@apache.org> Co-authored-by: Armin Braun <me@obrown.io> Co-authored-by: Mikhail Khludnev <mkhludnev@users.noreply.github.com> Co-authored-by: Marcus <marcuseagan@gmail.com> Co-authored-by: Benjamin Trent <ben.w.trent@gmail.com> Co-authored-by: Kaival Parikh <46070017+kaivalnp@users.noreply.github.com> Co-authored-by: Kaival Parikh <kaivalp2000@gmail.com> Co-authored-by: tang donghai <tangdhcs@gmail.com> Co-authored-by: Patrick Zhai <zhaih@users.noreply.github.com> Co-authored-by: Greg Miller <gsmiller@gmail.com> Co-authored-by: Jerry Chin <metrxqin@gmail.com> Co-authored-by: Patrick Zhai <zhai7631@gmail.com> Co-authored-by: Andrey Bozhko <andybozhko@gmail.com> Co-authored-by: Chris Hegarty <62058229+ChrisHegarty@users.noreply.github.com> Co-authored-by: Robert Muir <rmuir@apache.org> Co-authored-by: Jonathan Ellis <jbellis@datastax.com> Co-authored-by: Daniele Antuzi <daniele.antuzi@gmail.com> Co-authored-by: Alessandro Benedetti <a.benedetti@sease.io> Co-authored-by: Chaitanya Gohel <104654647+gashutos@users.noreply.github.com> Co-authored-by: Petr Portnov | PROgrm_JARvis <pportnov@ozon.ru> Co-authored-by: Tomas Eduardo Fernandez Lobbe <tflobbe@apache.org> Co-authored-by: Ignacio Vera <ivera@apache.org> Co-authored-by: fudongying <30896830+fudongyingluck@users.noreply.github.com> Co-authored-by: Chris Fournier <chris.fournier@shopify.com> Co-authored-by: gf2121 <52390227+gf2121@users.noreply.github.com> Co-authored-by: Adrien Grand <jpountz@gmail.com> Co-authored-by: Elia Porciani <e.porciani@sease.io> Co-authored-by: Peter Gromov <peter@jetbrains.com>

gf2121 added 2 commits May 22, 2023 13:56

cache sparse next doc

6a791d7

add CHANGES

7ddde17

zhaih approved these changes Jun 1, 2023

View reviewed changes

gf2121 added 2 commits June 1, 2023 17:45

feedback fix: remove integer.max_value

31c7c12

feedback fix: 65536->BLOCK_SIZE

3823a25

zhaih approved these changes Jun 2, 2023

View reviewed changes

jpountz approved these changes Jun 8, 2023

View reviewed changes

gf2121 merged commit 30eba6d into apache:main Jun 13, 2023
4 checks passed

gf2121 added a commit to gf2121/lucene that referenced this pull request Jun 13, 2023

Speed up IndexedDISI Sparse #AdvanceExactWithinBlock for tiny step ad…

60b7d38

…vance (apache#12324)

gf2121 mentioned this pull request Jun 13, 2023

Speed up IndexedDISI Sparse #AdvanceExactWithinBlock for tiny step advance (Backport 9x) #12364

Merged

gf2121 added a commit that referenced this pull request Jun 13, 2023

Speed up IndexedDISI Sparse #AdvanceExactWithinBlock for tiny step ad…

7f16909

…vance (#12324)

gautamworah96 mentioned this pull request Jul 26, 2023

Potential bug in IndexedDISI90#SPARSE->advanceExactWithinBlock #12465

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Speed up IndexedDISI Sparse #AdvanceExactWithinBlock for tiny step advance #12324

Speed up IndexedDISI Sparse #AdvanceExactWithinBlock for tiny step advance #12324

gf2121 commented May 22, 2023

zhaih left a comment

zhaih Jun 1, 2023

gf2121 Jun 1, 2023 •

edited

Loading

zhaih Jun 1, 2023

zhaih Jun 1, 2023

gf2121 Jun 1, 2023

zhaih left a comment

jpountz left a comment

Speed up IndexedDISI Sparse #AdvanceExactWithinBlock for tiny step advance #12324

Speed up IndexedDISI Sparse #AdvanceExactWithinBlock for tiny step advance #12324

Conversation

gf2121 commented May 22, 2023

zhaih left a comment

Choose a reason for hiding this comment

zhaih Jun 1, 2023

Choose a reason for hiding this comment

gf2121 Jun 1, 2023 • edited Loading

Choose a reason for hiding this comment

zhaih Jun 1, 2023

Choose a reason for hiding this comment

zhaih Jun 1, 2023

Choose a reason for hiding this comment

gf2121 Jun 1, 2023

Choose a reason for hiding this comment

zhaih left a comment

Choose a reason for hiding this comment

jpountz left a comment

Choose a reason for hiding this comment

gf2121 Jun 1, 2023 •

edited

Loading