zacharymorn commented on a change in pull request #240: URL: https://github.com/apache/lucene/pull/240#discussion_r689035631
########## File path: lucene/core/src/java/org/apache/lucene/search/TopFieldCollectorManager.java ########## @@ -0,0 +1,141 @@ +/* + * 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.search; + +import java.io.IOException; +import java.util.ArrayList; +import java.util.Collection; +import java.util.List; + +/** + * Create a TopFieldCollectorManager which uses a shared hit counter to maintain number of hits and + * a shared {@link MaxScoreAccumulator} to propagate the minimum score across segments if the + * primary sort is by relevancy. + * + * <p>Note that a new collectorManager should be created for each search due to its internal states. + */ +public class TopFieldCollectorManager implements CollectorManager<TopFieldCollector, TopFieldDocs> { + private final Sort sort; + private final int numHits; + private final FieldDoc after; + private final HitsThresholdChecker hitsThresholdChecker; + private final MaxScoreAccumulator minScoreAcc; + private final List<TopFieldCollector> collectors; + + public TopFieldCollectorManager(Sort sort, int numHits, FieldDoc after, int totalHitsThreshold) { + if (totalHitsThreshold < 0) { + throw new IllegalArgumentException( + "totalHitsThreshold must be >= 0, got " + totalHitsThreshold); + } + + this.sort = sort; + this.numHits = numHits; + this.after = after; + /* + nocommit + Should the following two be passed in instead? Possible custom initialization based on executor status and slices? + On the other hand, in a single-threaded environment, shared HitsThresholdChecker and MaxScoreAccumulator should be fast without lock contention anyway? Review comment: Thanks for the great suggestion! I just ran luceneutil, and it turns out that the changes do get slower due to this thread-safe `HitsThresholdChecker` (confirmed by playing with the atomic long inside a bit). Here is the full benchmark result: ``` TaskQPS baseline StdDevQPS my_modified_version StdDev Pct diff p-value HighTermTitleBDVSort 52.86 (14.0%) 42.72 (6.6%) -19.2% ( -34% - 1%) 0.000 HighTermMonthSort 31.66 (10.7%) 25.90 (6.3%) -18.2% ( -31% - -1%) 0.000 HighTermDayOfYearSort 24.00 (15.7%) 20.02 (8.5%) -16.6% ( -35% - 9%) 0.000 TermDTSort 20.22 (9.0%) 17.13 (6.6%) -15.3% ( -28% - 0%) 0.000 MedTerm 1270.02 (3.9%) 1208.29 (5.1%) -4.9% ( -13% - 4%) 0.001 HighTerm 941.14 (4.0%) 900.05 (5.9%) -4.4% ( -13% - 5%) 0.006 OrHighNotMed 531.63 (5.6%) 514.68 (6.2%) -3.2% ( -14% - 9%) 0.089 LowTerm 1112.50 (5.1%) 1081.48 (6.2%) -2.8% ( -13% - 8%) 0.118 MedSpanNear 12.50 (2.2%) 12.18 (2.2%) -2.5% ( -6% - 1%) 0.000 HighSpanNear 12.82 (1.7%) 12.52 (1.7%) -2.4% ( -5% - 1%) 0.000 OrHighNotLow 562.50 (6.9%) 550.15 (5.6%) -2.2% ( -13% - 11%) 0.269 OrHighNotHigh 577.10 (5.6%) 565.51 (6.7%) -2.0% ( -13% - 10%) 0.303 OrNotHighHigh 633.91 (6.7%) 621.85 (6.0%) -1.9% ( -13% - 11%) 0.345 HighIntervalsOrdered 6.94 (3.1%) 6.81 (3.2%) -1.9% ( -7% - 4%) 0.061 PKLookup 174.06 (4.2%) 171.17 (4.5%) -1.7% ( -9% - 7%) 0.230 OrHighLow 283.56 (5.3%) 279.05 (5.8%) -1.6% ( -12% - 10%) 0.365 MedIntervalsOrdered 19.00 (3.7%) 18.74 (3.6%) -1.4% ( -8% - 6%) 0.237 Fuzzy1 59.38 (11.6%) 58.65 (12.0%) -1.2% ( -22% - 25%) 0.740 HighSloppyPhrase 12.68 (2.7%) 12.53 (2.1%) -1.1% ( -5% - 3%) 0.135 Fuzzy2 63.95 (8.6%) 63.40 (9.0%) -0.9% ( -17% - 18%) 0.759 MedSloppyPhrase 17.53 (2.7%) 17.38 (2.2%) -0.8% ( -5% - 4%) 0.279 LowSloppyPhrase 23.90 (3.0%) 23.70 (2.8%) -0.8% ( -6% - 5%) 0.354 Respell 39.23 (3.1%) 38.90 (3.1%) -0.8% ( -6% - 5%) 0.403 AndHighMed 61.17 (3.8%) 60.72 (3.8%) -0.7% ( -8% - 7%) 0.542 OrHighMed 48.81 (3.0%) 48.46 (3.0%) -0.7% ( -6% - 5%) 0.444 LowIntervalsOrdered 14.64 (2.3%) 14.55 (2.3%) -0.6% ( -5% - 4%) 0.380 LowSpanNear 11.65 (1.6%) 11.57 (1.6%) -0.6% ( -3% - 2%) 0.223 AndHighHigh 16.57 (3.6%) 16.46 (4.0%) -0.6% ( -7% - 7%) 0.608 IntNRQ 17.25 (28.0%) 17.17 (27.3%) -0.5% ( -43% - 76%) 0.959 OrHighHigh 9.28 (2.5%) 9.24 (2.5%) -0.4% ( -5% - 4%) 0.635 MedPhrase 72.99 (3.3%) 72.75 (3.7%) -0.3% ( -7% - 6%) 0.763 LowPhrase 78.02 (3.5%) 77.77 (4.0%) -0.3% ( -7% - 7%) 0.786 BrowseMonthSSDVFacets 3.19 (1.8%) 3.19 (1.7%) -0.2% ( -3% - 3%) 0.653 BrowseDayOfYearSSDVFacets 3.01 (1.5%) 3.01 (1.2%) -0.0% ( -2% - 2%) 0.991 Wildcard 27.98 (2.8%) 27.98 (2.7%) 0.0% ( -5% - 5%) 0.981 AndHighLow 454.42 (5.4%) 455.10 (4.6%) 0.1% ( -9% - 10%) 0.925 HighPhrase 289.77 (5.1%) 290.54 (4.0%) 0.3% ( -8% - 9%) 0.854 OrNotHighLow 521.43 (6.5%) 523.38 (5.7%) 0.4% ( -11% - 13%) 0.846 Prefix3 39.82 (4.8%) 40.02 (3.1%) 0.5% ( -7% - 8%) 0.707 OrNotHighMed 560.49 (5.8%) 563.51 (5.5%) 0.5% ( -10% - 12%) 0.762 BrowseMonthTaxoFacets 0.99 (4.8%) 1.00 (4.6%) 0.6% ( -8% - 10%) 0.702 BrowseDateTaxoFacets 0.94 (4.6%) 0.94 (4.6%) 0.6% ( -8% - 10%) 0.683 BrowseDayOfYearTaxoFacets 0.94 (4.5%) 0.94 (4.5%) 0.7% ( -7% - 10%) 0.638 ``` To fix the performance regression, but still try to keep the logic simple then, I'm considering adding an additional flag `supportsConcurrency` to the collector managers' constructor like such: ``` public TopScoreDocCollectorManager(int numHits, ScoreDoc after, int totalHitsThreshold, boolean supportsConcurrency) { ... this.hitsThresholdChecker = supportsConcurrency ? HitsThresholdChecker.createShared(Math.max(totalHitsThreshold, numHits)) : HitsThresholdChecker.create(Math.max(totalHitsThreshold, numHits)); this.minScoreAcc = supportsConcurrency ? new MaxScoreAccumulator() : null; } public TopFieldCollectorManager(Sort sort, int numHits, FieldDoc after, int totalHitsThreshold, boolean supportsConcurrency) { ... this.hitsThresholdChecker = supportsConcurrency ? HitsThresholdChecker.createShared(Math.max(totalHitsThreshold, numHits)) : HitsThresholdChecker.create(Math.max(totalHitsThreshold, numHits)); this.minScoreAcc = supportsConcurrency ? new MaxScoreAccumulator() : null; } ``` What do you think about this approach? -- 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-unsubscr...@lucene.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org