Details
-
Improvement
-
Status: Closed
-
Minor
-
Resolution: Fixed
-
None
-
None
-
None
-
New
Description
While looking at LUCENE-10534, found that *FieldSource exists implementation after LUCENE-7407 can avoid value lookup when just checking for exists.
Flamegraphs - x axis = time spent as a percentage of time being profiled, y axis = stack trace bottom being first call top being last call
Looking only at the left most getValueForDoc highlight only (and it helps to make it bigger or download the original)
LongFieldSource#exists spends MOST of its time doing a LongFieldSource#getValueForDoc. LongFieldSource#getValueForDoc spends its time doing two things primarily:
- FilterNumericDocValues#longValue()
- advance()
This makes sense based on looking at the code (copied below to make it easier to see at once)
https://github.com/apache/lucene/blob/main/lucene/queries/src/java/org/apache/lucene/queries/function/valuesource/LongFieldSource.java#L72
private long getValueForDoc(int doc) throws IOException { if (doc < lastDocID) { throw new IllegalArgumentException( "docs were sent out-of-order: lastDocID=" + lastDocID + " vs docID=" + doc); } lastDocID = doc; int curDocID = arr.docID(); if (doc > curDocID) { curDocID = arr.advance(doc); } if (doc == curDocID) { return arr.longValue(); } else { return 0; } }
LongFieldSource#exists - doesn't care about the actual longValue. Just that there was a value found when iterating through the doc values.
https://github.com/apache/lucene/blob/main/lucene/queries/src/java/org/apache/lucene/queries/function/valuesource/LongFieldSource.java#L95
@Override public boolean exists(int doc) throws IOException { getValueForDoc(doc); return arr.docID() == doc; }
So putting this all together for exists calling getValueForDoc, we spent ~50% of the time trying to get the long value when we don't need it in exists. We can save that 50% of time making exists not care about the actual value and just return if doc == curDocID basically.
This 50% extra is exaggerated in MaxFloatFunction (and other places) since exists() is being called a bunch. Eventually the value will be needed from longVal(), but if we call exists() say 3 times for every longVal(), we are spending a lot of time computing the value when we only need to check for existence.
I found the same pattern in DoubleFieldSource, EnumFieldSource, FloatFieldSource, IntFieldSource, LongFieldSource. I put together a change showing what this would look like:
Simple JMH performance tests comparing the original FloatFieldSource to the new ones from PR #847.
Benchmark | Mode | Cnt | Score and Error | Units |
----------------------------------------------------------------- | ------- | ------------------ | ------- | |
MyBenchmark.testMaxFloatFunction | thrpt | 25 | 64.159 ± 2.031 | ops/s |
MyBenchmark.testNewMaxFloatFunction | thrpt | 25 | 94.997 ± 2.365 | ops/s |
MyBenchmark.testMaxFloatFunctionNewFloatFieldSource | thrpt | 25 | 123.191 ± 9.291 | ops/s |
MyBenchmark.testNewMaxFloatFunctionNewFloatFieldSource | thrpt | 25 | 123.817 ± 6.191 | ops/s |
MyBenchmark.testMaxFloatFunctionRareField | thrpt | 25 | 244.921 ± 6.439 | ops/s |
MyBenchmark.testNewMaxFloatFunctionRareField | thrpt | 25 | 239.288 ± 5.136 | ops/s |
MyBenchmark.testMaxFloatFunctionNewFloatFieldSourceRareField | thrpt | 25 | 271.521 ± 3.870 | ops/s |
MyBenchmark.testNewMaxFloatFunctionNewFloatFieldSourceRareField | thrpt | 25 | 279.334 ± 10.511 | ops/s |
Attachments
Attachments
Issue Links
- relates to
-
LUCENE-7407 Explore switching doc values to an iterator API
- Resolved
-
LUCENE-10534 MinFloatFunction / MaxFloatFunction calls exists twice
- Closed
- links to