TESTING: Optimize byte view comparison in multi groupby #19344
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Contributor
Author
|
I think there is more performance to be had by squeezing hashing for byteview |
Contributor
Author
|
run benchmarks |
Contributor
Yeah I think this could be a really nice optimization. |
Dandandan
reviewed
Dec 16, 2025
datafusion/common/src/hash_utils.rs
Outdated
| for (hash, &v) in hashes_buffer.iter_mut().zip(array.views().iter()) { | ||
| let view_len = v as u32; | ||
| // if the length is not inlined, then we need to hash the bytes as well | ||
| if view_len > 12 { |
Contributor
There was a problem hiding this comment.
Could also eliminate this branch when having no data buffers at all.
|
🤖 |
|
🤖: Benchmark completed Details
|
Contributor
|
run benchmark tpch |
alamb
changed the title
Contributor
Author
|
(BTW I am testing some other improvements to hash locally) My next plan is:
|
Contributor
Author
|
run benchmarks |
|
🤖 |
|
🤖: Benchmark completed Details
|
|
🤖 |
|
🤖: Benchmark completed Details
|
Contributor
Author
|
I broke the first part of this PR into Will put the hash improvements in their own PR |
Dandandan
reviewed
Dec 17, 2025
| return result; | ||
| result | ||
| } else { | ||
| self.do_equal_to_inner_values_only(lhs_row, array, rhs_row) |
Contributor
There was a problem hiding this comment.
perhaps this could be optimized as well (in the outer loop) for empty data buffers
Dandandan
reviewed
Dec 17, 2025
| ); | ||
|
|
||
| // if the input array has no nulls, can skip null check | ||
| for (&lhs_row, &rhs_row, equal_to_result) in iter { |
Contributor
There was a problem hiding this comment.
I wonder if we can get this to compile to a well-vectorized loop.
alamb
changed the title
Contributor
Author
|
I with the specialized hash function it gets even faster |
Contributor
Author
|
run benchmark tpch_mem |
Contributor
Author
|
show benchmark queue |
|
🤖 Hi @alamb, you asked to view the benchmark queue (#19344 (comment)).
|
|
Benchmark script failed with exit code 2. Last 10 lines of output: Click to expand |
1 similar comment
|
Benchmark script failed with exit code 2. Last 10 lines of output: Click to expand |
Contributor
Author
|
I have broken the changes in this PR into two chunks:
So closing this one down to work on them |
github-merge-queue bot
pushed a commit
that referenced
this pull request
Dec 20, 2025
…5% faster (#19413) ## Which issue does this PR close? <!-- We generally require a GitHub issue to be filed for all bug fixes and enhancements and this helps us generate change logs for our releases. You can link an issue to this PR using the GitHub syntax. For example `Closes #123` indicates that this PR will close issue #123. --> - Part of #18411 - Closes #19344 - Closes #19364 Note this is an alternate to #19364 ## Rationale for this change @camuel found a query where DuckDB's raw grouping is is faster. I looked into it and much of the difference can be explained by better vectorization in the comparisons and short string optimizations ## What changes are included in this PR? Optimize (will comment inline) ## Are these changes tested? By CI. See also benchmark results below. I tested manually as well Create Data: ```shell nice tpchgen-cli --tables=lineitem --format=parquet --scale-factor 100 ``` Run query: ```shell hyperfine --warmup 3 " datafusion-cli -c \"select l_returnflag,l_linestatus, count(*) as count_order from 'lineitem.parquet' group by l_returnflag, l_linestatus;\" " ``` Before (main): 1.368s ```shell Benchmark 1: datafusion-cli -c "select l_returnflag,l_linestatus, count(*) as count_order from 'lineitem.parquet' group by l_returnflag, l_linestatus;" Time (mean ± σ): 1.393 s ± 0.020 s [User: 16.778 s, System: 0.688 s] Range (min … max): 1.368 s … 1.438 s 10 runs ``` After (this PR) 1.022s ```shell Benchmark 1: ./datafusion-cli-multi-gby-try2 -c "select l_returnflag,l_linestatus, count(*) as count_order from 'lineitem.parquet' group by l_returnflag, l_linestatus;" Time (mean ± σ): 1.022 s ± 0.015 s [User: 11.685 s, System: 0.644 s] Range (min … max): 1.005 s … 1.052 s 10 runs ``` I have a PR that improves string view hashing performance too, see - #19374 ## Are there any user-facing changes? Faster performance
github-merge-queue bot
pushed a commit
that referenced
this pull request
Dec 20, 2025
## Which issue does this PR close? - builds on #19373 - part of #18411 - Broken out of #19344 - Closes #19344 ## Rationale for this change While looking at performance as part of #18411, I noticed we could speed up string view hashing by optimizing for small strings ## What changes are included in this PR? Optimize StringView hashing, specifically by using the inlined view for short strings ## Are these changes tested? Functionally by existing coverage Performance by benchmarks (added in #19373) which show * 15%-20% faster for mixed short/long strings * 50%-70% faster for "short" arrays where we know there are no strings longer than 12 bytes ``` utf8_view (small): multiple, no nulls 1.00 47.9±1.71µs ? ?/sec 4.00 191.6±1.15µs ? ?/sec utf8_view (small): multiple, nulls 1.00 78.4±0.48µs ? ?/sec 3.08 241.6±1.11µs ? ?/sec utf8_view (small): single, no nulls 1.00 13.9±0.19µs ? ?/sec 4.29 59.7±0.30µs ? ?/sec utf8_view (small): single, nulls 1.00 23.8±0.20µs ? ?/sec 3.10 73.7±1.03µs ? ?/sec utf8_view: multiple, no nulls 1.00 235.4±2.14µs ? ?/sec 1.11 262.2±1.34µs ? ?/sec utf8_view: multiple, nulls 1.00 227.2±2.11µs ? ?/sec 1.34 303.9±2.23µs ? ?/sec utf8_view: single, no nulls 1.00 71.6±0.74µs ? ?/sec 1.05 75.2±1.27µs ? ?/sec utf8_view: single, nulls 1.00 71.5±1.92µs ? ?/sec 1.28 91.6±4.65µs ``` <details><summary>Details</summary> <p> ``` Gnuplot not found, using plotters backend utf8_view: single, no nulls time: [20.872 µs 20.906 µs 20.944 µs] change: [−15.863% −15.614% −15.331%] (p = 0.00 < 0.05) Performance has improved. Found 13 outliers among 100 measurements (13.00%) 8 (8.00%) high mild 5 (5.00%) high severe utf8_view: single, nulls time: [22.968 µs 23.050 µs 23.130 µs] change: [−17.796% −17.384% −16.918%] (p = 0.00 < 0.05) Performance has improved. Found 7 outliers among 100 measurements (7.00%) 3 (3.00%) high mild 4 (4.00%) high severe utf8_view: multiple, no nulls time: [66.005 µs 66.155 µs 66.325 µs] change: [−19.077% −18.785% −18.512%] (p = 0.00 < 0.05) Performance has improved. utf8_view: multiple, nulls time: [72.155 µs 72.375 µs 72.649 µs] change: [−17.944% −17.612% −17.266%] (p = 0.00 < 0.05) Performance has improved. Found 11 outliers among 100 measurements (11.00%) 6 (6.00%) high mild 5 (5.00%) high severe utf8_view (small): single, no nulls time: [6.1401 µs 6.1563 µs 6.1747 µs] change: [−69.623% −69.484% −69.333%] (p = 0.00 < 0.05) Performance has improved. Found 6 outliers among 100 measurements (6.00%) 3 (3.00%) high mild 3 (3.00%) high severe utf8_view (small): single, nulls time: [10.234 µs 10.250 µs 10.270 µs] change: [−53.969% −53.815% −53.666%] (p = 0.00 < 0.05) Performance has improved. Found 5 outliers among 100 measurements (5.00%) 5 (5.00%) high severe utf8_view (small): multiple, no nulls time: [20.853 µs 20.905 µs 20.961 µs] change: [−66.006% −65.883% −65.759%] (p = 0.00 < 0.05) Performance has improved. Found 9 outliers among 100 measurements (9.00%) 7 (7.00%) high mild 2 (2.00%) high severe utf8_view (small): multiple, nulls time: [32.519 µs 32.600 µs 32.675 µs] change: [−53.937% −53.581% −53.232%] (p = 0.00 < 0.05) Performance has improved. Found 1 outliers among 100 measurements (1.00%) 1 (1.00%) high mild ``` </p> </details> ## Are there any user-facing changes? <!-- If there are user-facing changes then we may require documentation to be updated before approving the PR. --> <!-- If there are any breaking changes to public APIs, please add the `api change` label. -->
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Which issue does this PR close?
Rationale for this change
Testing some ideas to make gby faster
What changes are included in this PR?
Are these changes tested?
I benchmarked this manually like this
Main
This branch
Are there any user-facing changes?