Perf: Optimize CursorValues compare performance for StringViewArray (1.4X faster for sort-tpch Q11) by zhuqi-lucas · Pull Request #16509 · apache/datafusion

zhuqi-lucas · 2025-06-23T10:05:10Z

Which issue does this PR close?

Closes#16508

Rationale for this change

Add fast path for CursorValues compare performance for StringViewArray.

What changes are included in this PR?

Add fast path for CursorValues compare performance for StringViewArray if the data buffer is empty, which can avoid len compare for following logic.

Are these changes tested?

Yes

Are there any user-facing changes?

No

zhuqi-lucas · 2025-06-23T10:05:27Z

Test result:

--------------------
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Query ┃ main ┃ fast_path_view ┃ Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ Q1 │ 106.56 ms │ 106.53 ms │ no change │
│ Q2 │ 107.52 ms │ 104.28 ms │ no change │
│ Q3 │ 666.52 ms │ 677.66 ms │ no change │
│ Q4 │ 128.45 ms │ 127.80 ms │ no change │
│ Q5 │ 258.79 ms │ 260.39 ms │ no change │
│ Q6 │ 271.88 ms │ 271.11 ms │ no change │
│ Q7 │ 427.24 ms │ 415.86 ms │ no change │
│ Q8 │ 305.82 ms │ 303.95 ms │ no change │
│ Q9 │ 316.23 ms │ 316.75 ms │ no change │
│ Q10 │ 470.93 ms │ 467.82 ms │ no change │
│ Q11 │ 281.10 ms │ 197.90 ms │ +1.42x faster │
└──────────────┴───────────┴────────────────┴───────────────┘

alamb · 2025-06-23T15:05:11Z

 self.views().len()
 }

+ #[inline(always)]


For anyone else following along, I believe this is the same optimization as applied in
Perf: Optimize comparison kernels for inlined views arrow-rs#7731

alamb · 2025-06-23T16:40:27Z

🤖 ./gh_compare_branch.shBenchmark Script Running
Linux aal-dev 6.11.0-1015-gcp #15~24.04.1-Ubuntu SMP Thu Apr 24 20:41:05 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
Comparing fast_path_view (f22ca3c) to 2bf8441diff
Benchmarks: tpch_mem clickbench_partitioned clickbench_extended
Results will be posted here when complete

alamb · 2025-06-23T17:21:01Z

🤖: Benchmark completed

Details

Comparing HEAD and fast_path_view
--------------------
Benchmark clickbench_extended.json
--------------------
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━┓
┃ Query ┃ HEAD ┃ fast_path_view ┃ Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━┩
│ QQuery 0 │ 1988.57 ms │ 2142.83 ms │ 1.08x slower │
│ QQuery 1 │ 718.95 ms │ 707.35 ms │ no change │
│ QQuery 2 │ 1381.41 ms │ 1385.85 ms │ no change │
│ QQuery 3 │ 648.24 ms │ 686.89 ms │ 1.06x slower │
│ QQuery 4 │ 1372.37 ms │ 1370.82 ms │ no change │
│ QQuery 5 │ 15173.88 ms │ 15217.15 ms │ no change │
│ QQuery 6 │ 2082.96 ms │ 2023.77 ms │ no change │
│ QQuery 7 │ 2049.36 ms │ 2028.40 ms │ no change │
│ QQuery 8 │ 805.76 ms │ 820.48 ms │ no change │
└──────────────┴─────────────┴────────────────┴──────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Benchmark Summary ┃ ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ Total Time (HEAD) │ 26221.51ms │
│ Total Time (fast_path_view) │ 26383.55ms │
│ Average Time (HEAD) │ 2913.50ms │
│ Average Time (fast_path_view) │ 2931.51ms │
│ Queries Faster │ 0 │
│ Queries Slower │ 2 │
│ Queries with No Change │ 7 │
│ Queries with Failure │ 0 │
└───────────────────────────────┴────────────┘
--------------------
Benchmark clickbench_partitioned.json
--------------------
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Query ┃ HEAD ┃ fast_path_view ┃ Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ QQuery 0 │ 2.26 ms │ 2.37 ms │ no change │
│ QQuery 1 │ 34.52 ms │ 33.99 ms │ no change │
│ QQuery 2 │ 82.25 ms │ 81.73 ms │ no change │
│ QQuery 3 │ 91.96 ms │ 101.59 ms │ 1.10x slower │
│ QQuery 4 │ 670.29 ms │ 595.46 ms │ +1.13x faster │
│ QQuery 5 │ 924.74 ms │ 848.52 ms │ +1.09x faster │
│ QQuery 6 │ 2.29 ms │ 2.23 ms │ no change │
│ QQuery 7 │ 39.12 ms │ 39.60 ms │ no change │
│ QQuery 8 │ 875.61 ms │ 869.89 ms │ no change │
│ QQuery 9 │ 1203.04 ms │ 1148.95 ms │ no change │
│ QQuery 10 │ 256.35 ms │ 258.61 ms │ no change │
│ QQuery 11 │ 290.28 ms │ 285.61 ms │ no change │
│ QQuery 12 │ 911.90 ms │ 874.02 ms │ no change │
│ QQuery 13 │ 1264.11 ms │ 1274.50 ms │ no change │
│ QQuery 14 │ 835.77 ms │ 826.75 ms │ no change │
│ QQuery 15 │ 788.25 ms │ 793.46 ms │ no change │
│ QQuery 16 │ 1661.81 ms │ 1673.49 ms │ no change │
│ QQuery 17 │ 1613.20 ms │ 1632.61 ms │ no change │
│ QQuery 18 │ 2995.68 ms │ 2969.08 ms │ no change │
│ QQuery 19 │ 88.59 ms │ 84.39 ms │ no change │
│ QQuery 20 │ 1171.81 ms │ 1139.65 ms │ no change │
│ QQuery 21 │ 1320.01 ms │ 1283.02 ms │ no change │
│ QQuery 22 │ 2171.26 ms │ 2095.38 ms │ no change │
│ QQuery 23 │ 7442.25 ms │ 7410.28 ms │ no change │
│ QQuery 24 │ 472.94 ms │ 453.60 ms │ no change │
│ QQuery 25 │ 402.89 ms │ 395.16 ms │ no change │
│ QQuery 26 │ 526.97 ms │ 514.73 ms │ no change │
│ QQuery 27 │ 1579.76 ms │ 1525.91 ms │ no change │
│ QQuery 28 │ 13130.51 ms │ 11906.74 ms │ +1.10x faster │
│ QQuery 29 │ 539.36 ms │ 514.08 ms │ no change │
│ QQuery 30 │ 800.47 ms │ 777.46 ms │ no change │
│ QQuery 31 │ 822.87 ms │ 818.04 ms │ no change │
│ QQuery 32 │ 2525.29 ms │ 2559.95 ms │ no change │
│ QQuery 33 │ 3219.31 ms │ 3228.55 ms │ no change │
│ QQuery 34 │ 3323.28 ms │ 3244.18 ms │ no change │
│ QQuery 35 │ 1282.15 ms │ 1269.19 ms │ no change │
│ QQuery 36 │ 122.16 ms │ 122.21 ms │ no change │
│ QQuery 37 │ 53.36 ms │ 49.92 ms │ +1.07x faster │
│ QQuery 38 │ 119.63 ms │ 119.59 ms │ no change │
│ QQuery 39 │ 203.97 ms │ 196.17 ms │ no change │
│ QQuery 40 │ 42.37 ms │ 42.41 ms │ no change │
│ QQuery 41 │ 38.85 ms │ 38.38 ms │ no change │
│ QQuery 42 │ 31.93 ms │ 33.52 ms │ 1.05x slower │
└──────────────┴─────────────┴────────────────┴───────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Benchmark Summary ┃ ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ Total Time (HEAD) │ 55975.43ms │
│ Total Time (fast_path_view) │ 54134.99ms │
│ Average Time (HEAD) │ 1301.75ms │
│ Average Time (fast_path_view) │ 1258.95ms │
│ Queries Faster │ 4 │
│ Queries Slower │ 2 │
│ Queries with No Change │ 37 │
│ Queries with Failure │ 0 │
└───────────────────────────────┴────────────┘
--------------------
Benchmark tpch_mem_sf1.json
--------------------
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━┓
┃ Query ┃ HEAD ┃ fast_path_view ┃ Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━┩
│ QQuery 1 │ 103.20 ms │ 103.84 ms │ no change │
│ QQuery 2 │ 21.07 ms │ 20.58 ms │ no change │
│ QQuery 3 │ 32.68 ms │ 32.00 ms │ no change │
│ QQuery 4 │ 18.60 ms │ 18.70 ms │ no change │
│ QQuery 5 │ 50.00 ms │ 48.53 ms │ no change │
│ QQuery 6 │ 11.86 ms │ 11.95 ms │ no change │
│ QQuery 7 │ 86.84 ms │ 88.45 ms │ no change │
│ QQuery 8 │ 25.11 ms │ 25.03 ms │ no change │
│ QQuery 9 │ 54.79 ms │ 54.82 ms │ no change │
│ QQuery 10 │ 43.38 ms │ 42.97 ms │ no change │
│ QQuery 11 │ 11.47 ms │ 11.33 ms │ no change │
│ QQuery 12 │ 34.12 ms │ 34.13 ms │ no change │
│ QQuery 13 │ 25.94 ms │ 26.54 ms │ no change │
│ QQuery 14 │ 9.78 ms │ 9.75 ms │ no change │
│ QQuery 15 │ 18.83 ms │ 19.84 ms │ 1.05x slower │
│ QQuery 16 │ 18.81 ms │ 18.78 ms │ no change │
│ QQuery 17 │ 96.22 ms │ 95.89 ms │ no change │
│ QQuery 18 │ 191.84 ms │ 195.57 ms │ no change │
│ QQuery 19 │ 24.92 ms │ 25.16 ms │ no change │
│ QQuery 20 │ 31.73 ms │ 31.82 ms │ no change │
│ QQuery 21 │ 144.91 ms │ 148.56 ms │ no change │
│ QQuery 22 │ 15.07 ms │ 15.09 ms │ no change │
└──────────────┴───────────┴────────────────┴──────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
┃ Benchmark Summary ┃ ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
│ Total Time (HEAD) │ 1071.19ms │
│ Total Time (fast_path_view) │ 1079.35ms │
│ Average Time (HEAD) │ 48.69ms │
│ Average Time (fast_path_view) │ 49.06ms │
│ Queries Faster │ 0 │
│ Queries Slower │ 1 │
│ Queries with No Change │ 21 │
│ Queries with Failure │ 0 │
└───────────────────────────────┴───────────┘

alamb · 2025-06-23T17:21:03Z

🤖 ./gh_compare_branch.shBenchmark Script Running
Linux aal-dev 6.11.0-1015-gcp #15~24.04.1-Ubuntu SMP Thu Apr 24 20:41:05 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
Comparing fast_path_view (f22ca3c) to 2bf8441diff
Benchmarks: sort_tpch
Results will be posted here when complete

alamb · 2025-06-23T17:22:23Z

🤖: Benchmark completed

Details

Comparing HEAD and fast_path_view
--------------------
Benchmark sort_tpch.json
--------------------
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Query ┃ HEAD ┃ fast_path_view ┃ Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ Q1 │ 350.37 ms │ 347.24 ms │ no change │
│ Q2 │ 301.59 ms │ 300.50 ms │ no change │
│ Q3 │ 1164.84 ms │ 1187.22 ms │ no change │
│ Q4 │ 412.28 ms │ 410.15 ms │ no change │
│ Q5 │ 414.43 ms │ 407.09 ms │ no change │
│ Q6 │ 445.56 ms │ 447.95 ms │ no change │
│ Q7 │ 902.09 ms │ 888.45 ms │ no change │
│ Q8 │ 819.94 ms │ 777.72 ms │ +1.05x faster │
│ Q9 │ 848.95 ms │ 801.57 ms │ +1.06x faster │
│ Q10 │ 1226.96 ms │ 1196.23 ms │ no change │
│ Q11 │ 750.22 ms │ 694.11 ms │ +1.08x faster │
└──────────────┴────────────┴────────────────┴───────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓
┃ Benchmark Summary ┃ ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩
│ Total Time (HEAD) │ 7637.22ms │
│ Total Time (fast_path_view) │ 7458.23ms │
│ Average Time (HEAD) │ 694.29ms │
│ Average Time (fast_path_view) │ 678.02ms │
│ Queries Faster │ 3 │
│ Queries Slower │ 0 │
│ Queries with No Change │ 8 │
│ Queries with Failure │ 0 │
└───────────────────────────────┴───────────┘

alamb

Thank you @zhuqi-lucas -- very nice

Dandandan · 2025-06-23T21:48:01Z

🤖: Benchmark completed
Details

had a look at the "regressions", I think should not be impacted by this change (thus noise).

zhuqi-lucas added 2 commits June 23, 2025 17:42

Perf: Optimize CursorValues compare performance for StringViewArray
838492d

Merge remote-tracking branch 'upstream/main' into fast_path_view
374b8be

github-actionsBot added the physical-plan Changes to the physical-plan crate label Jun 23, 2025

zhuqi-lucas changed the title ~~Perf: Optimize CursorValues compare performance for StringViewArray (1.6 faster for sort-tpch Q11)~~Perf: Optimize CursorValues compare performance for StringViewArray (1.4X faster for sort-tpch Q11)Jun 23, 2025

This was referenced Jun 23, 2025
Perf: Optimize CursorValues compare performance for StringViewArray #16508
Closed
Perf: Optimize comparison kernels for inlined views apache/arrow-rs#7731
Merged

fix
f22ca3c

alamb reviewed Jun 23, 2025
View reviewed changes

alamb approved these changes Jun 23, 2025
View reviewed changes

Dandandan approved these changes Jun 23, 2025
View reviewed changes

comphead merged commit 15a8738 into apache:mainJun 24, 2025

zhuqi-lucas mentioned this pull request Jul 8, 2025
Upgrade arrow/parquet to 56.0.0 #16690
Merged

alamb added the performance Make DataFusion faster label Jul 18, 2025

Navigation Menu

Uh oh!

Perf: Optimize CursorValues compare performance for StringViewArray (1.4X faster for sort-tpch Q11) - #16509

Perf: Optimize CursorValues compare performance for StringViewArray (1.4X faster for sort-tpch Q11)#16509
comphead merged 3 commits into
apache:mainfrom
zhuqi-lucas:fast_path_view

zhuqi-lucas commented Jun 23, 2025

Uh oh!

zhuqi-lucas commented Jun 23, 2025

Uh oh!

alambJun 23, 2025

Uh oh!

alamb commented Jun 23, 2025

Uh oh!

alamb commented Jun 23, 2025

Uh oh!

alamb commented Jun 23, 2025

Uh oh!

alamb commented Jun 23, 2025

Uh oh!

alamb left a comment

Uh oh!

Dandandan commented Jun 23, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

zhuqi-lucas commented Jun 23, 2025

Which issue does this PR close?

Rationale for this change

What changes are included in this PR?

Are these changes tested?

Are there any user-facing changes?

Uh oh!

zhuqi-lucas commented Jun 23, 2025

Uh oh!

alambJun 23, 2025

Choose a reason for hiding this comment

Uh oh!

alamb commented Jun 23, 2025

Uh oh!

alamb commented Jun 23, 2025

Uh oh!

alamb commented Jun 23, 2025

Uh oh!

alamb commented Jun 23, 2025

Uh oh!

alamb left a comment

Choose a reason for hiding this comment

Uh oh!

Dandandan commented Jun 23, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants