perf(array): optimize performance of literal expression

[wangrunji0408]

I hereby agree to the terms of the Singularity Data, Inc. Contributor License Agreement.

What's changed and what's your intention?

• introduce Array::raw_iter to efficiently iterate over raw values regardless of null.
• introduce ArrayBuilder::append_n to efficiently append the same value multiple times.
• remove the head from Bitmap

With these optimizations, the evaluation time of literal expressions dropped from 2.2us to 150ns.

This PR also adds an optimized version of i32 + i32 using raw iterator in the bench, reducing time from 3.3us to 350ns.

Checklist

• I have written necessary rustdoc comments
• I have added necessary unit tests and integration tests
• All checks passed in ./risedev check (or alias, ./risedev c)

Refer to a related PR or issue link (optional)

#6868

[BugenZhao]

🚀

[BowenXiao1999]

Great Job, LGTM!

1. How should use the append_n in executor (e.g. build a chunk) or is it only for bench? Usually we do not how many values are the same before.
2. I saw perf(expr): add initial microbenchmark for expressions #6856 has mentioned some potential improvements. Are all kinds of improvements has been includeded in this pr or there is more?

[st1page]

3.3us to 350us

350ns?

[wangrunji0408]

3.3us to 350us

350ns?

Sure. Thanks for pointing it out. 🥵

[wangrunji0408]

How should use the append_n in executor (e.g. build a chunk) or is it only for bench? Usually we do not how many values are the same before.

It is a method in ArrayBuilder trait. There's also a new method append_datum_n in ArrayBuilderImpl. So they should be generally available. But I think the most use case would still be evaluating literal expressions.

I saw #6856 has mentioned some potential improvements. Are all kinds of improvements has been includeded in this pr or there is more?

Yes, that's all.

[jon-chuang]

Although we demonstrated that we can improve the i32 add performance possibly by vectorizing each of the steps (null check, arithmetic and check overflow), can I ask whether any benchmark demonstrates improvement due to the use of append_n?

I guess the intention is simply to provide an additional method that we can benchmark the usage of in the future?

[wangrunji0408]

Although we demonstrated that we can improve the i32 add performance possibly by vectorizing each of the steps (null check, arithmetic and check overflow), can I ask whether any benchmark demonstrates improvement due to the use of append_n?

Sure. The bench "expr/constant" (literal expression). Previously it could only append the same value one by one. Now it can append them all at once.

[codecov]

Codecov Report

Merging #6877 (6d8c3a2) into main (c188e5f) will decrease coverage by 0.01%.
The diff coverage is 71.08%.

@@            Coverage Diff             @@
##             main    #6877      +/-   ##
==========================================
- Coverage   73.16%   73.15%   -0.02%     
==========================================
  Files        1033     1033              
  Lines      164925   165019      +94     
==========================================
+ Hits       120665   120714      +49     
- Misses      44260    44305      +45     

Flag	Coverage Δ
rust	73.15% <71.08%> (-0.02%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
src/common/src/array/iterator.rs	79.45% <0.00%> (-18.86%)	⬇️
src/expr/src/expr/expr_binary_nonnull.rs	62.60% <ø> (ø)
src/common/src/array/utf8_array.rs	90.76% <22.22%> (-2.52%)	⬇️
src/common/src/array/struct_array.rs	87.38% <53.84%> (-0.99%)	⬇️
src/common/src/array/bool_array.rs	90.40% <62.50%> (-2.69%)	⬇️
src/common/src/array/primitive_array.rs	87.00% <64.70%> (-2.48%)	⬇️
src/common/src/array/list_array.rs	91.43% <70.00%> (-0.71%)	⬇️
src/common/src/array/bytes_array.rs	84.37% <86.95%> (-0.71%)	⬇️
src/common/src/array/mod.rs	73.52% <100.00%> (+0.50%)	⬆️
src/common/src/buffer/bitmap.rs	95.94% <100.00%> (+0.12%)	⬆️
... and 6 more

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[BugenZhao]

LGTM!!!