mbutrovich commented on PR #9683:
URL: https://github.com/apache/arrow-rs/pull/9683#issuecomment-6947405540

   Changes since the benchmarks at 03ee81c, which the description now replaces 
with runs at 3bf1b4a:
   
   - cd3e438 handles rows that end within 12 bytes after the shared prefix. The 
key now holds only those bytes, after leading zeros, and the radix passes start 
after the zeros. Before, a short row took its last key bytes from the rows 
after it, so equal rows kept splitting on their neighbors' bytes. Against 
43d06ee, which adds the new cases, `[i32(0..10)]` is 3.29x to 3.82x faster at 
4096 rows, 2.55x to 2.87x at 32768, and 1.50x to 1.59x at 1M, and `[i32_opt]`, 
`[i64_opt]`, and `[f64_opt]` are 1.05x to 1.30x faster 
([Neoverse-V2](https://github.com/apache/arrow-rs/pull/9683#issuecomment-6937259227),
 
[aligned](https://github.com/apache/arrow-rs/pull/9683#issuecomment-6937255002);
 [Xeon](https://github.com/apache/arrow-rs/pull/9683#issuecomment-6937272774), 
[aligned](https://github.com/apache/arrow-rs/pull/9683#issuecomment-6937257527)).
   - cd3e438 also raised `FALLBACK_THRESHOLD` from 32 to 64, and 3bf1b4a puts 
it back. At 64, `[i32]` and `[i32, i32_opt]` at 8192 and 2M rows were 1.11x to 
1.16x faster on Neoverse-V2, but `[i32, i32_opt]` at 8192 rows was 0.80x to 
0.88x on the Xeon 
([Neoverse-V2](https://github.com/apache/arrow-rs/pull/9683#issuecomment-6935964651),
 
[aligned](https://github.com/apache/arrow-rs/pull/9683#issuecomment-6935941480);
 [Xeon](https://github.com/apache/arrow-rs/pull/9683#issuecomment-6935973525), 
[aligned](https://github.com/apache/arrow-rs/pull/9683#issuecomment-6935985471)).
   


-- 
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: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to