andygrove commented on PR #11482:
URL: https://github.com/apache/arrow-rs/pull/11482#issuecomment-6950897200

   Thanks for the measurements, @pierre-warnier. b5d5f87f2 makes the same jiff 
1.0 changes you describe, including the allowlist growing to 49 zones.
   
   I reran the bench on an AMD Ryzen 9 7950X3D, pinned to one core, and updated 
the Benchmarks section of the description with the results.
   
   - `date_part` before 2038 is within 1% in `America/Los_Angeles` and 
`Europe/London` here, so the slowdown you saw on the i9-14900F does not 
reproduce on this machine.
   - The string cast before 2038 is 9% to 10% slower in those zones, so that 
part does reproduce.
   
   On the `TzOffset` size guess, I tried an 8-byte `Tz` and a 12-byte 
`TzOffset`, the same sizes as with chrono-tz, using a `u16` index into the zone 
table. It measured the same, and it made the cast in `Asia/Tokyo` 20% slower, 
so I have not pushed it. Aligning every function to 64 bytes in both builds did 
not change the result either.
   
   A profile puts all of the extra time in chrono's `write_rfc3339`, which runs 
the same code on the same inputs with either engine. The jiff build runs about 
10% fewer instructions per row, with fewer branch misses, but at a lower IPC. I 
have not found the cause yet.
   
   I agree the gate on #9183 should include kernel-level rows for 1970 to 2038, 
since a lookup-only gate would pass today.
   


-- 
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