[MCOL-5140] Improve min/max to use vectorized updates (AMD+ ARM) Created: 2022-06-16 Updated: 2022-08-10 Resolved: 2022-08-10 |
|
| Status: | Closed |
| Project: | MariaDB ColumnStore |
| Component/s: | PrimProc |
| Affects Version/s: | 6.3.1 |
| Fix Version/s: | 22.08.1 |
| Type: | New Feature | Priority: | Major |
| Reporter: | Roman | Assignee: | Daniel Lee (Inactive) |
| Resolution: | Fixed | Votes: | 0 |
| Labels: | performance | ||
| Sprint: | 2021-17 |
| Assigned for Review: | |
| Description |
|
The code in primitives/linux-port/column.cpp:vectorizedFiltering suffers calculating min/max values for a block in a scalar manner that takes place in vectWriteRIDValues/vectWriteColValues. According with my observations we can reduce the timings 20% down removing this deficiency. |
| Comments |
| Comment by Andrey Piskunov (Inactive) [ 2022-07-15 ] |
|
Done, after testing (on 10**6 values, for a 30 s period) the increase shown is from 3826 queries to 5612, about 45%. |
| Comment by Roman [ 2022-08-04 ] |
|
4QA this optimization should benefit query processing(that has filters on numbers or column types that are reduced to numbers) in general. The query example is mentioned in the description. |
| Comment by Daniel Lee (Inactive) [ 2022-08-10 ] |
|
Build verified: 22.08-1 (#5243) Closed by regression test |