[MCOL-4564] Performance issues for DBT3 queries #17, #18 Created: 2021-02-26 Updated: 2021-06-28 Resolved: 2021-03-03 |
|
| Status: | Closed |
| Project: | MariaDB ColumnStore |
| Component/s: | ExeMgr |
| Affects Version/s: | 5.5.2 |
| Fix Version/s: | 5.5.2 |
| Type: | Bug | Priority: | Blocker |
| Reporter: | Daniel Lee (Inactive) | Assignee: | Daniel Lee (Inactive) |
| Resolution: | Fixed | Votes: | 0 |
| Labels: | None | ||
| Issue Links: |
|
||||||||||||||||
| Sprint: | 2021-4 | ||||||||||||||||
| Description |
|
Build tested: 5.5.2 (Drone build #1718) Performed 10gb DBT3 performance on the build and noticed couple issues. #18 took 56 times longer to complete. Query execution time: 5.5.1-1 Disk run 5.5.1-1 Cache run 5.5.2-1 Disk run 5.5.2-1 Cache run Query #17 Query #18 |
| Comments |
| Comment by Roman [ 2021-03-01 ] |
|
Plz review. |
| Comment by Daniel Lee (Inactive) [ 2021-03-01 ] |
|
Build verified: 5.5.2 (Drone #1776) New elapsed time 5.5.2-1 Disk run 5.5.2-1 Cache run |
| Comment by Roman [ 2021-03-01 ] |
|
perf top output comparison b/w two MCS versions implicitly confirmed that the number of hash collisions is a way higher for 5.5.2 then for 5.5.1. AggregateComparator::operator() (this symbol is used by the hashmap used in GROUP BY processing) overhead was a way higher for 5.5.2 (5.5.2 uses CHARSET_INFO::hash_sort whilst 5.5.1 uses homebrew Murmur3 implementation) This tells me that hash values distributions produced by CHARSET_INFO::hash_sort are far from uniform distributions comparing with distrubutions produced by Murmur3 hash function. |
| Comment by Daniel Lee (Inactive) [ 2021-03-03 ] |
|
Build verified: 5.5.2 (Drone #1793) Verified again with the latest cron build. |