[MCOL-1882] Aggregates performance degradation between 1.1 and 1.2 Created: 2018-11-13  Updated: 2020-08-25  Resolved: 2018-11-14

Status: Closed
Project: MariaDB ColumnStore
Component/s: PrimProc
Affects Version/s: 1.2.1
Fix Version/s: Icebox

Type: Bug Priority: Major
Reporter: Roman Assignee: Andrew Hutchings (Inactive)
Resolution: Duplicate Votes: 0
Labels: None

Issue Links:
Duplicate
duplicates MCOL-1793 regr_slope() and regr_r2() produce in... Closed

 Description   

There is a major performance degradation between 1.2 and 1.1. Here is the test bed.

MariaDB [test]> desc t1;
+-------+--------+------+-----+---------+-------+
| Field | Type   | Null | Key | Default | Extra |
+-------+--------+------+-----+---------+-------+
| a     | double | YES  |     | NULL    |       |
| b     | double | YES  |     | NULL    |       |
+-------+--------+------+-----+---------+-------+
2 rows in set (0.00 sec)
MariaDB [test]> select count(*) from t1;
+-----------+
| count(*)  |
+-----------+
| 560023752 |
+-----------+
1 row in set (2.46 sec)

Version 1.1 timings for the query

MariaDB [test]> select max(a),max(b) from t1;
selec count(*) from t1;
+--------+--------+
| max(a) | max(b) |
+--------+--------+
|      1 |      1 |
+--------+--------+
1 row in set (14.66 sec)

1.2 timings for the query with the same hardware, standart lib, boost

MariaDB [test]> select max(a),max(b) from t1;
selec count(*) from t1;
+--------+--------+
| max(a) | max(b) |
+--------+--------+
|      1 |      1 |
+--------+--------+
1 row in set (27.53 sec)

According with the assembly transcript of the code the problem caused by excesive locking for atomic ref counter increase whilst running shared_pointer copy constructor in rowaggregation.cpp by RowAggregation::updateEntry().



 Comments   
Comment by Roman [ 2018-11-14 ]

Close this as David.Hall is fixing the issue in MCOL-1793.

Generated at Thu Feb 08 02:32:05 UTC 2024 using Jira 8.20.16#820016-sha1:9d11dbea5f4be3d4cc21f03a88dd11d8c8687422.