timestamp: Thu 2011-12-01 15:12:10 +0100
Fix for Bug#13430436 PERFORMANCE DEGRADATION IN SYSBENCH ON INNODB DUE TO ICP
When running sysbench on InnoDB there is a performance degradation due
to index condition pushdown (ICP). Several of the queries in sysbench
have a WHERE condition that the optimizer uses for executing these
queries as range scans. The upper and lower limit of the range scan
will ensure that the WHERE condition is fulfilled. Still, the WHERE
condition is part of the queries' condition and if ICP is enabled the
condition will be pushed down to InnoDB as an index condition.
Due to the range scan's upper and lower limits ensure that the WHERE
condition is fulfilled, the pushed index condition will not filter out
any records. As a result the use of ICP for these queries results in a
performance overhead for sysbench. This overhead comes from using
resources for determining the part of the condition that can be pushed
down to InnoDB and overhead in InnoDB for executing the pushed index
condition.
With the default configuration for sysbench the range scans will use
the primary key. This is a clustered index in InnoDB. Using ICP on a
clustered index provides the lowest performance benefit since the
entire record is part of the clustered index and in InnoDB it has the
highest relative overhead for executing the pushed index condition.
The fix for removing the overhead ICP introduces when running sysbench
is to disable use of ICP when the index used by the query is a
clustered index.
When WL#6061 is implemented this change should be re-evaluated.
Due to this bug the function SEL_IMERGE::or_sel_tree_with_checks()
could build an inconsistent merge tree if one of the SEL_TREEs in the
resulting index merge happened to contain a full key range.
This could trigger an assertion failure.
The function key_and() erroneously called SEL_ARG::increment_use_count()
when SEL_ARG::incr_refs() should had been called. This could lead to
wrong values of use_count for some SEL_ARG trees.
A bug in the code of the function key_or could lead to a situation
when performing of an OR operation for one index changes the result
the operation for another index. This bug is fixed with this patch.
Also corrected the specification and the code of the function
or_sel_tree_with_checks.
revno: 2876.47.174
revision-id: jorgen.loland@oracle.com-20110519120355-qn7eprkad9jqwu5j
parent: mayank.prasad@oracle.com-20110518143645-bdxv4udzrmqsjmhq
committer: Jorgen Loland <jorgen.loland@oracle.com>
branch nick: mysql-trunk-11765831
timestamp: Thu 2011-05-19 14:03:55 +0200
message:
BUG#11765831: 'RANGE ACCESS' MAY INCORRECTLY FILTER
AWAY QUALIFYING ROWS
The problem was that the ranges created when OR'ing two
conditions could be incorrect. Without the bugfix,
"I <> 6 OR (I <> 8 AND J = 5)" would create these ranges:
"NULL < I < 6",
"6 <= I <= 6 AND 5 <= J <= 5",
"6 < I < 8",
"8 <= I <= 8 AND 5 <= J <= 5",
"8 < I"
While the correct ranges is
"NULL < I < 6",
"6 <= I <= 6 AND 5 <= J <= 5",
"6 < I"
The problem occurs when key_or() ORs
(1) "NULL < I < 6, 6 <= I <= 6 AND 5 <= J <= 5, 6 < I" with
(2) "8 < I AND 5 <= J <= 5"
The reason for the bug is that in key_or(), SEL_ARG *tmp is
used to point to the range in (1) above that is merged with
(2) while key1 points to the root of the red-black tree of
(1). When merging (1) and (2), tmp refers to the "6 < I"
part whereas the root is the "6 <= ... AND 5 <= J <= 5" part.
key_or() decides that the tmp range needs to be split into
"6 < I < 8, 8 <= I <= 8, 8 < I", in which next_key_part of the
second range should be that of tmp. However, next_key_part is
set to key1->next_key_part ("5 <= J <= 5") instead of
tmp->next_key_part (empty). Fixing this gives the correct but
not optimal ranges:
"NULL < I < 6",
"6 <= I <= 6 AND 5 <= J <= 5",
"6 < I < 8",
"8 <= I <= 8",
"8 < I"
A second problem can be seen above: key_or() may create
adjacent ranges that could be replaced with a single range.
Fixes for this is also included in the patch so that the range
above becomes correct AND optimal:
"NULL < I < 6",
"6 <= I <= 6 AND 5 <= J <= 5",
"6 < I"
Merging adjacent ranges like this gives a slightly lower cost
estimate for the range access.
- Set the default
- Adjust the testcases so that 'new' tests are run with optimizations turned on.
- Pull out relevant tests from "irrelevant" tests and run them with optimizations on.
- Run range.test and innodb.test with both mrr=on and mrr=off
- "Using MRR" is no longer shown with range access.
- Instead, both range and BKA accesses will show one of the following:
= "Rowid-ordered scan"
= "Key-ordered scan"
= "Key-ordered Rowid-ordered scan"
depending on whether DS-MRR implementation will do scan keys in order, rowids in order,
or both.
- The patch also introduces a way for other storage engines/MRR implementations to
pass information to EXPLAIN output about the properties of employed MRR scans.
Also:
Changed the value of TIME_FOR_COMPARE_ROWID to make it the same as for MWL 21.
Changed some queries in range_vs_index_merge.test to make them generate
the same plans as earlier.
When ORing two AND-OR formulas the range optimizer could miss
a conjunct in one of AND-OR formulas in the result. If the
index merge union plan to access a table is formed by this
formula that, in general, is not inferred from the original
where/on condition,the query could return an incorrect result set.
A crash may happenin the cases when the range optimizer tried to OR
two index merge such that the second one contained less range trees
than the first one.
The bug was introduced by the patch of MWL#24:
"index_merge: fair choice between index_merge union and range access".
A crash may happenin the cases when the range optimizer tried to OR
two index merge such that the second one contained less range trees
than the first one.
The bug was introduced by the patch of MWL#24:
"index_merge: fair choice between index_merge union and range access".
that blocked building index merge plans for the queries with where
conditions of the form
(key1|2_p1=c AND range(key1_p2)) OR (key1|2_p1=c AND range(key2_p2)).
The problem was discovered by Sergey Petrunia when he reviewed the patch
for WL#24.
Added new test cases. One of them failed to produce an index merge plan
before the patch was applied.
"index_merge: fair choice between index_merge union and range access"
mysql-test/include/world.inc:
A new include file to upload the world database.
mysql-test/include/world_schema.inc:
A new include file to create tables of the world database.
mysql-test/r/index_merge_myisam.result:
The results for test cases testing the optimizations added in WL#24 for MyISAM.
mysql-test/r/range_vs_index_merge.result:
The results for test cases testing the optimizations added in WL#24 for InnoDB.
mysql-test/t/range_vs_index_merge.test:
Test cases to test the optimizations added in WL#24 for MyISAM.
mysql-test/t/range_vs_index_merge_innodb.test:
Test cases to test the optimizations added in WL#24 for InnoDB.
sql/sql_list.h:
Fixed a bug that did not allow adding a non-empty list to an empty list.
