Description:
In the build of 5.1.40, all runs of "stress.ddl_myisam" on i5os (32 bit) failed with a timeout.
Quoting one report:
=====
stress.ddl_myisam [ fail ] timeout after 900 seconds
Test ended at YYYY-MM-DD HH:MM:SS
Test case timeout after 900 seconds
== /PATH/mysql-test/var/log/ddl_myisam.log ==
CREATE INDEX IDX ON t1 (f2);
DROP INDEX IDX ON t1;
CREATE UNIQUE INDEX IDX ON t1 (f2);
DROP INDEX IDX ON t1;
# Subtest 8A (one connection, no PREPARE/EXECUTE)
# connection action
# default: CREATE INDEX IDX ON t1 (f2)
# default: SHOW CREATE TABLE t1
# default: DROP INDEX IDX ON t1
# default: CREATE UNIQUE INDEX IDX ON t1 (f2)
# default: SHOW CREATE TABLE t1
# default: DROP INDEX IDX ON t1
# Subtest 8B (one connection, use PREPARE/EXECUTE)
# connection action
# default: CREATE INDEX IDX ON t1 (f2)
# default: SHOW CREATE TABLE t1
# default: DROP INDEX IDX ON t1
# default: CREATE UNIQUE INDEX IDX ON t1 (f2)
# default: SHOW CREATE TABLE t1
# default: DROP INDEX IDX ON t1
== /PATH/mysql-test/var/tmp/analyze-timeout-mysqld.1.err ==
mysqltest: Could not open connection 'default' after 500 attempts: 2002 Can't connect to local MySQL server through socket '/PATH/mysqld.1.sock' (79)
=====
i5os (32 bit) is the only platform showing this problem,
but i5os (64 bit) is a very close follower:
community 898145
enterprise-commercial-advanced 837693
enterprise-commercial-classic 871122
enterprise-commercial-pro 870510
enterprise-gpl-pro.log 871359
So the distance to the timeout limit (900000) is very small.
For comparison, AIX is faster by a factor of (more than) 10:
32bit community 77692
enterprise-commercial-advanced 82412
enterprise-commercial-classic 85340
enterprise-commercial-pro 87008
enterprise-gpl-advanced 84280
enterprise-gpl-pro 86198
64bit community 85815
enterprise-commercial-advanced 81140
enterprise-commercial-classic 79759
enterprise-commercial-pro 84401
enterprise-gpl-advanced 84373
enterprise-gpl-pro 85128
How to repeat:
Run the test suite ...
Suggested fix:
Not sure what to do.
I might say "speed up i5os", but that is beyond our powers I fear.
Given that the 64 bit build passes the test in all configurations but has times so close to the limit, I do not think there is some real problem with the 32 bit build.
To me, it rather seems like the platform is slower for 32 bit than for 64 bit applications (I hope the machines are identical, hardware-wise).
I cannot provide history data of 5.1 tests on that platform, because we had a problem with socket name length there (or rather, with Perl checking it), so these tests were not run - and I don't have times.
However, I checked the AIX times of all 5.1.3* builds for it, and I don't see an increase, so I have no indication that the test was grown.
I still propose to check whether we can shrink or split the test.