While running sysbench 0.4 "unit" tests using a separate client machine, I noticed that this particular test "simple ranges" runs much slower on MariaDB 5.5 than it does on MySQL 5.5. If I'm using original Innodb rather than builtin XtraDB, the speed is back and life is good.

This makes me suspect one of the optimizer changes that causes the slowdown.

The numbers in QPS from 15 minutes,64 users,1Mio rows test:

MySQL-5.5.25 (bzr) : 30372.36
MariaDB-5.5.24 (bzr) : 25081.54 <-- The slowdown
MariaDB-5.5.24/plugin : 31404.73

The "simple ranges" query in question is doing, according to the manual

SELECT c FROM sbtest WHERE id BETWEEN N AND M

(id is primary key, c is not a key)

I run sysbench 0.4 like this

sysbench --num-threads=64 --test=oltp --oltp-table-size=1000000 --oltp-dist-type=uniform --oltp-table-name=sbtest --init-rng=off --max-requests=0 --max-time=900 --mysql-user=root --mysql-port=3306 --mysql-host=<HOST> --mysql-db=test --myisam-max-rows=1000000 --mysql-table-engine=INNODB --oltp-point-selects=0 --oltp-simple-ranges=1 --oltp-sum-ranges=0 --oltp-order-ranges=0 --oltp-distinct-ranges=0 --oltp-skip-trx=on --oltp-read-only=on <MODE>

(where <MODE> is either prepare/run/clean, and <HOST> is IP address of the server where mysqld is running)

Note: I can also reproduce the effect when running sysbench locally on the same machine as mariadb server, however difference less noticable(because of noise sysbench client adds)