cat /sys/block/sda/queue/scheduler noop anticipatory [deadline] cfqThe load injector in this case is the DBMS_RESOURCE_MANAGER.CALIBRATE_IO. System is a two node RAC (with Dell EqualLogic SAN) running on RHEL 5(2.6.18-308.24.1.el5). Test was simply to set the elevator on the /etc/grub.conf (no configuration changes on the elevator parameters), reboot the server after each time elevator is set and once all nodes and DB instances are up run the CALIBRATE_IO function (same set of inputs were used in all cases). This was a new build and nothing else (application related) was running on the system. By no means this is an extensive test, just wanted to see what will be effect changing elevator alone (all others remaining the same).
Below is the output from the three runs.
ELEVATOR MAX_IOPS MAX_MBPS MAX_PMBPS LATENCY DURATION START_TIME END_TIME ---------- ---------- ---------- ---------- ---------- ------------ ------------------------------ ------------------------------ cfq 4970 206 171 11 7:58.585988 17-JAN-13 10.35.25.114131 AM 17-JAN-13 10.43.23.700119 AM deadline 4988 224 144 11 5:32.195376 17-JAN-13 01.34.14.094488 PM 17-JAN-13 01.39.46.289864 PM noop 5016 206 145 11 7:14.051605 17-JAN-13 02.05.07.657428 PM 17-JAN-13 02.12.21.709033 PM
It seems each elevator is a winner in some category.
MAX IOPS (winner noops)
MAX MBPS (winner deadline)
MAX PMBPS (winner cfq)