IO_sched-speedtest.txt

This test script is designed to measure the performance of I/O
schedulers in blk or of blk-mq alone, where no I/O scheduler is
available yet.  The script automatically checks whether blk or blk-mq
is in use, and, in the first case, measures performance with all
available schedulers.

In more detail, the script measures the maximum number of IOPS
achievable with each scheduler (or only with blk-mq), and for each of
the following I/O types:
- Sequential read
- Sequential write
- Random read
- Random write

As for blk, by comparing the performance of a scheduler with that of
NOOP, it is possible to have an idea, by difference, of the overhead
entailed by the scheduler.

As for inner workings, the script generates, with fio, I/O for the
null_blk virtual device.  The latter is configured to simply receive
IOs as fast as possible and acknowledge their completion immediately.

In more detail, here are the parameters used for null_blk:
- queue_mode=1|2
- irqmode=0
- completion_nsec=0
- nr_devices=1

Further details can be found in the script itself.

Finally, we report below the output of the script for a system with
the SMP support disabled (to measure more precisely the overhead on a
CPU, and rule out issues related to parallel execution). As for the
hardware configuration, the system was equipped with a 2.3 GHz CPU and
16 GB of DD3 RAM at 1600 MHz. The version of BFQ used in the test is
v8r11.

Sample output:

Unit of measure: KIOPS			Time: 60s		Device: nullb0
Number of parallel threads: 1
SCHEDULER  SEQREAD  SEQWRITE  RANDREAD  RANDWRITE
noop       899K     484K      719K      468K
deadline   881K     488K      797K      458K
cfq        436K     375K      411K      359K
bfq        319K     242K      254K      214K

Unit of measure: KIOPS			Time: 60s		Device: nullb0
Number of parallel threads: 1
SCHEDULER  SEQREAD  SEQWRITE  RANDREAD  RANDWRITE
blk-mq     974K     543K      892K      511K