Fedora Developers Are Trying To Figure Out The Best Linux I/O Scheduler

[phoronix]

Phoronix: Fedora Developers Are Trying To Figure Out The Best Linux I/O Scheduler

Fedora developers are working on trying to figure out the best default behavior moving forward for their I/O scheduler selection...

http://www.phoronix.com/scan.php?page=news_item&px=Fedora-Future-IO-Scheduler

[Candy]

At the end they'll come up with an own implementation made by the systemd guys.

[ms178]

There was a debate on the Kernel mailing list about who should actually decide on the default and which it should be. The Kernel developers should at least in theory have an information bonus as they are the ones developing and testing these schedulers in the first place. But appearently there is not a single best solution for all workloads and use cases and I wonder how this could be handled best.

[tildearrow]

It is best to have 2 default schedulers: one for classic hard disks and one for SSD's.

[starshipeleven]

At the end they'll come up with an own implementation made by the systemd guys.

It will be great.

[reavertm]

Or for server/database and desktop workflow.

[xorbe]

What are the commands for examining the current configuration, and changing the selection?

[Grogan]

What are the commands for examining the current configuration, and changing the selection?

dmesg | grep sched
io scheduler noop registered
io scheduler deadline registered
[B]io scheduler cfq registered (default)[/B]

Shows you the ones available (I don't build them all) and the "default" is the one in use unless overridden.

To change i/o schedulers, you do it on the kernel command line with the elevator=schedulername parameter

It can also be done per-disk, for example, for /dev/sda

echo 'schedulername' > /sys/block/sda/queue/scheduler

Where schedulername is one of the available i/o schedulers.

P.S. You can query that with:

cat /sys/block/sda/queue/scheduler

It will return something like:

noop deadline (cfq)

With the one in brackets being the one in use.

[Yae8xahch]

Or for server/database and desktop workflow.

The thing is not that easy, though...

We have...

A) one sequential read
B) several sequential reads
C) one non-sequential read
D) many non-sequential reads (that might turn out to be some As and Bs)
E) one sequential write (that might invalidate an A, B, C or D)
F) several sequential writes
G) one non-sequential write
H) many non-sequential writes (that might turn out trigger an E)

It's all not that easy...

Given what we have today in either a) choice and b) tunables given for each scheduler it's a good experience we are "complaining" about...

[Charlie68]

The great thing would be to give users a graphical tool where they can select what they want.
host image free