LWN: Comments on "I/O scheduling for single-queue devices"

I/O scheduling for single-queue devices

timokokk — Tue, 16 Oct 2018 12:27:44 +0000

Seems that there is only one list for IO schedulers, both normal and multi-queue schedulers are put on the same list. You can only choose default for the normal schedulers, not for the MQ ones. So apparently the default for mq-scheduler is thus nop as there is no way to set it up at compile time.

I/O scheduling for single-queue devices

josh — Mon, 15 Oct 2018 15:11:00 +0000

$ cat /sys/block/nvme0n1/queue/scheduler
[none] mq-deadline

I/O scheduling for single-queue devices

timokokk — Mon, 15 Oct 2018 12:30:56 +0000

For me it shows as:

$ cat /sys/block/nvme0n1/queue/scheduler
[none] mq-deadline kyber bfq

Running still 4.18.1

I/O scheduling for single-queue devices

jan.kara — Mon, 15 Oct 2018 09:26:13 +0000

What does your /sys/block/<dev>/queue/scheduler show?

I/O scheduling for single-queue devices

linusw — Sat, 13 Oct 2018 19:47:10 +0000

I am sorry that I (Linus Walleij) have not posted a v2 yet, but it's on my laptop and it enforces mq-deadline on SMR devices while still letting single queue devices use BFQ by default. I will post it next week or so, just waiting for the debate to conclude. I am anyways going to maintain that patch for myself, so who knows, I might go and lobby it into a few distributions where it makes sense, if that is what it takes. Blessed are the meek, for they shall inherit the earth.

I/O scheduling for single-queue devices

marcH — Sat, 13 Oct 2018 19:33:54 +0000

> That's not an entirely realistic assessment of a lot of practical embedded development - while people *can* go in and tweak things to their heart's content and some will have the time to do that there's a lot of small teams pulling together entire systems who rely fairly heavily on defaults, focusing most of their effort on the bits of code they directly wrote.

OK, but do we know at least how the small, "early adopters" minority configure their systems and why?

Changing a default setting without *any* prior, "from the field" experience would be wrong. Sorry if I missed something.

I/O scheduling for single-queue devices

josh — Sat, 13 Oct 2018 17:46:53 +0000

For a device that does have multiple queues, does the kernel use "none" or "mq-deadline" by default? I have a recent NVMe drive, and the kernel seems to default to "none"; I'm wondering if that's the right default.

I/O scheduling for single-queue devices

post-factum — Fri, 12 Oct 2018 17:50:15 +0000

> Oleksandr Natalenko was not impressed by the embedded-systems argument

Just for the record, that Oleksandr Natalenko guy (aka me) is using BFQ since the very early days of its existing (2009, I believe? BFQ was definitely merged into pf-kernel for v2.6.31, at least) and has nothing against its proven superiority. It is just the automated choice for everyone that is questionable.