LWN: Comments on "Two new block I/O schedulers for 4.12"

Two new block I/O schedulers for 4.12

leodag — Mon, 05 Dec 2022 14:30:40 +0000

You can also find these devices listed under /sys/class/block with their more familiar names:

$ ls -l /sys/class/block
total 0
lrwxrwxrwx 1 root root 0 dez 5 11:22 dm-0 -> ../../devices/virtual/block/dm-0/
lrwxrwxrwx 1 root root 0 dez 5 11:22 nvme0n1 -> ../../devices/pci0000:00/0000:00:01.2/0000:02:00.0/nvme/nvme0/nvme0n1/
lrwxrwxrwx 1 root root 0 dez 5 11:22 nvme0n1p1 -> ../../devices/pci0000:00/0000:00:01.2/0000:02:00.0/nvme/nvme0/nvme0n1/nvme0n1p1/
lrwxrwxrwx 1 root root 0 dez 5 11:22 nvme0n1p2 -> ../../devices/pci0000:00/0000:00:01.2/0000:02:00.0/nvme/nvme0/nvme0n1/nvme0n1p2/

Two new block I/O schedulers for 4.12

dtlin — Thu, 11 May 2017 06:37:30 +0000

For reference, there is documentation at /usr/src/linux/Documentation/block/switching-sched.txt.

Two new block I/O schedulers for 4.12

micka — Wed, 10 May 2017 21:58:11 +0000

Thank you, this seems sensible. I could not find any hint about that in the article.

Two new block I/O schedulers for 4.12

zdzichu — Mon, 08 May 2017 10:32:40 +0000

Linux supports pluggable, per-device I/O schedulers; you can change them with `cat`:

$ find /sys -name scheduler -exec grep . {} +
/sys/devices/pci0000:00/0000:00:0b.0/ata1/host0/target0:0:0/0:0:0:0/block/sda/queue/scheduler: noop [deadline] cfq
/sys/devices/pci0000:00/0000:00:09.0/virtio3/host8/target8:0:0/8:0:0:2/block/sdc/queue/scheduler: noop deadline [cfq]
/sys/devices/pci0000:00/0000:00:0c.0/virtio5/block/vda/queue/scheduler: mq-deadline kyber [bfq] none

Two new block I/O schedulers for 4.12

micka — Mon, 08 May 2017 09:01:27 +0000

Do I understand correctly that you have only one such scheduler at any time? So if you have both slow and fast devices on a system, which one do you use?

difference between deadline & kyber

dud225 — Sat, 29 Apr 2017 03:58:53 +0000

This new scheduler seems to have similarities with deadline. As far as I understand kyber is able to dynamically adapt the size of the queues to serve requests at low latencies whereas deadline will be doing as much as it can.
Is that correct ? Are there other differences ?

Interaction with btrfs

masoncl — Wed, 26 Apr 2017 13:00:26 +0000

It should work pretty well, but we'll definitely look into any latency problems. For the most part we're able to tag latency sensitive IOs and get good results from the schedulers. On the btrfs side, we try to make sure the btrfs IO threads aren't creating priority inversions of their own.

Interaction with btrfs

DG — Tue, 25 Apr 2017 13:15:31 +0000

Unscientific observation : It works well.

I used to see spikes in latency from BTRFS on an internal backup server as it performed rsync --inplace based backups and deleted old snapshots ... which were bad enough to make me try BFQ.

Random details: vanilla 4.9.x kernel + BFQ. Two WD 3TiB Red disks in a BTRFS raid 1 filesystem. 4 core i5 processor with 32gb of RAM. FS contains 150 readonly snapshots. Currently using about 1.79TiB.

Interaction with btrfs

zdzichu — Tue, 25 Apr 2017 09:41:22 +0000

I wonder how those schedulers are going to interact with btrfs? Those schedulers try to ensure fairness between *processes*. But in my observations, when using btrfs most work seem to be dispatched from kworker, btrfs-worker, btrfs-cleaner, btrfs-transaction and other btrfs-* kernel threads, not from original processes.
This btrfs behaviour makes observing I/O traffic troublesome already. I fear it will wreak havoc with fairness.

Two new block I/O schedulers for 4.12

liam — Tue, 25 Apr 2017 00:38:18 +0000

http://marc.info/?l=linux-block&m=148978871820916&...

Latency falls from 8ms to 1ms (the target set by kyber).

Two new block I/O schedulers for 4.12

jhoblitt — Mon, 24 Apr 2017 19:49:17 +0000

Are there benchmarks for kyber? A quick search didn't turn up anything relevant.