Linux 4.2 Will Tweak The CFQ Scheduler For SSDs To Offer Better Performance

[zeroepoch]

It would be nice to see comparisons with deadline. That's what I've been using so far based on Michael's benchmarking.

[arti]

[COLOR=#000000][roberth@Desktop ~]$ sudo echo "0" > /sys/devices/pci0000:00/0000:00:1f.2/ata5/host4/target4:0:0/4:0:0:0/block/sda/queue/iosched/slice_idle[/COLOR]
An error occurred while redirecting file '/sys/devices/pci0000:00/0000:00:1f.2/ata5/host4/target4:0:0/4:0:0:0/block/sda/queue/iosched/slice_idle' open: Permission denied

EDIT: Okay maybe not use sudo for that. But how do I make it an udev rule so it gets renable on restarts?

the trick is to do something like this

$ echo 0 > sudo tee /file/somewhere/that/only/root/can/write/to

because `sudo echo 0 > file` will run echo as root but that stdout redirection part is ran by your shell that is still running as your user

[xeekei]

I use CFQ for mechanical drives and deadline for SSDs right now. I also use deadline for XFS even for mechanical drives since that's what the XFS wiki recommended.

[uid313]

Why not just the deadline or noop scheduler?

[gens]

Why not just the deadline or noop scheduler?

'cuz its cloud
cloud needs cgroups IO limiting that needs CFQ

[chilek]

But how do I make it an udev rule so it gets renable on restarts?

/etc/udev/rules.d/40-ssd.rules:

[COLOR=#222222][FONT=verdana]ACTION=="add|change", KERNEL=="sd[a-z]", ATTR{queue/rotational}=="0", ATTR{queue/iosched/[/FONT][/COLOR]slice_idle[COLOR=#222222][FONT=verdana]}="0"[/FONT][/COLOR]

[liam]

[COLOR=#000000][roberth@Desktop ~]$ sudo echo "0" > /sys/devices/pci0000:00/0000:00:1f.2/ata5/host4/target4:0:0/4:0:0:0/block/sda/queue/iosched/slice_idle[/COLOR]
An error occurred while redirecting file '/sys/devices/pci0000:00/0000:00:1f.2/ata5/host4/target4:0:0/4:0:0:0/block/sda/queue/iosched/slice_idle' open: Permission denied

EDIT: Okay maybe not use sudo for that. But how do I make it an udev rule so it gets renable on restarts?

Use /etc/tmpfiles.d/ instead of udev. I'm assuming you're using a systemd based distro.

[Guest]

Is this option available and ready to use in 3.16?

For ssd's, deadline io scheduler seemed to always give better performance, I think that kernel option (grub commandline) is 'elevator=deadline' or alternatively one can disable io scheduler completely by 'elevator=noop'. While CFQ can make sense in heavily multi-tasked io (or on hdds), most workloads on ssd's don't really make too much use of it.

[rdnetto]

It would be nice to see comparisons with deadline. That's what I've been using so far based on Michael's benchmarking.

It'd also be interesting to see if CFQ has reached parity with BFQ yet. BFQ was significantly better than CFQ (and even deadline) last time it was benchmarked, but it was never mainlined. (Despite this, Sabayon has it as their default scheduler.) That was a while ago though (3.14), so it's possible CFQ has caught up.

[GreatEmerald]

It'd also be interesting to see if CFQ has reached parity with BFQ yet. BFQ was significantly better than CFQ (and even deadline) last time it was benchmarked, but it was never mainlined. (Despite this, Sabayon has it as their default scheduler.) That was a while ago though (3.14), so it's possible CFQ has caught up.

Why hasn't it been mainlined, anyway... And this change is about throughput, it won't help latency, which is what BFQ does.