Linux 4.20 I/O Scheduler Benchmarks On NVMe SSD Storage

[jtoaspern]

Thanks for the NVME Benchnarks Michael, I was looking forward to them.

[dwagner]

Isn't CFQ still the default scheduler? If so, shouldn't it be one candidate for benchmarks of schedulers?

[Michael]

Isn't CFQ still the default scheduler? If so, shouldn't it be one candidate for benchmarks of schedulers?

As stated in the article, the default for NVMe SSD storage on Ubuntu and most distros is 'none'

[Almindor]

Default is none or deadline-mq. I just switched to bfq/lowlat and it seems to be making a visible impact. Just a feel tho, didn't run any benchmarks.

[paolo]

paolo

Remind me: Is BFQ compiled with low_latency by default or do I need to specify some kernel configuration flag before it's turned on?

Yep, I confirm what Micheal already wrote: low latency is no by default. That is, by default,
/sys/block/DEVICE/queue/iosched/low_latency = 1

The idea is that BFQ works well, also for throughput, even with low latency on. Results seem to show that we might be getting there.

[snadrus]

Is it possible to use BFQ_LL for reads but None for writes?

The benchmarks suggest such a mode would win the most.

[duby229]

Is it possible to use BFQ_LL for reads but None for writes?

The benchmarks suggest such a mode would win the most.

No, because IO schedulers are all about managing latency amd write access is still where most latency is.

[paolo]

No, because IO schedulers are all about managing latency amd write access is still where most latency is.

BTW, we are also working on an improvement for BFQ, which might boost writes rather remarkably ...

[gnartz]

Nice find. According to PTS, my startup times went down by a factor of 4..5 after switching from none to bfq-mq with low latency enabled. I'm using Intel Optane 900p under Ubuntu 18.04 with self-compiled liquorix 4.18.20 kernel.

[jpg44]

Isn't CFQ still the default scheduler? If so, shouldn't it be one candidate for benchmarks of schedulers?

CFQ from what I have seen is default for regular hard drives.