Re: [PATCH RFC 09/22] block, cfq: replace CFQ with the BFQ-v0 I/O scheduler

[Linus Walleij]

On Thu, Apr 14, 2016 at 6:29 PM, Tejun Heo <tj@xxxxxxxxxx> wrote:
> On Thu, Apr 14, 2016 at 12:23:14PM +0200, Paolo Valente wrote:
> ...
>> >> 3) Stable(r) and low latency for soft real-time applications
(...)
>> Goals 2-4 are obtained by granting a higher share of the throughput
>> to the applications to privilege.
(...)
> 4) makes sense as a lot of that workload would be at least
> quasi-sequential but I can't tell why 2) and 3) would depend on
> bandwidth based scheduling.  They're about recognizing workloads which
> can benefit from low latency and treating them accordingly.  Why would
> making the underlying scheduling time based change that?

For (3) Isn't that fairly intuitive?

Media people (think vlc, mplayer, gstreamer and whatnot) have
a framed media format, and need the following to process it
and guarantee an enjoyable media stream with audio/video:

- a certain number of bits/s from the storage device or network

- a certain number of MIPS or FLOPS or whatnot from the
  CPU before a certain deadline (a determinate time slice of
  computation time)

What they don't need is an allocated time slice on a block
device. So the BFQ is asking the right question
for that kind of applications: how many bits/s can I get,
and it delivers accordingly.

For the record I have successfully reproduced Paulo's reported
benefits from this test case:
https://github.com/Algodev-github/S/blob/master/video_playing_vs_commands/video_play_vs_comms.sh

Less frames *are* dropped!

And I have been booting BFQ kernels for my development laptop
for the last months simply because I like to have a music stream
running while working. With a stock kernel, my heavy git operations
(like git fetch && git reset --hard origin/master on linux-next)
makes the audio skip and hang for seconds. With the BFQ-patched
kernel it does *not* happen. This is subjective though, based
on intuition.

Yours,
Linus Walleij