Re: [PATCH 02/20] blkio: Change CFQ to use CFS like queue time stamps

[Corrado Zoccolo <czoccolo@gmail.com>]

Hi Vivek,
On Wed, Nov 4, 2009 at 11:25 PM, Vivek Goyal <vgoyal@redhat.com> wrote:
> Thanks. I am looking at your patches right now. Got one question about
> following commit.
>
> ****************************************************************
> commit a6d44e982d3734583b3b4e1d36921af8cfd61fc0
> Author: Corrado Zoccolo <czoccolo@gmail.com>
> Date:   Mon Oct 26 22:45:11 2009 +0100
>
>    cfq-iosched: enable idling for last queue on priority class
>
>    cfq can disable idling for queues in various circumstances.
>    When workloads of different priorities are competing, if the higher
>    priority queue has idling disabled, lower priority queues may steal
>    its disk share. For example, in a scenario with an RT process
>    performing seeky reads vs a BE process performing sequential reads,
>    on an NCQ enabled hardware, with low_latency unset,
>    the RT process will dispatch only the few pending requests every full
>    slice of service for the BE process.
>
>    The patch solves this issue by always performing idle on the last
>    queue at a given priority class > idle. If the same process, or one
>    that can pre-empt it (so at the same priority or higher), submits a
>    new request within the idle window, the lower priority queue won't
>    dispatch, saving the disk bandwidth for higher priority ones.
>
>    Note: this doesn't touch the non_rotational + NCQ case (no hardware
>    to test if this is a benefit in that case).
> *************************************************************************
>
[snipping questions I answered in the combo mail]
> On top of that, even if we don't idle for RT reader, we will always
> preempt BE reader immediately and get the disk. The only side affect
> is that on rotational media, disk head might have moved and bring the
> overall throughput down.
You bring down throughput, and also increase latency,  not only on
rotational media, so you may not want to enable it on servers.
Without low_latency, I saw this bug in current 'fairness' policy in
CFQ, so this patch fixes it.
>
> So my concern is that with this idling on last queue, we are targetting
> fairness issue for the random seeky readers with thinktime with-in 8ms.
> That can be easily solved by setting low_latency=1. Why are we going
> to this lenth then?
Maybe on the servers where you want to run RT tasks you don't want the
aforementioned drawbacks of low_latency.
Since I was going to change the implications of low_latency in
following patches, I fixed the 'bug' here, so I was free to change the
implementation in the following, without reintroducing this bug (it
was present for long, before being fixed by the introduction of
low_latency).

Thanks
Corrado
>
> Thanks
> Vivek
>