Re: [PATCH 12/35] writeback: scale down max throttle bandwidth on concurrent dirtiers

[Wu Fengguang <fengguang.wu@intel.com>]

On Tue, Dec 14, 2010 at 03:00:05PM +0800, Wu Fengguang wrote:
> On Tue, Dec 14, 2010 at 09:21:19AM +0800, Yan Zheng wrote:
> > On Mon, Dec 13, 2010 at 10:46 PM, Wu Fengguang <fengguang.wu@intel.com> wrote:
> > > This will noticeably reduce the fluctuaions of pause time when there are
> > > 100+ concurrent dirtiers.
> > >
> > > The more parallel dirtiers (1 dirtier => 4 dirtiers), the smaller
> > > bandwidth each dirtier will share (bdi_bandwidth => bdi_bandwidth/4),
> > > the less gap to the dirty limit ((C-A) => (C-B)), the less stable the
> > > pause time will be (given the same fluctuation of bdi_dirty).
> > >
> > > For example, if A drifts to A', its pause time may drift from 5ms to
> > > 6ms, while B to B' may drift from 50ms to 90ms.  It's much larger
> > > fluctuations in relative ratio as well as absolute time.
> > >
> > > Fig.1 before patch, gap (C-B) is too low to get smooth pause time
> > >
> > > throttle_bandwidth_A = bdi_bandwidth .........o
> > >                                              | o <= A'
> > >                                              |   o
> > >                                              |     o
> > >                                              |       o
> > >                                              |         o
> > > throttle_bandwidth_B = bdi_bandwidth / 4 .....|...........o
> > >                                              |           | o <= B'
> > > ----------------------------------------------+-----------+---o
> > >                                              A           B   C
> > >
> > > The solution is to lower the slope of the throttle line accordingly,
> > > which makes B stabilize at some point more far away from C.
> > >
> > > Fig.2 after patch
> > >
> > > throttle_bandwidth_A = bdi_bandwidth .........o
> > >                                              | o <= A'
> > >                                              |   o
> > >                                              |     o
> > >    lowered max throttle bandwidth for B ===> *       o
> > >                                              |   *     o
> > > throttle_bandwidth_B = bdi_bandwidth / 4 .............*   o
> > >                                              |       |   * o
> > > ----------------------------------------------+-------+-------o
> > >                                              A       B       C
> > >
> > > Note that C is actually different points for 1-dirty and 4-dirtiers
> > > cases, but for easy graphing, we move them together.
> > >
> > > Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
> > > ---
> > >  mm/page-writeback.c |   16 +++++++++++++---
> > >  1 file changed, 13 insertions(+), 3 deletions(-)
> > >
> > > --- linux-next.orig/mm/page-writeback.c 2010-12-13 21:46:14.000000000 +0800
> > > +++ linux-next/mm/page-writeback.c      2010-12-13 21:46:15.000000000 +0800
> > > @@ -587,6 +587,7 @@ static void balance_dirty_pages(struct a
> > >        unsigned long background_thresh;
> > >        unsigned long dirty_thresh;
> > >        unsigned long bdi_thresh;
> > > +       unsigned long task_thresh;
> > >        unsigned long long bw;
> > >        unsigned long period;
> > >        unsigned long pause = 0;
> > > @@ -616,7 +617,7 @@ static void balance_dirty_pages(struct a
> > >                        break;
> > >
> > >                bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh, nr_dirty);
> > > -               bdi_thresh = task_dirty_limit(current, bdi_thresh);
> > > +               task_thresh = task_dirty_limit(current, bdi_thresh);
> > >
> > >                /*
> > >                 * In order to avoid the stacked BDI deadlock we need
> > > @@ -638,14 +639,23 @@ static void balance_dirty_pages(struct a
> > >
> > >                bdi_update_bandwidth(bdi, start_time, bdi_dirty, bdi_thresh);
> > >
> > > -               if (bdi_dirty >= bdi_thresh || nr_dirty > dirty_thresh) {
> > > +               if (bdi_dirty >= task_thresh || nr_dirty > dirty_thresh) {
> > >                        pause = MAX_PAUSE;
> > >                        goto pause;
> > >                }
> > >
> > > +               /*
> > > +                * When bdi_dirty grows closer to bdi_thresh, it indicates more
> > > +                * concurrent dirtiers. Proportionally lower the max throttle
> > > +                * bandwidth. This will resist bdi_dirty from approaching to
> > > +                * close to task_thresh, and help reduce fluctuations of pause
> > > +                * time when there are lots of dirtiers.
> > > +                */
> > >                bw = bdi->write_bandwidth;
> > > -
> > >                bw = bw * (bdi_thresh - bdi_dirty);
> > > +               do_div(bw, bdi_thresh / BDI_SOFT_DIRTY_LIMIT + 1);
> > > +
> > > +               bw = bw * (task_thresh - bdi_dirty);
> > >                do_div(bw, bdi_thresh / TASK_SOFT_DIRTY_LIMIT + 1);
> > 
> > Maybe changing this line to "do_div(bw, task_thresh /
> > TASK_SOFT_DIRTY_LIMIT + 1);"
> > is more consistent.
> 
> I'll show you another consistency of "shape" :)
> 
> http://www.kernel.org/pub/linux/kernel/people/wfg/writeback/slides/light-dirtier-control-line.svg
> http://www.kernel.org/pub/linux/kernel/people/wfg/writeback/slides/heavy-dirtier-control-line.svg
> 
> In the above two figures, the overall control lines for light/heavy
> dirtier tasks have exactly the same shape -- it's merely shifted in
> the X axis direction. So the current form is actually more simple.

Sorry it's not the overall control lines that's simply shifted, but
the task control line.

bdi control line:
> > >                bw = bw * (bdi_thresh - bdi_dirty);
> > > +               do_div(bw, bdi_thresh / BDI_SOFT_DIRTY_LIMIT + 1);

task control line:
> > > +               bw = bw * (task_thresh - bdi_dirty);
> > >                do_div(bw, bdi_thresh / TASK_SOFT_DIRTY_LIMIT + 1);

The use of bdi_thresh in the last line makes sure all task control
lines are of the same slope.

Thanks,
Fengguang