Re: [RFC v2] writeback: add elastic bdi in cgwb bdp

[Jan Kara <jack@suse.cz>]

On Fri 15-11-19 11:32:40, Hillf Danton wrote:
> 
> On Thu, 14 Nov 2019 13:17:46 +0100 Jan Kara wrote:
> > 
> > On Sat 26-10-19 18:46:56, Hillf Danton wrote:
> > > 
> > > The elastic bdi is the mirror bdi of spinning disks, SSD, USB and
> > > other storage devices/instruments on market. The performance of
> > > ebdi goes up and down as the pattern of IO dispatched changes, as
> > > approximately estimated as below.
> > > 
> > > 	P = j(..., IO pattern);
> > > 
> > > In ebdi's view, the bandwidth currently measured in balancing dirty
> > > pages has close relation to its performance because the former is a
> > > part of the latter.
> > > 
> > > 	B = y(P);
> > > 
> > > The functions above suggest there may be a layer violation if it
> > > could be better measured somewhere below fs.
> > > 
> > > It is measured however to the extent that makes every judge happy,
> > > and is playing a role in dispatching IO with the IO pattern entirely
> > > ignored that is volatile in nature.
> > > 
> > > And it helps to throttle the dirty speed, with the figure ignored
> > > that DRAM in general is x10 faster than ebdi. If B is half of P for
> > > instance, then it is near 5% of dirty speed, just 2 points from the
> > > figure in the snippet below.
> > > 
> > > /*
> > >  * If ratelimit_pages is too high then we can get into dirty-data overload
> > >  * if a large number of processes all perform writes at the same time.
> > >  * If it is too low then SMP machines will call the (expensive)
> > >  * get_writeback_state too often.
> > >  *
> > >  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
> > >  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
> > >  * thresholds.
> > >  */
> > > 
> > > To prevent dirty speed from running away from laundry speed, ebdi
> > > suggests the walk-dog method to put in bdp as a leash seems to
> > > churn less in IO pattern.
> > > 
> > > V2 is based on next-20191025.
> > 
> > Honestly, the changelog is still pretty incomprehensible as Andrew already
> > mentioned. Also I completely miss there, what are the benefits of this work
> > compared to what we currently have.
> > 
> Hey Jan
> 
> In the room which has been somewhere between 3% and 5% for bdp since
> 143dfe8611a6 ("writeback: IO-less balance_dirty_pages()") a bdp is
> proposed with target of surviving tests like LTP without regressions
> introduced, so overall the concerned benefit is that bdp is becoming
> more diverse if the diversity under linux/fs is good for the 99%.

What do you mean by "balance_dirty_pages() is becoming more diverse"?

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR