Re: [PATCH 01/10] mm: allow swappiness that prefers anon over file

[Minchan Kim]

On Wed, Jun 08, 2016 at 11:58:12AM -0400, Johannes Weiner wrote:
> On Wed, Jun 08, 2016 at 09:06:32AM +0900, Minchan Kim wrote:
> > On Tue, Jun 07, 2016 at 10:18:18AM -0400, Johannes Weiner wrote:
> > > On Tue, Jun 07, 2016 at 09:25:50AM +0900, Minchan Kim wrote:
> > > > On Mon, Jun 06, 2016 at 03:48:27PM -0400, Johannes Weiner wrote:
> > > > > --- a/Documentation/sysctl/vm.txt
> > > > > +++ b/Documentation/sysctl/vm.txt
> > > > > @@ -771,14 +771,20 @@ with no ill effects: errors and warnings on these stats are suppressed.)
> > > > >  
> > > > >  swappiness
> > > > >  
> > > > > -This control is used to define how aggressive the kernel will swap
> > > > > -memory pages.  Higher values will increase agressiveness, lower values
> > > > > -decrease the amount of swap.  A value of 0 instructs the kernel not to
> > > > > -initiate swap until the amount of free and file-backed pages is less
> > > > > -than the high water mark in a zone.
> > > > > +This control is used to define the relative IO cost of cache misses
> > > > > +between the swap device and the filesystem as a value between 0 and
> > > > > +200. At 100, the VM assumes equal IO cost and will thus apply memory
> > > > > +pressure to the page cache and swap-backed pages equally. At 0, the
> > > > > +kernel will not initiate swap until the amount of free and file-backed
> > > > > +pages is less than the high watermark in a zone.
> > > > 
> > > > Generally, I agree extending swappiness value good but not sure 200 is
> > > > enough to represent speed gap between file and swap sotrage in every
> > > > cases. - Just nitpick.
> > > 
> > > How so? You can't give swap more weight than 100%. 200 is the maximum
> > > possible value.
> > 
> > In old, swappiness is how agressively reclaim anonymous pages in favour
> > of page cache. But when I read your description and changes about
> > swappiness in vm.txt, esp, *relative IO cost*, I feel you change swappiness
> > define to represent relative IO cost between swap storage and file storage.
> > Then, with that, we could balance anonymous and file LRU with the weight.
> > 
> > For example, let's assume that in-memory swap storage is 10x times faster
> > than slow thumb drive. In that case, IO cost of 5 anonymous pages
> > swapping-in/out is equal to 1 file-backed page-discard/read.
> > 
> > I thought it does make sense because that measuring the speed gab between
> > those storages is easier than selecting vague swappiness tendency.
> > 
> > In terms of such approach, I thought 200 is not enough to show the gab
> > because the gap is started from 100.
> > Isn't it your intention? If so, to me, the description was rather
> > misleading. :(
> 
> The way swappiness works never actually changed.
> 
> The only thing that changed is that we used to look at referenced
> pages (recent_rotated) and *assumed* they would likely cause IO when
> reclaimed, whereas with my patches we actually know whether they are.
> But swappiness has always been about relative IO cost of the LRUs.
> 
> Swappiness defines relative IO cost between file and swap on a scale
> from 0 to 200, where 100 is the point of equality. The scale factors
> are calculated in get_scan_count() like this:
> 
>   anon_prio = swappiness
>   file_prio = 200 - swappiness
> 
> and those are applied to the recorded cost/value ratios like this:
> 
>   ap = anon_prio * scanned / rotated
>   fp = file_prio * scanned / rotated
> 
> That means if your swap device is 10 times faster than your filesystem
> device, and you thus want anon to receive 10x the refaults when the
> anon and file pages are used equally, you do this:
> 
>   x + 10x = 200
>         x = 18 (ish)
> 
> So your file priority is ~18 and your swap priority is the remainder
> of the range, 200 - 18. You set swappiness to 182.
> 
> Now fill in the numbers while assuming all pages on both lists have
> been referenced before and will likely refault (or in the new model,
> all pages are refaulting):
> 
>   fraction[anon] = ap      = 182 * 1 / 1 = 182
>   fraction[file] = fp      =  18 * 1 / 1 =  18
>      denominator = ap + fp =    182 + 18 = 200
> 
> and then calculate the scan target like this:
> 
>   scan[type] = (lru_size() >> priority) * fraction[type] / denominator
> 
> This will scan and reclaim 9% of the file pages and 90% of the anon
> pages. On refault, 9% of the IO will be from the filesystem and 90%
> from the swap device.

Thanks for the detail example. Then, let's change the example a little bit.

A system has big HDD storage and SSD swap.

HDD:    200 IOPS
SSD: 100000 IOPS
>From https://en.wikipedia.org/wiki/IOPS

So, speed gap is 500x.
x + 500x = 200
If we use PCIe-SSD, the gap will be larger.
That's why I said 200 is enough to represent speed gap.
Such system configuration is already non-sense so it is okay to ignore such
usecases?