Re: Addendum to Daniel Phillips [RFC] use-once patch

[Daniel Phillips <phillips@bonn-fries.net>]

On Friday 27 July 2001 11:57, Balbir Singh wrote:
> Addendum to Daniel Phillips [RFC] use-once patch
> ------------------------------------------------
>
> A while ago, Daniel Phillips, posted his use use once patch. I used
> it and found it useful. I have been thinking of something similar.
> Let me describe what I have been thinking, this is in-line with
> page-aging and the working set model.
>
> As per the working set model, we use locality of reference, to keep
> constantly used pages in memory. It is for sure that after a period
> of time, these pages that were being used constantly, would no longer
> be required (since we would be done with that piece of code or data).
> We would like to evict these pages since soon.
>
> To illustrate :-
>
> I have used a PAGE_MAX_USE principle (my own from what I know), which
> states that most of the pages (**except shared pages**), would be
> used for a maximum of PAGE_MAX_USE (some constant > 0). We look at
> pages that are very frequently used and then after some number of
> times (PAGE_MAX_USE) they have been used, we "victimize" them. This
> may be wrong, since the page may be required for more than the number
> of times (PAGE_MAX_USE), we think it is required. In that case, it
> will be paged back in (when required) and reused again for
> PAGE_MAX_USE times before being victimized again.
>
> Below is a small patch for proof of concept
> -------------------------------------------
>
>
> --- vmscan.c.org        Fri Jul 27 14:27:06 2001
> +++ vmscan.c    Fri Jul 27 14:32:38 2001
> @@ -43,10 +43,20 @@
>
>         /* Don't look at this pte if it's been accessed recently. */
>         if (ptep_test_and_clear_young(page_table)) {
> -               page->age += PAGE_AGE_ADV;
> -               if (page->age > PAGE_AGE_MAX)
> -                       page->age = PAGE_AGE_MAX;
> -               return;
> +
> +               /*
> +                * If the page has been at PAGE_AGE_MAX for a while,
> may be
> +                * it is the best candidate for swapping.
> +                */
> +               if ((page->age > PAGE_AGE_MAX) && (page_count(page)
> <= 1)) {
> +                       page->age = PAGE_AGE_START;
> +               } else {
> +                       page->age += PAGE_AGE_ADV;
> +                       if (page->age > PAGE_AGE_MAX) {
> +                               page->age = PAGE_AGE_MAX;
> +                       }
> +                       return;
> +               }
>         }

I noticed your good benchmark results below, but I'm having some 
trouble understanding how this works.  How can page->age ever become 
greater than PAGE_AGE_MAX?  Also, I don't see any reference to 
PAGE_MAX_USE.  Comments?

> System Configuration
> =====================
>
> Single processor celeron system with 128 MB of RAM, running
> Linux-2.4.7pre6 with Daniel's patch applied (running X windows at the
> time of compilation, with GNOME).
>
> time for creating clean bzImage *before* patch
> ==============================================
>
> real    28m40.492s
> user    22m43.450s
> sys     2m44.490s
>
>
> time for creating clean bzImage *after* patch
> =============================================
>
> real    26m37.011s
> user    21m56.350s
> sys     2m28.060s

Bash seems to have a built-in "time" command that isn't nearly as 
useful as the GNU version, which you'd find in /usr/bin/time.  The GNU 
version tells us, among other things, how many swaps occured.  Also, 
check the list for Marcelo's mm statistics patch.  I'm not sure what 
the integration status is on that.

> The system, seemed to respond faster (or I might be feeling so).
>
> I am also planning to run some standard benchmark (I need to figure
> out, which one, or you could guide me). If you like the idea, I will
> post the benchmark results also to you (soon!). This patch is a
> simple implementation of the idea, I could come out with a more
> comprehensive solution if required.

I imagine your system is swapping during your kernel build due to 
memory pressure created by gnome and X.  If you show the swapping 
statistics from GNU time maybe we can suggest a more predictable way of 
creating a similar load.  I always run my benchmark tests in text mode, 
by the way, just to try to eliminate some variables and get more 
consistent timing results.

Did you repeat your timing measurements several times, and did you 
start each test with a clean reboot?

--
Daniel