On Tue, May 19, 2009 at 08:28:28PM +0800, KOSAKI Motohiro wrote:
> Hi
> 
> 2009/5/19 Wu Fengguang <fengguang.wu@intel.com>:
> > On Tue, May 19, 2009 at 04:06:35PM +0800, KOSAKI Motohiro wrote:
> >> > > > Like the console mode, the absolute nr_mapped drops considerably - to 1/13 of
> >> > > > the original size - during the streaming IO.
> >> > > >
> >> > > > The delta of pgmajfault is 3 vs 107 during IO, or 236 vs 393 during the whole
> >> > > > process.
> >> > >
> >> > > hmmm.
> >> > >
> >> > > about 100 page fault don't match Elladan's problem, I think.
> >> > > perhaps We missed any addional reproduce condition?
> >> >
> >> > Elladan's case is not the point of this test.
> >> > Elladan's IO is use-once, so probably not a caching problem at all.
> >> >
> >> > This test case is specifically devised to confirm whether this patch
> >> > works as expected. Conclusion: it is.
> >>
> >> Dejection ;-)
> >>
> >> The number should address the patch is useful or not. confirming as expected
> >> is not so great.
> >
> > OK, let's make the conclusion in this way:
> >
> > The changelog analyzed the possible beneficial situation, and this
> > test backs that theory with real numbers, ie: it successfully stops
> > major faults when the active file list is slowly scanned when there
> > are partially cache hot streaming IO.
> >
> > Another (amazing) finding of the test is, only around 1/10 mapped pages
> > are actively referenced in the absence of user activities.
> >
> > Shall we protect the remaining 9/10 inactive ones? This is a question ;-)
> 
> Unfortunately, I don't reproduce again.
> I don't apply your patch yet. but mapped ratio is reduced only very little.

mapped ratio or absolute numbers? The ratio wont change much because
nr_mapped is already small.

> I think smem can show which library evicted.  Can you try it?
> 
> download:  http://www.selenic.com/smem/
> usage:   ./smem -m -r --abbreviate

Sure, but I don't see much change in its output (see attachments).

smem-console-0 is collected after fresh boot,
smem-console-1 is collected after the big IO.

> We can't decide 9/10 is important or not. we need know actual evicted file list.

Right. But what I measured is the activeness. Almost zero major page
faults means the evicted 90% mapped pages are inactive during the
long 300 seconds of IO.

Thanks,
Fengguang

> > Or, shall we take the "protect active VM_EXEC mapped pages" approach,
> > or Christoph's "protect all mapped pages all time, unless they grow
> > too large" attitude? A I still prefer the best effort VM_EXEC heuristics.
> >
> > 1) the partially cache hot streaming IO is far more likely to happen
> > A  on (file) servers. For them, evicting the 9/10 inactive mapped
> > A  pages over night should be acceptable for sysadms.
> >
> > 2) for use-once IO on desktop, we have Rik's active file list
> > A  protection heuristics, so nothing to worry at all.
> >
> > 3) for big working set small memory desktop, the active list will
> > A  still be scanned, in this situation, why not evict some of the
> > A  inactive mapped pages? If they have not been accessed for 1 minute,
> > A  they are not likely be the user focus, and the tight memory
> > A  constraint can only afford to cache the user focused working set.
> >
> > Does that make sense?