On Sat, 10 Feb 2001, Rik van Riel wrote:
> On Sat, 10 Feb 2001, Mike Galbraith wrote:
> > On Sat, 10 Feb 2001, Rik van Riel wrote:
> > > On Sat, 10 Feb 2001, Marcelo Tosatti wrote:
> > > > On Sat, 10 Feb 2001, Mike Galbraith wrote:
> > > >
> > > > > This change makes my box swap madly under load.
> > > >
> > > > Swapped out pages were not being counted in the flushing limitation.
> > > >
> > > > Could you try the following patch?
> > >
> > > Marcelo's patch should do the trick wrt. to making page_launder()
> > > well-behaved again. It should fix the problems some people have
> > > seen with bursty swap behaviour.
> >
> > It's still reluctant to shrink cache. I'm hitting I/O saturation
> > at 20 jobs vs 30 with ac5. (difference seems to be the delta in
> > space taken by cache.. ~same space shows as additional swap volume).
>
> Indeed, to "fix" that we'll need to work at refill_inactive().
If this reluctance to munch cache can be relaxed a little, I think
we'll see the end of a long standing problem. I often see a scenario
wherein we flush everything flushable, then steal the entire cache
before doing any paging. The result (we hit a wall) is a mondo swapout
followed immediately by swapping it all right back in. We seem to have
done a complete turnaround wrt paging vs flush/cache reap preference,
and that does effectively cure this scenario.. but methinks optimal
(-ENOENT?) lies somewhere in between.
> However, I am very much against tuning the VM for one particular
> workload. If you can show me that this problem also happens under
> other workloads we can work at changing it, but I don't think it's
> right to optimise the VM for a specific workload...
I'll watch behavior under other loads. (I don't have enough network
capacity to do anything stressful there, and whatever load I pick
has to be compute bound as to not end up benchmarking my modest I/O
capacity.. suggestions welcome. I use make -j primarily because it
doesn't need much I/O bandwidth for itself, but does allocate quite
a bit.. that leaves most I/O capacity free for vm usage)
Something else I see while watching it run: MUCH more swapout than
swapin. Does that mean we're sending pages to swap only to find out
that we never need them again?
-Mike
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [EMAIL PROTECTED]
Please read the FAQ at http://www.tux.org/lkml/
