[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: More VM balancing issues..
>
>
> Oh, I found another problem: when the VM balancing was rewritten, the
> "pages_low" thing was still calculated, but nothing actually USED it.
2.3.50 did.
This is one of the things I pointed out in the recent balancing patch.
The patch I sent to Christopher tries to fix it and go back to 2.3.50
bahavior.
Again, Documentation/vm/balance has comments about this which were
true till 2.3.50.
AFAIR, Andrea put this stuff into 2.3, round about 2.3.40 or so timeframe.
>
> So we had three water-marks: "enough for anything", "low on memory" and
> "critical".
>
> And we somehow lost the "low on memory" and only used the "enough" and
> "critical" to do all comparisons.
This was also done properly in 2.3.50, and changed in the recent patch.
Maybe its time that tweaks to the balancing code get documented in
Documentation/vm/balance. Its easier all around to keep track of
what's happening.
Kanoj
>
> Which makes for a _very_ choppy balance, and is definitely wrong.
>
> The behaviour should be something like:
> - whenever we dip below "low", we wake up kswapd. kswapd remains awake
> (for that zone) until we reach "enough".
> - whenever we dip below "critical", we start doing synchronous memory
> freeing ourselves. We continue to do that until we reach "low" again
> (at which point kswapd will still continue in the background, but we
> don't depend on the synchronous freeing any more).
>
> but for some time we appear to have gotten this wrong, and lost the "low"
> mark, and used the "critical" and "high" marks only.
>
> Or maybe somebody did some testing and decided to disagree with the old
> three-level thing based on actual numbers? The only coding I've done has
> been based on "this is how I think it should work, and because I'm always
> right it's obviously the way it _should_ work". Which is not always the
> approach that gets the best results ;)
>
> Linus
>
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@kvack.org. For more info on Linux MM,
> see: http://www.linux.eu.org/Linux-MM/
>
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org. For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/