Re: [RFC][PATCH] split file and anonymous page queues #3

From: Nikita Danilov
Date: Wed Mar 21 2007 - 11:29:38 EST


Rik van Riel writes:
> Rik van Riel wrote:
> > Nikita Danilov wrote:
> >
> >> Probably I am missing something, but I don't see how that can help. For
> >> example, suppose (for simplicity) that we have swappiness of 100%, and
> >> that fraction of referenced anon pages gets slightly less than of file
> >> pages. get_scan_ratio() increases anon_percent, and shrink_zone() starts
> >> scanning anon queue more aggressively. As a result, pages spend less
> >> time there, and have less chance of ever being accessed, reducing
> >> fraction of referenced anon pages further, and triggering further
> >> increase in the amount of scanning, etc. Doesn't this introduce positive
> >> feed-back loop?
> >
> > It's a possibility, but I don't think it will be much of an
> > issue in practice.
> >
> > If it is, we can always use refaults as a correcting
> > mechanism - which would have the added benefit of being
> > able to do streaming IO without putting any pressure on
> > the active list, essentially clock-pro replacement with
> > just some tweaks to shrink_list()...
>
> As an aside, due to the use-once algorithm file pages are at a
> natural disadvantage already. I believe it would be really
> hard to construct a workload where anon pages suffer the positive
> feedback loop you describe...

That scenario works for file queues too. Of course, all this is but a
theoretical speculation at this point, but I am concerned that

- that loop would tend to happen under various border conditions,
making it hard to isolate, diagnose, and debug, and

- long before it becomes explicitly visible (say, as an excessive cpu
consumption by scanner), it would ruin global lru ordering, degrading
overall performance.

Generally speaking, multi-queue replacement mechanisms were tried in the
past, and they all suffer from the common drawback: once scanning rate
is different for different queues, so is the notion of "hotness",
measured by scanner. As a result multi-queue scanner fails to capture
working set properly.

Nikita.


>
> --
> Politics is the struggle between those who want to make their country
> the best in the world, and those who believe it already is. Each group
> calls the other unpatriotic.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/