Re: [RFC PATCH V3] mm readahead: Fix the readahead fail in case ofempty numa node

[Raghavendra K T]

On 01/07/2014 03:43 AM, Andrew Morton wrote:
> On Mon,  6 Jan 2014 15:51:55 +0530 Raghavendra K T <raghavendra.kt@xxxxxxxxxxxxxxxxxx> wrote:
>
> > +	/*
> > +	 * Readahead onto remote memory is better than no readahead when local
> > +	 * numa node does not have memory. We sanitize readahead size depending
> > +	 * on free memory in the local node but limiting to 4k pages.
> > +	 */
> > +	return local_free_page ? min(sane_nr, local_free_page / 2) : sane_nr;
> >   }
>
> So if the local node has two free pages, we do just one page of
> readahead.
>
> Then the local node has one free page and we do zero pages readahead.
>
> Assuming that bug(!) is fixed, the local node now has zero free pages
> and we suddenly resume doing large readahead.
>
> This transition from large readahead to very small readahead then back
> to large readahead is illogical, surely?

Hi Andrew, Thanks for having a look at this.

You are correct that there is a transition from small readahead to
large once we have zero free pages.
I am not sure I can defend well, but 'll give a try :).

Hoping that we have evenly distributed cpu/memory load, if we have very
less free+inactive memory may be we are in really bad shape already.

But in the case where we have a situation like below [1] (cpu does not 
have any local memory node populated) I had mentioned
earlier where we will have to depend on remote node always,
is it not that sanitized readahead onto remote memory seems better?

But having said that I am not able to get an idea of sane implementation
to solve this readahead failure bug overcoming the anomaly you pointed
:(.  hints/ideas.. ?? please let me know.


[1]: IBM P730
----------------------------------
# numactl -H
available: 2 nodes (0-1)
node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 
23 24 25 26 27 28 29 30 31
node 0 size: 0 MB
node 0 free: 0 MB
node 1 cpus:
node 1 size: 12288 MB
node 1 free: 10440 MB
node distances:
node   0   1
0:  10  40
1:  40  10

--
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/