[lkml]   [2011]   [Jun]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH 2/3] mm: make the threshold of enabling THP configurable
On Tue, Jun 21, 2011 at 01:16:00AM +0800, Cong Wang wrote:
> ??? 2011???06???21??? 00:59, Mel Gorman ??????:
> >On Tue, Jun 21, 2011 at 12:34:29AM +0800, Amerigo Wang wrote:
> >>Don't hard-code 512M as the threshold in kernel, make it configruable,
> >>and set 512M by default.
> >>
> >
> >I'm not seeing the gain here either. This is something that is going to
> >be set by distributions and probably never by users. If the default of
> >512 is incorrect, what should it be? Also, the Kconfig help message has
> >spelling errors.
> >
> Sorry for spelling errors, I am not an English speaker.
> Hard-coding is almost never a good thing in kernel, enforcing 512
> is not good either. Since the default is still 512, I don't think this
> will affect much users.
> I do agree to improve the help message, like Dave mentioned in his reply,
> but I don't like enforcing a hard-coded number in kernel.
> BTW, why do you think 512 is suitable for *all* users?

Fragmentation avoidance benefits from tuning min_free_kbytes to a higher
value and minimising fragmentation-related problems is crucial if THP is
to allocate its necessary pages.

THP tunes min_free_kbytes automatically and this value is in part
related to the number of zones. At 512M on a single node machine, the
recommended min_free_kbytes is close to 10% of memory which is barely
tolerable as it is. At 256M, it's 17%, at 128M, it's 34% so tuning the
value lower has diminishing returns as the performance impact of giving
up such a high percentage of free memory is not going to be offset by
reduced TLB misses. Tuning it to a higher value might make some sense
if the higher min_free_kbytes was a problem but it would be much more
rational to tune it as a sysctl than making it a compile-time decision.

Mel Gorman

 \ /
  Last update: 2011-06-21 11:39    [W:0.067 / U:3.180 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site