On 2016-01-07 13:34:51 -0500, Tom Lane wrote:
> It's fairly well established that the implementation of transparent
> huge pages in Linux kernels from the 2.6-or-so era sucks, and you're
> best off turning it off if you care about consistency of performance.
I think the feature wasn't introduced in original 2.6 kernels (3.2 or
so?), but red hat had backported it to their 2.6.32 kernel.
> I am not sure whether modern kernels have improved this area.
I think the problem has largely been solved around 3.16. Around 4.1 I
could still reproduce problems, but the regressions were only in the
sub 10% range in my test workload.
Andres
