[Beowulf] big read triggers migration and slow memory IO?

[Kilian Cavalotti]

Hi David,

> Well, that seems to be it, but not quite with the same symptoms you
> observed.  khugepaged never showed up, and "perf top" never revealed
> _spin_lock_irqsave.

Ah, very interesting. Reading the thread, I was thinking about this,
but the symptoms I've seen in the past quite didn't match yours
(khugepaged was a big red flag in my case).
So, as I suggested on James' blog, to make things even better, you can
choose to enable huge pages defragmentation only for applications that
explicitly want to use hugepages (with MADV_HUGEPAGE).

#  echo madvise > /sys/kernel/mm/transparent_hugepage/defrag

See https://www.kernel.org/doc/Documentation/vm/transhuge.txt for details.

Cheers,
-- 
Kilian