Re: [RFC][PATCH 0/6] more detailed per-process transparenthugepage statistics

[Andrea Arcangeli]

Hello,

On Tue, Feb 08, 2011 at 09:54:34AM -0800, Dave Hansen wrote:
> Just FYI, I did some profiling on a workload that constantly split and
> joined pages.  Very little of the overhead was in the scanning itself,
> so I think you're dead-on here.

Yep, my way to deduce it has been to set both to 100%, and check the
rate of increase of
/sys/kernel/mm/transparent_hugepage/khugepaged/full_scans vs
/sys/kernel/mm/ksm/full_scans and the differences is enormous. So a
100% CPU ksmd scan can probably be followed more than well with a 1%
CPU khugepaged scan and probably achieve the exact same hugepage ratio
of a 100% khugepaged scan. The default khugepaged scan is super
paranoid (it has to be, considering the default ksm scan is
zero). Maybe we can still increase the default pages_to_scan a bit. I
suspect most of the current cost should be in the scheduler and that
only accounts for 1 kthread schedule event every 10 sec.
--
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/