Re: linux memory settings on OL 6.2 ( red hat compatible )
From: Matthias Hoys <matthias.hoys_at_gmail.com>
Date: Mon, 10 Sep 2012 01:16:14 -0700 (PDT)
Message-ID: <573b0bfe-6ef7-4116-9c92-ed72d8fb8dd1_at_googlegroups.com>
On Saturday, September 8, 2012 2:15:58 AM UTC+2, Mladen Gogala wrote:
> On Thu, 06 Sep 2012 11:06:05 -0700, John Hurley wrote:
>
>
>
> > echo never > /sys/kernel/mm/redhat_transparent_hugepage/enabled echo
>
> > never > /sys/kernel/mm/redhat_transparent_hugepage/defrag
>
>
>
>
>
> You are smart. Transparent huge pages mechanism is still in its infancy.
>
> I have the same thing in my rc.local script. Speaking of sys, you should
>
> also enable deadline scheduler for the disks that will be used by Oracle:
>
>
>
>
>
> echo deadline >/sys/block/sdb/queue/scheduler
>
>
>
> Substitute your own devices for "sdb".
>
>
>
>
>
> >
>
> > We are using huge pages for Oracle SGA ( so automatic sizing crud turned
>
> > off ... all manual settings ). Aside from the usual settings for kernel
>
> > stuff in /etc/sysctl.conf for an oracle database we have ( again ...
>
> > preliminary testing ) these settings in effect now:
>
> >
>
> > #
>
> > vm.swappiness = 10 vm.vfs_cache_pressure = 200 vm.pagecache = 1 5 10
>
> > vm.min_free_kbytes = 81920 vm.nr_hugepages = 28672 EOF
>
>
>
>
>
> With such a large memory, you don't want to throw out inodes and
>
> directory entries quickly. I would set something like
>
> vm.vfs_cache_pressure = 20 to retain them as long as possible. That will
>
> give you the benefit of the caching. The "open" system service is an
>
> expensive operation and with 96GB of RAM you can afford to keep them
>
> around a little longer. The "pagecache" parameter is obsolete and should
>
> be omitted. You should use dirty_* parameters instead. Also, you want to
>
> set page-cluster to 5. That will cause swapper to try swapping out chunks
>
> of 32 (2**5) pages in a single IO operation, if possible. I would also
>
> set overcommit_memory to 1, to prevent OOM killer from activating.
>
>
Date: Mon, 10 Sep 2012 01:16:14 -0700 (PDT)
Message-ID: <573b0bfe-6ef7-4116-9c92-ed72d8fb8dd1_at_googlegroups.com>
On Saturday, September 8, 2012 2:15:58 AM UTC+2, Mladen Gogala wrote:
> On Thu, 06 Sep 2012 11:06:05 -0700, John Hurley wrote:
>
>
>
> > echo never > /sys/kernel/mm/redhat_transparent_hugepage/enabled echo
>
> > never > /sys/kernel/mm/redhat_transparent_hugepage/defrag
>
>
>
>
>
> You are smart. Transparent huge pages mechanism is still in its infancy.
>
> I have the same thing in my rc.local script. Speaking of sys, you should
>
> also enable deadline scheduler for the disks that will be used by Oracle:
>
>
>
>
>
> echo deadline >/sys/block/sdb/queue/scheduler
>
>
>
> Substitute your own devices for "sdb".
>
>
>
>
>
> >
>
> > We are using huge pages for Oracle SGA ( so automatic sizing crud turned
>
> > off ... all manual settings ). Aside from the usual settings for kernel
>
> > stuff in /etc/sysctl.conf for an oracle database we have ( again ...
>
> > preliminary testing ) these settings in effect now:
>
> >
>
> > #
>
> > vm.swappiness = 10 vm.vfs_cache_pressure = 200 vm.pagecache = 1 5 10
>
> > vm.min_free_kbytes = 81920 vm.nr_hugepages = 28672 EOF
>
>
>
>
>
> With such a large memory, you don't want to throw out inodes and
>
> directory entries quickly. I would set something like
>
> vm.vfs_cache_pressure = 20 to retain them as long as possible. That will
>
> give you the benefit of the caching. The "open" system service is an
>
> expensive operation and with 96GB of RAM you can afford to keep them
>
> around a little longer. The "pagecache" parameter is obsolete and should
>
> be omitted. You should use dirty_* parameters instead. Also, you want to
>
> set page-cluster to 5. That will cause swapper to try swapping out chunks
>
> of 32 (2**5) pages in a single IO operation, if possible. I would also
>
> set overcommit_memory to 1, to prevent OOM killer from activating.
>
>
Mladen,
Are there any tools to measure/tune the performance impact of these VM subsystem settings? Do you have any idea what the performance gain can be on a optimal configured system with Oracle running on it?
Thanks,
Matthias
http://matthiashoys.wordpress.com
Received on Mon Sep 10 2012 - 03:16:14 CDT