Skip site navigation (1)Skip section navigation (2)
Date:      Wed, 13 Jun 2007 14:22:04 -0500
From:      Craig Boston <craig@xfoil.gank.org>
To:        Kris Kennaway <kris@obsecurity.org>
Cc:        current@freebsd.org
Subject:   Re: ZFS tuning tips?
Message-ID:  <20070613192204.GB82412@nowhere>
In-Reply-To: <20070613183347.GA54210@rot13.obsecurity.org>
References:  <20070613160835.GA6461@nowhere> <20070613183347.GA54210@rot13.obsecurity.org>

next in thread | previous in thread | raw e-mail | index | archive | help
On Wed, Jun 13, 2007 at 02:33:48PM -0400, Kris Kennaway wrote:
> I sent mail about this last week.

Are you referring to the "ZFS on 32-bit CPUs" topic?  I saw that, but
lowering maxvnodes to 75000 still results in panics for me.  Lowering it
further works, but results in poor performance.  I with there was a more
quantitative way to calculate it; right now it seems like taking shots
in the dark to see what works.

Missed the VM_KMEM_SIZE_MAX though, I'll try playing with that and see
if it helps.

As far as I understand most of the problem on i386 is KVA -- mostly I'm
thinking if there are ways to optimize the kernel memory space.  Maybe
cutting back on usage of other kernel allocations by lowering some
limits of things I'm not fully using.

I'm not 100% sure I understand why it's a problem even with <=2GB
memory.  I thought the kernel had a full 2GB of address space to itself
(or is the problem the 320/400M that's available for kernel malloc?).

Just wild speculation, and I'm probably way off here, but would it be
possible / worth it to try to teach ZFS to allocate VM objects rather
than carving directly out of kernel memory?  How does the FreeBSD buffer
cache handle it?

Craig



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20070613192204.GB82412>