[geeks] One, two, three, panic!

[Jonathan C. Patschke]

On Mon, 16 Mar 2009, Phil Stracchino wrote:

> So the doc that says keep vm.kmem_size_max UNDER 400M is wrong?

For ZFS?  Absolutely.

> I hadn't come across the "known problems" doc Michael Turner referenced.
> Frankly, reading that really makes it sound as though even on amd64,
> ZFS on FreeBSD really isn't ready to be actually used yet.

Well, there is the message that pops up in dmesg when you load the ZFS
module:

   WARNING: ZFS is considered to be an experimental feature in FreeBSD.

It can be made to work reliably; it just can't be trusted to regulate
itself yet.  ZFS wasn't stable right away in Solaris, either.  Roughly two
years passed between ZFS's inclusion in Sun's "ten steps ahead"
advertisement and the first release of Solaris 10 that had some basic
version of ZFS included.

> Going by that document, there is no such thing as a really stable
> FreeBSD+ZFS configuration, particularly on i386.

Correct.  You can make it stable by constraining its memory usage, but it
really wants a 64-bit ABI on which to run.  ZFS was designed with 64-bit
processors in mind; it heavily uses 64-bit arithmetic and doesn't blink at
asking for half a gig of memory at a go.

The configuration I posted WorksForMe(tm), even if I really want to beat
up on the disks, but it's very conservative and is tuned for stability
rather than performance.  I can't have my stuff crashing while I'm
overseas for weeks at a stretch.

>> This information sounds uselessly out-of-date.  Where did you find it?  I
>> can't remember the last time I installed FreeBSD on a system with -less-
>> than 1GB memory.
>
> It was about the third or fourth hit (and the first with actual
> information) on that error.

I finally found the document you were talking about; unfortunately, it's
the Handbook. :(

That information was added back in 2004, and most likely describes FreeBSD
4.10 or some early release of 5.  If it really does document 4.10, the
kernel's changed enough since then that FreeBSD 7 may be regarded as a
wholly different OS.

> That's what I'd have assumed, yes, but the documented fix said set it
> 400M *or lower* but not to exceed 400M.
>
> It seemed odd to me, but, well, it was documented...

It's really unfortunate that it's still in the handbook.  Unmaintained
documentation can be worse than none at all.

-- 
Jonathan Patschke ( "They don't have the right to read a book out loud."
Elgin, TX         (                  --Paul Aiken
USA               (                    Executive Director, Authors Guild