[nycbug-talk] Fwd: Re: OpenSolaris dead

Pete Wright pete at nomadlogic.org
Thu Aug 19 12:57:49 EDT 2010

On Tue, Aug 17, 2010 at 12:40:52AM -0400, Charles Sprickman wrote:
> On Mon, 16 Aug 2010, Jason Dixon wrote:
> Slightly OT, I'm looking for a benchmark that deals more with iops rather 
> than throughput and has a fairly simple output - just something that 
> sweeps through various sizes and then spits out a summary of how many io 
> operations/second it recorded.

netapp's postmark might be worth taking a look.  also iozone is pretty
decent.  netapp pulled the technical report that details the reasoning
behind postmark, but the source should still be available in the ports tree.  

> I am somewhat concerned about FreeBSD in general and specifically what the 
> plans are there for ZFS (both in general and  the impact of oracle perhaps 
> not ever releasing any additional code).  I've seen lots of folks 
> experimenting with FreeBSD just for ZFS, many that have never given 
> FreeBSD a second thought.  They seem to avoid the official mailing lists 
> and work with each other to tune and debug things either on the FBSD 
> forums or other general tech forums, so I don't know how aware the core 
> group of FBSD developers are of this growing group of users coming to FBSD 
> strictly for ZFS (and finding a pretty decent OS to go along with it).

IANAL - but I would reckon that the CDL bits of zfs (and dtrace for that
matter) will continue to be released to the community.  

as I read the posting on the opensol mailing list was that oracle wanted
to ensure they could capitalize fully on solaris (and support) and
ensure that features being developed in opensol are made into official
products (solairs 11 for exaple) more quickly.

i think oracle is moving in the direction of creating appliances built
on Sun hardware, OS's and bundling that layer on top of Oracle application 
frameworks and databases.  i think the "exadata" product is a good
example of this thinking.  it is a datawarehousing applaince build on
sun hardware (including their cool flash disk modules), run's solairs 10
and hosts an oracle database.

another example that comes to mind is the openstorage platform.

i guess what i am saying is that the memo seemed to me to try to focus
the development effort of solaris into products that oracle can make
money on.

> It seems like even with ZFS now being offically labelled as 
> production-ready on FBSD, there's very little documentation popping up, 
> and the few people that understand the internals are not really available 
> to help out those who want to document it.  The classic case is the whole 
> issue of running ZFS on i386 - tuning it to not panic is not easy, and the 
> manpages are still just straight rips from open solaris - there's no 
> FreeBSD-specific info in there.  The wiki (http://wiki.freebsd.org/ZFS) is 
> somewhat dated and does not really explain the i386 tuning issue well. 
> It's also not open to editing.  There are a few tricks to getting i386 
> stable, but it's not readily apparent.  The biggest hint I found (and I've 
> heard this may have changed) is that the arc_max value is NOT a hard 
> limit, but a high water mark where a thread to flush the arc to disk kicks 
> off.  On slower/older hardware with unpredictable disk load, the time it 
> takes for that thread to do what it needs to can fall behind the rate at 
> which some process is writing data that lands in the ARC.  When that 
> happens, boom!  I had much better luck tuning this once I realized that I 
> could routinely load the box in a way that the arc could grow to double 
> it's "max" size.  In some cases that meant just going with better hardware 
> (and 64-bit, where zfs is very stable), and in others it could be worked 
> around by setting the ARC max very low - in essence giving the flush to 
> disk thread a little bit of a head start.

I agree the freebsd specific documentation certainly needs some help,
but fortunately there is decent documentation on the internals of ZFS
published by sun, which I have found quite helpful in the past.

> There's also the issue of sysinstall, especially if you want to do a zfs 
> on root setup.
pcbsd makes this easy - although I am personally not convinced of the
benefit of running ZFS on a root filesystem.  I'm a big fan of keeping
things simple and easy to fix when things go south - I find that UFS is
fast and well understood.  its more of a personal preference not really
based on any data though...

having said that I *do* see the utility of running ZFS on data volumes


Pete Wright
pete at nomadlogic.org

More information about the talk mailing list