Performance 4.x vs. 6.x

Thu Oct 12 12:26:32 PDT 2006

What a load...

Here's a report...

I have over 800 nodes installed in the field with FreeBSD 6.0 running
as routers on silly little 1.3Ghz machines with 256MB of RAM.  They
run Apache/PHP/wSSL enabled, MySQL, dual-firewall with custom NetGraph
module for Wireless MAC authentication.  The company does over 180k a
month in subscribers in the trucking industry in the US.

The company has TWO network administrators who do very little during
the day because the machines NEVER die.  If they do, 99.9% of the time
its hardware related.

I built those systems in 2 months and they support remote rollout of a
new operating system snapshot and they're preparing to rollout 7.0
when its stable.  I no longer work there - only on occassion when they
need assistance.

Internally, I have 50 FreeBSD machines hosting over 600 complex web
applications that my firm has built over the last 11 years using ONLY
FreeBSD.  Currently, they're all running FreeBSD 6.0 and later and "I"
am the only network administrator in the company.  If I was running
anything else (which, we do run some Windows machines and they are the
bain of my existence...) I would be too busy to do anything else.

One of our largest systems has redundant load-balancers with three
presentation boxes serving web pages out of memory - again, Apache
w/PHP.  These boxes build 200+ page 300dpi PDF documents for high
school year books (including LOTS of 300+ dpi student and faculty
images).  They're supported by two mid-sized database machines, one
read, one write (replicated, obviously) that do 200 to 500 queries per
second at busy times during the day.  Graphic data is all stored on
SATA data storage systems, which after a bit of tweaking scale really
well using NFS and Jumbo Frames - bound multiple NICs with the ng_fec
module (thank you thank you guys)...

Oh yeah, forgot to mention, once the system was setup, I haven't had
to touch it - and even "braver" yet, these 2 load balancers, 3
presentation machines, 2 database machines and 2 1.4TB data storage
boxes ALL run 7.0-CURRENT.  Call me stupid, brave, whatever - but 7.0
, with the snapshot release I got is the fastest I have ever seen
FreeBSD run, regardless of the fact the hardware is fast.  I've tuned
each machine using the online docs and a bit of help from PHK and Juli
Malette...

Interesting stat - from 10 other machines, I used ab to toss some hits
at these boxes.  Like:

ab -n 1000 -c 20 <url>

The page hit was a test page that did reading and writing, several
times to the database and read an image, used MagickWand to resample
them and write the image back.

The average time for the test took 4 to 5 seconds.  I achieved around
~220 requests per second per test machine with 75 to 100ms per
request.

I don't want to feed the trolls either, but sometimes performance is
achieved because you take the time to read and don't just install the
OS "as-is" and expect it to work well on all hardware.  When
configured properly, in my opinion, FreeBSD kicks ass.

D

On 10/12/06, Eric Anderson <anderson at centtech.com> wrote:
> On 10/12/06 09:19, Danial Thom wrote:
> >
> > --- Alexander Leidinger <Alexander at Leidinger.net>
> > wrote:
> >
> >> Quoting Dan Lukes <dan at obluda.cz> (from Thu, 12
> >> Oct 2006 09:43:20 +0200):
> >>
> >> [moved from security@ to performance@]
> >>
> >>>     The main problem is - 6.x is still not
> >> competitive replacement for
> >>> 4.x. I'm NOT speaking about old unsupported
> >> hardware - I speaked about
> >>> performance in some situation and believe in
> >> it's stability.
> >>
> >> You can't be sure that a committer has the
> >> resources to setup an
> >> environment where he is able to reproduce your
> >> performance problems.
> >> You on the other hand have hands-on experience
> >> with the performance
> >> problem. If you are able to setup a -current
> >> system (because there are
> >> changes which may affect performance already,
> >> and it is the place
> >> where the nuw stuff will be developt) which
> >> exposes the bad behavior,
> >> you could make yourself familiar with the pmc
> >> framework
> >> (http://wiki.freebsd.org/PmcTools, I'm sure
> >> jkoshy@ will help if you
> >> have questions) and point out the bottlenecks
> >> on current@ and/or
> >> performance@ (something similar happened for
> >> MySQL, and now we have a
> >> webpage in the wiki about it). Without such
> >> reports, we can't handle
> >> the issue.
> >>
> >> Further discussion about this should happen in
> >> performance@ or current at ...
> >>
> >> Bye,
> >> Alexander.
> >>
> >
> > Maybe its just time for the entire FreeBSD team
> > to come out of its world of delusion and come to
> > terms with what every real-life user of FreeBSD
> > knows: In how ever many years of development,
> > there is still no good reason to use anything
> > other than FreeBSD 4.x except that 4.x doesn't
> > support a lot of newer harder. There is no
> > performance advantage in real world applications
> > with multiple processors, and the performance is
> > far worse with 1 processor.
> >
> > The right thing to do is to port the SATA support
> > and new NIC support back to 4.x and support both.
> > 4.x is far superior on a Uniprocessor system and
> > FreeBSD-5+ may be an entire re-write away from
> > ever being any good at MP. Come to terms with it,
> > PLEASE, because it is the case and saying
> > otherwise won't change it.
> >
> > My prediction is that a  year from now we'll all
> > be using DragonflyBSD and you guys will be
> > looking for a new bunch of beta-test guinea pigs.
>
> My prediction is that a year from now single processor systems are going
> to look like 386's to the rest of the world using multi-proc with
> FreeBSD-6 or 7, meanwhile enjoying the increased filesystem performance
> gained from non-giant-locked UFS2, the GEOM tools, etc, etc..
>
> Anyway, people should stop complaining, and start offering up hardware,
> net connections, and man power to support a cvs repo/packages/etc for
> the 4.x tree if they want it.  That's what people do, and that's the
> beauty of open source.
>
>
> Eric
>
>
>
>
> --
> ------------------------------------------------------------------------
> Eric Anderson        Sr. Systems Administrator        Centaur Technology
> Anything that works is better than anything that doesn't.
> ------------------------------------------------------------------------
> _______________________________________________
> freebsd-performance at freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-performance
> To unsubscribe, send any mail to "freebsd-performance-unsubscribe at freebsd.org"
>

-- 
Derrick T. Woolworth, President
ServeTheWeb, LLC.  http://www.ServeTheWeb.com