SMP Performance (Was: Re: Are hardware vendors starting to bail ... )

[User Freebsd]

On Thu, 13 Jul 2006, Tamouh H. wrote:

> I have to put my two cents here:
>
> 1) I agree with few posters that FreeBSD performance have been lacking 
> behind. I've reported few issues on performance list and many did. We 
> offered few pre-production servers for performance testing, but the 
> answer we keep getting is:
>
> a. It is either your hardware sucks
> b. your benchmark application sucks

'k, here's to all the performance folks ... how should someone test 
performance?

a. actually doesn't apply, as long as your performance testing is being 
done apples to apples as far as hardware is concerned ... if I create a 
dual-boot system, with FreeBSD 4.x and FreeBSD 6.x on a machine, and run 
*accepted performance / benchmark applications*, and compare those 
results, one would hope that 6.x performance fater/better then 4.x ...

> 2) Regarding SMP, few posts talked about disabling hyper-thread and SMP 
> because it causes a performance degradation. On production hosting 
> server, the experience was otherwise though. Without HT and SMP, the 
> server would sky rocket in resource consumption. This has been tested on 
> FBSD 5.4 i386

Personally, I've never found HT to be a performance boost, and I run 9 
'production hosting servers' ... I can actually feel the difference 
between turning it on/off ... not sure what you mean by 'sky rocket in 
resource consumption', but all my Dual Xeon servers have HTT disabled, and 
I'm not noticing anything odd ... if you could elaborate on how you are 
seeing this, I can check on my machine to see if I see similar ...

> 3) I'm also frustrated like many with the rapid advancement in release 
> jumps. We barely started 5.x to conclude it does not live up to 
> expectations, so now 6.x is suppoused to be the good version, yet 7.x is 
> going to come out soon and probably in less than a year 6.x will be 
> considered inadequate.

As to this one ... 5.x built up a very very bad reputation for itself, so 
basically 'skipping' that one makes sense ... I know I wouldn't trust a 
new version of 5.x coming out ... 6.x, other then the file system 
deadlocks which I'm trying to provide suitable DDB traces for, I've not 
noticed anything wrong with 6.x ...

The jump from 6.x to 7.x does seem a bit ... quick ... but, then again, 
7.x hasn't been released yet, and I think its safe to say that we all know 
that in software development, 'release estimates' are almost never 
accurate ...

The problem, as I see it, is that until the OS gets used in "real life 
production environments", some of the more obscure bugs don't get found 
... on a simple production server, not doing much, I doubt anyone would 
ever see the file system deadlocks ... but, there are several of us that 
are running it in production with heavy loads that do ... but it takes a 
good load on the machine to trigger it, and I doubt any of the developers 
have that to work with, and/or can easily simulate the 'randomness' of a 
production environment ...

----
Marc G. Fournier           Hub.Org Networking Services (http://www.hub.org)
Email . scrappy at hub.org                              MSN . scrappy at hub.org
Yahoo . yscrappy               Skype: hub.org        ICQ . 7615664