Subject | AW: [IBO] Re: IB6 |
---|---|
Author | Kaputnik |
Post date | 2000-12-01T18:28:18Z |
Hi,
I'm actually running a dual XEON Box with IB6 and what Wayne wrote, is
almost true.
Well, performance doesn't degrade half, but there is definitely a
performance-loss when running anormal setup.
I have a small tool to generate test-data for IB-databases, and this can
show the true performance of a IBserver really good.
When I run with 2 processors, and I get e.g 6.000 inserts/second, then I
will get around 6.500-7.000 inserts/second when I set the Affinity of the
IBProcess to one processor. That makes some 8%-16% performance increase.
The reason is, that WinNT has some really bad load-balancing algorythm. As
soon as it sees, that there is a discrepance in the loads of the processors,
it will swap the most expensive process to the less loaded processor. And as
there is mostly only the IB-process running, WinNT will always swap this
process between the CPU's. Swapping a process has a small overhead
(tranfserring the CPU's cache and so on) and so the performance goes down.
Linux won't have these problems, as the Load-balancer is far better. The IB
classic architecture spawns seperate processes for every user, so that a
multi-processor system under linux should be normally setup with the classic
architecture.
If you are running Win2K, you will get some problems with IBAffinity, but
you can resolve them, when you set up a different user for the service, as
you can't set up a processes afinity, when it is logged in as the
system-user.
CU, Kaputnik
(Nick Josipovic)
nick@... <mailto:nick@...>
kap@... <mailto:kap@...>
-----------------------------------------------------------------------
superior Client/Server programming:
www.IBObjects.com <http://www.ibobjects.com/>
a nice Tool for Interbase:
www.InterbaseWorkbench.com <http://www.interbaseworkbench.com/>
I'm actually running a dual XEON Box with IB6 and what Wayne wrote, is
almost true.
Well, performance doesn't degrade half, but there is definitely a
performance-loss when running anormal setup.
I have a small tool to generate test-data for IB-databases, and this can
show the true performance of a IBserver really good.
When I run with 2 processors, and I get e.g 6.000 inserts/second, then I
will get around 6.500-7.000 inserts/second when I set the Affinity of the
IBProcess to one processor. That makes some 8%-16% performance increase.
The reason is, that WinNT has some really bad load-balancing algorythm. As
soon as it sees, that there is a discrepance in the loads of the processors,
it will swap the most expensive process to the less loaded processor. And as
there is mostly only the IB-process running, WinNT will always swap this
process between the CPU's. Swapping a process has a small overhead
(tranfserring the CPU's cache and so on) and so the performance goes down.
Linux won't have these problems, as the Load-balancer is far better. The IB
classic architecture spawns seperate processes for every user, so that a
multi-processor system under linux should be normally setup with the classic
architecture.
If you are running Win2K, you will get some problems with IBAffinity, but
you can resolve them, when you set up a different user for the service, as
you can't set up a processes afinity, when it is logged in as the
system-user.
CU, Kaputnik
(Nick Josipovic)
nick@... <mailto:nick@...>
kap@... <mailto:kap@...>
-----------------------------------------------------------------------
superior Client/Server programming:
www.IBObjects.com <http://www.ibobjects.com/>
a nice Tool for Interbase:
www.InterbaseWorkbench.com <http://www.interbaseworkbench.com/>
> -----Ursprüngliche Nachricht-----
> Von: xomp@... [mailto:xomp@...]
> Gesendet: Freitag, 1. Dezember 2000 18:34
> An: IBObjects@egroups.com
> Betreff: Re: [IBO] Re: IB6
>
>
> Thank U Helen,
>
> I did my test. Going to last row with IB_WISQL on a large data set
>
> One Cpu. CPU Usuage 100% (pain in neck to remove it)
> Two Cpu's. CPU Usuage 50%
>
> Overall on single query a slight performance improvement, but
> leaving my 'hands' free to do other things without affecting the
> speed of the query executing at the same time.
>
> Why do they start those rumors :)
>
> Best regards,
> Hans
>
> ==================================
>
> Helen Borrie wrote:
> >
> > At 09:23 AM 01-12-00 -0700, you wrote:
> > >Hello,
> > >
> > >Does anyone know if the attached message bares any truth. I run
> > >a dual pentium setup as a server. If so, does it appply to
> > >NT and Linux both ? Personally I haven't noticed any degradation,
> > >but I haven't compared results yet to an after a 'yank a processor'
> > >setup.
> > >
> > >Yes, I know my question doesn't appear IBOjects related, but maybe
> > >it might be and it could impact the way we plan our systems :)
> >
> > Windows only, I believe. The obvious way to check it out is to
> monitor the
> > CPU usage - the reported behaviour is for CPU usage to jump
> from 0 to 100%
> > like a see-saw. I think it affects Superserver worse than Classic (or
> > might be vice versa).
> >
> > H.
> > All for Open and Open for All
> > InterBase Developer Initiative · http://www.interbase2000.org
> > _______________________________________________________
> >
>
>
>
>