Re: New to reiserfs

Re: New to reiserfs

[Dieter Nützel]

On Monday 13 May 2002 10:07, Hans Reiser wrote:
> >>4.- I have a PIII 1 GHz, 512 MB RAM, SCSI server where I run sendmail,
> >>bind and squid. Will I have problems with performance if I migrate to
> >>reiserfs?
> >>
> sendmail will be lower in performance when using reiserfs unless you use 
> the latest not yet released patches and turned data journaling on.

More speed numbers on this to come with related post.
> Do you really push the performance limits of the machine as a mail server?

You mean if it is worth?

> squid will benefit from your putting all files in one directory.

So I should clobber my squid partition/dir and reconfigure to put all files 
into only one directory and let the squid hierarchy alone?

Thanks,
	Dieter

-- 
Dieter Nützel
Graduate Student, Computer Science

University of Hamburg
Department of Computer Science
@home: Dieter.Nuetzel@hamburg.de

Re: Re: New to reiserfs

[Joe Cooper]

Dieter Nützel wrote:
> On Monday 13 May 2002 10:07, Hans Reiser wrote:

>>squid will benefit from your putting all files in one directory.

Hi Hans and all,

Out of curiosity...When did this become true?  I haven't tested in some 
time, but way back in 2.4.0pre days it was very far from reality 
(all-in-one-dir was about 20-30% slower than the standard default of 
x/256/256).

I'll be happy to run some Polygraph runs of this...I wasn't aware 
anything had changed, however.  Perhaps someone has a pointer for the 
relevant changelog entries?
-- 
Joe Cooper <joe@swelltech.com>
http://www.swelltech.com
Web Caching Appliances and Support

Re: Re: New to reiserfs

[Hans Reiser]

Joe Cooper wrote:

> Dieter Nützel wrote:
>
>> On Monday 13 May 2002 10:07, Hans Reiser wrote:
>
>
>>> squid will benefit from your putting all files in one directory.
>>
>
> Hi Hans and all,
>
> Out of curiosity...When did this become true?  I haven't tested in 
> some time, but way back in 2.4.0pre days it was very far from reality 
> (all-in-one-dir was about 20-30% slower than the standard default of 
> x/256/256).
>
> I'll be happy to run some Polygraph runs of this...I wasn't aware 
> anything had changed, however.  Perhaps someone has a pointer for the 
> relevant changelog entries?

Perhaps I misremember, Nikita, please comment.

Hans

Re: New to reiserfs

[Hans Reiser]

Dieter Nützel wrote:

>On Monday 13 May 2002 10:07, Hans Reiser wrote:
>  
>
>>>>4.- I have a PIII 1 GHz, 512 MB RAM, SCSI server where I run sendmail,
>>>>bind and squid. Will I have problems with performance if I migrate to
>>>>reiserfs?
>>>>
>>>>        
>>>>
>>sendmail will be lower in performance when using reiserfs unless you use 
>>the latest not yet released patches and turned data journaling on.
>>    
>>
>
>More speed numbers on this to come with related post.
>  
>
>>Do you really push the performance limits of the machine as a mail server?
>>    
>>
>
>You mean if it is worth?
>
>  
>
>>squid will benefit from your putting all files in one directory.
>>    
>>
>
>So I should clobber my squid partition/dir and reconfigure to put all files 
>into only one directory and let the squid hierarchy alone?
>
>Thanks,
>	Dieter
>
>  
>
I don't understand you.  

You should put squid into one big directory (not using their kludge that 
avoids making directories too large by using a tree)  unless you have 
multiple processors (in which case the VFS directory giant lock will 
hurt you).

Hans

Re: New to reiserfs

[Dieter Nützel]

On Monday 13 May 2002 22:57, Hans Reiser wrote:
> Dieter Nützel wrote:
> >On Monday 13 May 2002 10:07, Hans Reiser wrote:
> >>>>4.- I have a PIII 1 GHz, 512 MB RAM, SCSI server where I run sendmail,
> >>>>bind and squid. Will I have problems with performance if I migrate to
> >>>>reiserfs?
> >>
> >>sendmail will be lower in performance when using reiserfs unless you use
> >>the latest not yet released patches and turned data journaling on.
> >
> >More speed numbers on this to come with related post.
> >
> >>Do you really push the performance limits of the machine as a mail
> >> server?
> >
> >You mean if it is worth?

Here I don't understand you that's why I asking.

> >>squid will benefit from your putting all files in one directory.
> >
> >So I should clobber my squid partition/dir and reconfigure to put all
> > files into only one directory and let the squid hierarchy alone?
> >
> >Thanks,
> >	Dieter
>
> I don't understand you.
>
> You should put squid into one big directory (not using their kludge that
> avoids making directories too large by using a tree)

So you understand me right...;-)

> unless you have multiple processors (in which case the VFS directory giant
> lock will hurt you).

BKL?

Where is the limit 2, 4, or even more CPUs?
I'll have two in some days.

Thanks,
	Dieter

PS You know that I'm on the preemption+lock-break, O(1)-scheduler track...;-)

Re: New to reiserfs

[Chris Mason]

On Mon, 2002-05-13 at 17:24, Dieter Nützel wrote:

> >
> > You should put squid into one big directory (not using their kludge that
> > avoids making directories too large by using a tree)
> 
> So you understand me right...;-)
> 
> > unless you have multiple processors (in which case the VFS directory giant
> > lock will hurt you).
> 
> BKL?
> 
> Where is the limit 2, 4, or even more CPUs?
> I'll have two in some days.

You probably want one directory per squid process.  The VFS layer does
semaphore locking to prevent races between people reading the directory
and people changing the directory.  If you've just got one big
directory, all the squid processes will end up waiting on each other.

More directories mean more time wasted changing the directory stat data,
and ram wasted on reading the stat data.  You see this more with
hundreds or thousands of directories.

-chris