Re: ping delay (was: Re: [uml-devel] Does HZ really need to default to 20?)

[Steve T. <stevegt@TerraLuna.Org>]

On Wed, Jun 27, 2001 at 10:38:34AM -0500, Jeff Dike wrote:
> stevegt@TerraLuna.Org said:
> > What host kernel version are you running above?  From what I've seen
> > so far, on host RH 7 distros w/2.2.X, ping works badly with HZ=20, and
> > works great with HZ=100.   
> 
> My last reply was actually slightly unfair :-)  

It's okay -- I survived; builds character.  ;-)

> I see the ping problem on one of my filesystems (debian) and not on
> another (slackware) - the log I showed you was from the slackware
> fs.  So, that's holding the host and uml version constant and
> varying the filesystem.  So, I agree that there's a problem here,
> but it's not caused by the choice of HZ (although setting HZ == 100
> may cover it over somehow).

Looks like a debian thing then.  Is your debian filesystem potato
and/or related to the root_fs_debian2.2_small you have posted on
sourceforge?  

> > Did you get a chance to do any of that yet?  This, coupled with
> > some stuff you said last year about wanting to get rid of the
> > tracing thread, sounded like that's where you were going with it.   
> 
> That is where I'm going, but the support doesn't exist yet, either
> in the host or UML.

Cool.  Okay -- let us know if you ever need any help with either
politics or technology to make it happen; this sounds like it would be
a huge performance boost (Factor of 2? 3?  Am I thinking right?).  

The other major improvement would still be multithreading block device
I/O, right?  

> > Has anyone already done any statistical comparisons of the impact
> > of different uml HZ values on the host machine?  I.E. I wonder how
> > much of a performance difference 20 vs. 50 vs. 100 really makes? 
> 
> Not that I know of.

I'll see what I can do.

Steve
-- 
                        .       .    `   *    
Steve Traugott   ` .  *  +                       
Infrastructure Architect            + `     
stevegt@TerraLuna.Org    '   *  .   '  +`   *   
http://www.stevegt.com/