tech-toolchain archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: build.sh: set creation is serialized



On Wed, Apr 30, 2008 at 09:00:28PM +0100, Andrew Doran wrote:
> On Wed, Apr 30, 2008 at 08:29:02PM +0200, Julio M. Merino Vidal wrote:
> 
> > On Apr 30, 2008, at 20:12 , Andrew Doran wrote:
> > 
> > >In src/src/distrib/sets/Makefile:
> > >
> > >---
> > >revision 1.56
> > >date: 2007/03/10 20:34:57;  author: apb;  state: Exp;  lines: +5 -2
> > >branches:  1.56.4;  1.56.10;  1.56.12;
> > >Serialise making the "sets" tarballs in the maketars target.
> > >On many systems, making them in parallel causes very poor
> > >system performance due to disk I/O.  (Requested by jmmv.)
> > >---
> > >
> > >On my development system this adds a lot of time to a release  
> > >build, and I'd
> > >like to revert it. My system is not affected, but if it were I  
> > >don't see
> > >why any of these would not be a valid solution:
> > >
> > >o use smaller value for -j
> > >o analyze the problem on the affected systems
> > >o buy more spindles
> > 
> > Affected system: laptop, -j2.  Problem analysis: slow disk.   
> > Solution: none.  (Well, I upgraded to 7200RPM after that change was  
> > made, but I am still pretty sure that parallel creation is slower  
> > than serial one here.)
> > 
> > If you want to change it, could it be possible to add some optional  
> > tunable to serialize the set creation?  That'd be enough for me.
> 
> Hmm. How much memory does this system have, and what is kern.maxvnodes?

Disk access is not a major factor on the system I'm testing with,
presumably because maxvnodes and the number of pages dedicated to
file data are fairly high. Perhaps we could tune it at runtime using
sysctl if doing a native build.

Andrew


Home | Main Index | Thread Index | Old Index