Unhappy with recent make(1) changes
Scott Long
scottl at freebsd.org
Mon Jan 12 16:28:42 PST 2004
On Tue, 13 Jan 2004, Ruslan Ermilov wrote:
> On Tue, Jan 13, 2004 at 12:38:05AM +0200, Ruslan Ermilov wrote:
> [...]
> > : old new main.c + new main.c +
> > : old job.h new job.h
> > :
> > : -B: 10,72s 10,67s 10,64s
> > : 10,57s 10,64s 10,70s
> > : 10,70s 10,80s 10,63s
> > :
> > : -j4: 13,53s 12,15s 1m3,89s
> > : 13,40s 12,02s 1m7,68s
> > : 13,23s 12,70s 1m5,73s
> >
> > old: main.c,v 1.85 job.h,v 1.20
> > new: main.c,v 1.86 job.h,v 1.21
> >
> Hmm, I've tested this on fresh -CURRENT and -STABLE UP boxes,
> and on SMP bento, and I cannot reproduce this behavior with
> new make(1).
>
> It also seems to depend heavily on the location of "src/"
> and of the value of MAKEOBJDIRPREFIX. I get best results
> with MAKEOBJDIRPREFIX pointing to NFS homedir, but the
> result in this case is unstable:
>
> old make new make
>
> /usr/src + 4,99s 29,85s
> /var/tmp/ru/obj
>
> /var/tmp/ru/src + 5,05s 29,80s
> /var/tmp/ru/obj
>
> /usr/src + 5,48s 14,45s 6,65s
> /home/ru/obj
>
> /var/tmp/ru/src + 5,34s 12,84s 2,74s
> /home/ru/obj
>
> In any case, this seems to be an issue specific to
> this machine, so nevermind this post.
>
>
> Cheers,
Try disabling HTT and seeing how that changes the results. Without a
good scheduler, HTT is a pessimization on CPU-intensive tasks like this.
It's quite possible that the mandatory delay that happened in the old
make(1) actually helped the HTT case a bit. When I tested DES's change,
I turned off HTT in order to not be influenced by it. Maybe switching to
ULE (which is HTT-aware, in theory) would help?
Scott
More information about the freebsd-current
mailing list