Re: lockf in installworld -- not a good idea

On Fri, Sep 29, 2006 at 10:20:53PM +0400, Ruslan Ermilov wrote:
On Fri, Sep 29, 2006 at 11:09:42AM -0600, M. Warner Losh wrote:
[...]
I tend to agree with that basic philosophy. From other items in the
thread, it was clear this came up in the context of build release,
which benefits from -j usually. The installworld phase in that should
be as robust as possible as well, since otherwise we have issues with
the actual release. Unless it is a big win (more than a few percent),
I'd imagine the right fix is to the release target to not do a
parallel installworld. I know that in the build scripts that I wrote
in 3.x days and have ported forward since then I've never done a
parallel install, due to it rarely working reliably in that (long)
time span...

They are safe to do nowadays. I'll do some measurements on real
SMP with the memory-based DESTDIR, and let you know the numbers.

The data (real time seconds) was taken on the following system:

CPU: AMD Athlon(tm) 64 X2 Dual Core Processor 3800+ (1995.01-MHz K8-class CPU)
real memory = 4227792896 (4031 MB)
avail memory = 4081033216 (3891 MB)
FreeBSD/SMP: Multiprocessor System Detected: 2 CPUs

Basically, doing installworld -jX doesn't make any sense.

: x installworld-B
: + installworld-j1
: +--------------------------------------------------------------------------+
: |xxx + + +|
: ||A| |_MA___||
: +--------------------------------------------------------------------------+
: N Min Max Median Avg Stddev
: x 3 40.21 40.82 40.53 40.52 0.30512293
: + 3 56.72 58.45 57.26 57.476667 0.88511769
: Difference at 95.0% confidence
: 16.9567 +/- 1.50052
: 41.8476% +/- 3.70317%
: (Student's t, pooled s = 0.662017)

: x installworld-B
: + installworld-j2
: +--------------------------------------------------------------------------+
: |xx x + + +|
: ||A_| |___AM___||
: +--------------------------------------------------------------------------+
: N Min Max Median Avg Stddev
: x 3 40.21 40.82 40.53 40.52 0.30512293
: + 3 55.17 57.29 56.44 56.3 1.0669114
: Difference at 95.0% confidence
: 15.78 +/- 1.77852
: 38.9437% +/- 4.38924%
: (Student's t, pooled s = 0.784666)

: x installworld-B
: + installworld-j4
: +--------------------------------------------------------------------------+
: |xxx + + + |
: ||A| |_____A_M___||
: +--------------------------------------------------------------------------+
: N Min Max Median Avg Stddev
: x 3 40.21 40.82 40.53 40.52 0.30512293
: + 3 59.44 63.28 62.18 61.633333 1.9775068
: Difference at 95.0% confidence
: 21.1133 +/- 3.2069
: 52.106% +/- 7.91437%
: (Student's t, pooled s = 1.41486)


Cheers,
--
Ruslan Ermilov
ru@xxxxxxxxxxx
FreeBSD committer

Attachment: pgphVKOnqC99o.pgp
Description: PGP signature

Relevant Pages

  • Re: Installworld broken on RELENG_6 by libc commit?
    ... My installworld of RELENG_6 from a few hours ago failed with this ... error (from memory): ... This broke everything that depended on libncurses, ... so it should be no problem to overwrite libc. ...
    (freebsd-stable)
  • Re: Make installworld problems with BETA4 through 6
    ... > And I seem to have found the reason. ... at least I wasn't able to reproduce any ... a complete "installworld" output would help narrow this problem ... FreeBSD committer ...
    (freebsd-current)
  • Re: BETA1 5.3 make world DESTDIR failure
    ... and specify DESTDIR when doing make installworld. ... I'm finishing testing a patch for Makefile.inc1. ... Ruslan Ermilov ... FreeBSD committer ...
    (freebsd-current)
  • Re: make installworld for non-root user
    ... retrieving revision 1.6 ... Not until you call installworld with "make -fMakefile.inc1 ...". ... > during buildworld and just install the compiled files during installworld. ... FreeBSD committer ...
    (freebsd-hackers)
  • Re: bind9: /etc/namedb symlink problem
    ... > It looks like the first time you do a make installworld it works right. ... > causes the make to fail. ... FreeBSD committer ...
    (freebsd-current)