building with MAKEFLAGS=-j8

Hi, as promised in another thread on the development mailing list, I
made some tests about building LFS-7.6-rc1 using MAKEFLAGS=-j8.

Here are my findings (you can skip to the bottom if you want to read
just the conclusion).

I stopped at the end of section 6, so I didn't build the kernel, and I
skipped GRUB too, because I won't use it.

My host system is Funtoo current, running on a Lenovo Thinkpad T440s.

I did 2 automated builds with no MAKEFLAGS setting and /mnt/lfs mounted
noatime, 1 with no MAKEFLAGS and /mnt/lfs mounted atime, and 3 with
noatime and MAKEFLAGS=-j8 exported as environment variable for both root
and the lfs user.

As noted in the book, the following tests are already known to require
-j1, and I built them by the book:
    attr (chapter 6.21)
    acl  (chapter 6.22)
    vim  (chapter 6.70)

My build times for the temporary tools (including stripping, chapters
5.4 - 5.35, about 33 SBUs according to the book) have been:
    no MAKEFLAGS  : ~ 48m
    MAKEFLAGS=-j8 : ~ 31m

My build times for the base system (excluding grub, chapters 6.7 - 6.70,
about 177 SBUs according to the book) have been:
    no MAKEFLAGS  : ~ 4h 20m
    MAKEFLAGS=-j8 : ~ 2h 30m

Mounting atime didn't produce any significant difference.

Here are the relevant differences that I found in the tests logs:

====================
glibc (chapter 6.09)
--------------------
    In some (not all) of the noatime builds I had several warnings like:

        make[2]: Warning: File '/dev/null' has modification time 2368
s in the future
        make[2]: warning:  Clock skew detected.  Your build may be incomplete.

    I didn't other atime builds to check the consistency.

    In some tests I also had some failures that seem uncorrelated with
    the -j settings, and uncorrelated with the fails, too.


==================
gcc (chapter 6.17)
------------------
    In one of the "no MAKEFLAGS" builds, I had 3 unexpected failures in
    the "gcc tests".

    The number of passed tests differ (it's bigger with -j8), even
    though the expected failures and the skipped ones are the same:

    no MAKEFLAGS :

            === libstdc++ Summary ===
        # of expected passes      9804
        # of expected failures    41
        # of unsupported tests    277

    MAKEFLAGS=-j8 :

            === libstdc++ Summary ===
        # of expected passes      9823
        # of expected failures    41
        # of unsupported tests    277


==========================================
coreutils, test as "nobody" (chapter 6.29)
------------------------------------------
    There are two summaries in the test log. The first ones differ,
    while the seconds are the same:

    no MAKEFLAGS :
        ========================================
        Testsuite summary for GNU coreutils 8.23
        ========================================
        # TOTAL: 564
        # PASS:  446
        # SKIP:  118
        # XFAIL: 0
        # FAIL:  0
        # XPASS: 0
        # ERROR: 0

        ========================================
        Testsuite summary for GNU coreutils 8.23
        ========================================
        # TOTAL: 306
        # PASS:  293
        # SKIP:  13
        # XFAIL: 0
        # FAIL:  0
        # XPASS: 0
        # ERROR: 0

    MAKEFLAGS=-j8 :
        ========================================
        Testsuite summary for GNU coreutils 8.23
        ========================================
        # TOTAL: 564
        # PASS:  442
        # SKIP:  122
        # XFAIL: 0
        # FAIL:  0
        # XPASS: 0
        # ERROR: 0

        ========================================
        Testsuite summary for GNU coreutils 8.23
        ========================================
        # TOTAL: 306
        # PASS:  293
        # SKIP:  13
        # XFAIL: 0
        # FAIL:  0
        # XPASS: 0
        # ERROR: 0


===================
perl (chapter 6.42)
-------------------
    The number of "tests" is always different from one build to the
    next, regardless of the -j setting:

        u=3.93  s=1.01  cu=253.65  cs=15.37  scripts=2258  tests=698696

    I suppose the other numbers are times, so I expect them to change,
    and "scripts" is always the same, but "tests"?  I don't know.

    With "MAKEFLAGS=-j8" I had many warnings like:

        make[1]: warning: jobserver unavailable: using -j1.  Add '+'
to parent make rule.


====================
gperf (chapter 6.51)
--------------------
    With "MAKEFLAGS=-j8" this test shows some failures like:

        gcc: error: test.o: No such file or directory
        Makefile:83: recipe for target 'check-ada' failed
        make[1]: *** [check-ada] Error 1

    and finally fails with:

        Makefile:57: recipe for target 'check' failed
        make: *** [check] Error 2


========================================================================
CONCLUSIONS
-----------
The speed-up with -j8 is huge, and the only tests that actually fail are
attr, acl, vim and gperf.  coreutils skips 4 more tests with -j8.

For all of these tests, adding -j1 to the test line fixes the issues,
even when the build used -j8, which means that the build phase is
actually OK with -j8, only the tests need -j1.


-- Emanuele Rusconi
-- 
http://lists.linuxfromscratch.org/listinfo/lfs-chat
FAQ: http://www.linuxfromscratch.org/blfs/faq.html
Unsubscribe: See the above information page