On 9/11/13 5:04 AM, Paul Eggleton wrote:
Hi Mark,
On Tuesday 10 September 2013 21:04:29 Mark Hatle wrote:
Time measurements with:
MACHINE = "qemux86"
PACKAGE_CLASSES = "package_rpm"
EXTRA_IMAGE_FEATURES = "dev-pkgs staticdev-pkgs doc-pkgs dbg-pkgs
ptest-pkgs" BB_NUMBER_THREADS ?= "8"
PARALLEL_MAKE ?= "-j 8"
image: core-image-sato
Quad Core i7 workstation...
With all of the recipes built, and just executing the do_rootfs:
Before this change:
real 26m2.541s
user 21m8.458s
sys 7m9.683s
After this change:
real 14m43.436s
user 15m57.525s
sys 5m16.412s
This is a significant performance boost when using the 'complementary'
installs and attemptonly installs. The performance is by doing the
complementary installs as a batch instead of one at a time, ignoring
failures.
This is a significant improvement, well done! My concern about not catching
failures stands though, which is why I hadn't implemented this yet. How do we
ensure that genuine failures are caught with this mode?
The code only ignores failures in the resolution of the dependencies. As far as
I'm aware, this is the correct behavior in the attempt only mode(s). If the
dependencies can't be resolved for some reason, we simply throw away that item
and move on.
Any errors that occur during package installation -will- now cause a failure to
exist, and they will be logged to the attemptonly log-file.
In the previous implementation, any error when installing an attemptonly package
would be ignored, even if it turned out to be an installation problem and not a
dependency issue. So I think this is an improvement as well.
(For some of my testing I intentionally created dependency errors by adding
things to the PACKAGE_EXCLUDE, and the system did what it was supposed to.
Ignored the items that couldn't be installed.)
--Mark
Cheers,
Paul
_______________________________________________
Openembedded-core mailing list
[email protected]
http://lists.openembedded.org/mailman/listinfo/openembedded-core
