Re: [ptxdist] Speed up targetinstall of certain packages

[Michael Olbrich <m.olbrich@pengutronix.de>]

Hi,

On Tue, Nov 08, 2022 at 01:10:32PM +0100, Christian Melki wrote:
> So I did an experiment with bash loadables (builtins).
> Turns out bash has a bunch of builtin examples.
> 
> I had ptxdist build me a bash-5.2 for x86_64.
> With the "loadables" make target.
> 
> $ diff -urN ../lib/ptxdist-2022.11.0/rules/bash.make rules/bash.make
> --- ../lib/ptxdist-2022.11.0/rules/bash.make	2022-10-21 16:54:50.000000000 +0200
> +++ rules/bash.make	2022-11-08 12:51:38.609794625 +0100
> @@ -77,6 +77,16 @@
>  	--$(call ptx/wwo, PTXCONF_BASH_CURSES)-curses
>  # ----------------------------------------------------------------------------
> +# Compile
> +# ----------------------------------------------------------------------------
> +
> +$(STATEDIR)/bash.compile:
> +	@$(call targetinfo)
> +	@$(call world/compile, BASH)
> +	@$(call compile, BASH, loadables)
> +	@$(call touch)
> +
> +# ----------------------------------------------------------------------------
> 
> Since mkdir is slightly less than half of the external binary calls I had it replaced.
> I copied the scripts/lib/ptxd_make_xpkg_pkg.sh to my build path
> and created a shell-loadables directory there.
> 
> $ cp platform-secplatform-x86_64/build-target/bash-5.2/examples/loadables/mkdir scripts/lib/shell-loadables/
> 
> $ ls -1  scripts/lib/shell-loadables/
> mkdir
> 
> $ diff -urN ../lib/ptxdist-2022.11.0/scripts/lib/ptxd_make_xpkg_pkg.sh scripts/lib/ptxd_make_xpkg_pkg.sh
> --- ../lib/ptxdist-2022.11.0/scripts/lib/ptxd_make_xpkg_pkg.sh	2022-10-21 16:54:50.000000000 +0200
> +++ scripts/lib/ptxd_make_xpkg_pkg.sh	2022-11-08 13:08:18.074550829 +0100
> @@ -682,6 +682,8 @@
>  ptxd_install_file() {
>      local cmd="file"
> +    #export PS4='+[${EPOCHREALTIME}][${BASH_SOURCE}:${LINENO}]: ${FUNCNAME[0]:+${FUNCNAME[0]}(): }'; set -x;
> +    enable -f scripts/lib/shell-loadables/mkdir mkdir
>      ptxd_install_file_impl "$@" ||
>      ptxd_install_error "install_file failed!"
>  }
> 
> without mkdir builtin:
> finished target timezone.targetinstall.post
> 
> real	0m17,043s
> user	0m13,522s
> sys	0m4,138s
> 
> with mkdir builtin:
> finished target timezone.targetinstall.post
> 
> real	0m12,345s
> user	0m9,899s
> sys	0m2,916s

So loadable builtins in bash is interesting, but not something we can rely
on at this time. But this whole thing got me thinking. Most mkdirs are
don't actually do anything. They are just in case the directory does not
exist yet. So we can do that:

if [ ! -d  "${dir}" ]; then
	mkdir -p "${dir}"
fi

That avoid most of the mkdirs and the speedup is considerably. Similar to
the builtin I think.  And I found some more things to improve:

The 'flock' is only needed when building packages in parallel, so we can
skip it for things like "ptxdist drop foo.compile; ptxdist targetinstall
foo", and that's where the slow targetinstall hurts the moset.

Do the chmod/chown with the 'install' when possible.

And some stuff to improve the ptxdist startup time.

For "time ptxdist targetinstall timezone" I get:

Before:
real    0m29.167s
user    0m19.389s
sys     0m9.903s

After:
real    0m8.887s
user    0m5.954s
sys     0m2.470s

This stuff should hit master pretty soon.

Thanks for the inspiration :-).

Michael

> On 11/8/22 11:13 AM, Christian Melki wrote:
> > 
> > 
> > On 11/4/22 8:12 PM, Alexander Dahl wrote:
> > > Hello Christian,
> > > 
> > > Am Fri, Nov 04, 2022 at 03:37:02PM +0000 schrieb Gieseler, Christian:
> > > > Hello,
> > > > 
> > > > I have question regarding the speedup of daily work.
> > > > 
> > > > We have frontend and backend of our webgui deployed with separate packages. Only task of these package is to deploy the files with
> > > > 
> > > > @$(call install_tree, web-frontend, -, -, $(WEB_FRONTEND_DIR)/var-www/, /var/www/,no)
> > > > 
> > > > Compile and install stages are empty. The just call targetinfo and touch to skip the stages.
> > > > 
> > > > The frontend depends on the backend and the backend obviously depends on our application which is called by the backend.
> > > > So our web-frontend.in file looks like this:
> > > > ## SECTION=project_specific
> > > > 
> > > > config WEB_FRONTEND
> > > >     bool
> > > >     select APP_LAYER
> > > >     select WEB_BACKEND
> > > >     prompt "e-mode Web Frontend"
> > > >     help
> > > > 
> > > > As expected if i clean and compile APP_LAYER the targetinstallstage of Backend and Frontend are executed again. However this is only a Run-Time only dependency. It is a third-party archive and install_tree takes quite some time even on fast build hosts. Even it if is just a minute it is annoying to spend the time waiting during image creation. Trying to solve that i found "if RUNTIME"  für Run-Time only Dependencys in the documentation here:
> > > > 
> > > > https://www.ptxdist.org/doc/daily_work_section.html#controlling-package-dependencies-in-more-detail
> > > > 
> > > > So my expectation would be that if i change the webfrontend.in file like this:
> > > > 
> > > > config WEB_FRONTEND
> > > >     bool
> > > >     select APP_LAYER    if RUNTIME
> > > >     select WEB_BACKEND   if RUNTIME
> > > >     prompt "e-mode Web Frontend"
> > > >     help
> > > 
> > > That sounds reasonable and I would have done it the same.
> > > 
> > > > The "if RUNTIME" would make sure that the targetinstall stage is not executed again if i just execute a "ptxdist clean app-layer" followed by a "ptxdist images". Same with ptxdist clean root; ptxdist images. It is clear that all targetinstall stages are executed again, but i would expect that the web-frontend is deployed earlier if no build dependency is given.
> > > > 
> > > > Am i missing something, oder is the "if RUNTIME" Switch not working properly in my ptxdist-2018.12 version? Or does it have no effect on targetinstall stages?
> > > 
> > > Not sure how that should behave.  However if you want to speed up the
> > > build: make sure you call ptxdist with -q or --quiet parameter.  The
> > > output on screen takes suprisingly much time, even with modern
> > > terminals, and especially when doing targetinstall of many many files
> > > (as usually the case with web frontends. been there, done that.)
> > > 
> > > Greets
> > > Alex
> > > 
> > 
> > I have a slight disagreement here. I don't think the console is slow.
> > So I did some investigation, mostly since the slowness bugs me too.
> > I did a:
> > export PS4='+[${EPOCHREALTIME}][${BASH_SOURCE}:${LINENO}]: ${FUNCNAME[0]:+${FUNCNAME[0]}(): }'; set -x;
> > as time measurement and trace in the shellscript in question (mostly scripts/lib/ptxd_make_xpkg_pkg.sh).
> > This pretty linear progressive cpu consumption, albeit the big chunks were due to calling of external binaries.
> > 
> > Then I did a timezone package clean and targetinstall.
> > When targetinstalling a zone there was calls to several binaries.
> > 
> > For one zone I had external calls to (in order):
> > echo, mkdir, printf(?), flock, ls, rm, mkdir, flock, mkdir, flock, mkdir, flock, mkdir, flock, mkdir, flock, mkdir, mkdir, mkdir, mkdir, mkdir, install, install, chmod, chmod, chown, echo.
> > 
> > Some of these could be internal builtins, but the consumed time suggested otherwise.
> > Either way. Each install took about 26 ms and I could account the majority of that time in forking external programs and waiting for them to return.
> > 
> > So my conclusion is: The whole thing is a bit slow and bash doesn't help.
> > 
> > Regards,
> > Christian
> > 
> > > > 
> > > > Thanks for any feedback.
> > > > BR,
> > > > Christian
> > > 
> > 
> 
> 

-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |