Hi,
Are you guys still there?
I think that we are saving about 30 min to an hour on the nightly builds now… Did anybody else run some benchmarks?
-Michael
On 6 Mar 2019, at 16:36, Michael Tremer michael.tremer@ipfire.org wrote:
Hi,
Sorry for not being able to get around to this earlier, but I have invested some time on this:
First of all, I merged Wolfgang’s patch. He is right, the build is too slow. It sometimes looks a bit like it is not taking enough advantage of fast systems. On the other hand with have small systems, which are (in their own right) quite fast. Those are the ARM systems. They usually have 1G of RAM. 2GB when you are lucky.
So I adjusted make.sh a little bit and removed composing the MAKETUNING variable from it.
Now, make.sh only figures out the value that usually comes after the -j. It is set in DEFAULT_PARALLELISM.
In the individual LFS files, we can now set a new variable called MAX_PARALLELISM which is optional. In boost, I now calculate this from the amount of memory that is available. That will make sure that we have a cap on this that will be high on systems that have the memory and low on those that don’t. Before, we hardcoded -j2 which of course runs very slow on large systems.
The code is here:
https://git.ipfire.org/?p=people/ms/ipfire-2.x.git;a=shortlog;h=refs/heads/f...
Please review. I am now running this on a couple of builders and will see what comes out. Please do the same and send me feedback.
Best, -Michael
On 21 Feb 2019, at 17:28, Wolfgang Apolinarski wolfgang.apolinarski@ipfire.org wrote:
Hi Arne,
of course I did delete the ccache between my multiple builds to verify the commit. To play it safe, I hard reset using git, checked that all folders (build, ccache, etc.) are gone and restarted the build after gettoolchain and downloadsrc.
Sorry, I did not know that there are still a lot of machines with low ram that we are using to build ipfire. My hope was that these machines have disappeared over the last years. Additionally, I do not have the ability to test all ipfire arches, as mentioned, the patch needs testing by others.
I also thought of adding a manual parameter for developers that somehow manually sets the parameters for parallel build, but since the MAKETUNING parameter already exists and is used with other packages this would just somehow mess everything up.
For my builds, I can just manually apply the patch to speed up the full build. My thought was that sharing the patch would make sense for others.
-----Ursprüngliche Nachricht----- Von: Arne Fitzenreiter Arne.Fitzenreiter@ipfire.org Gesendet: Donnerstag, 21. Februar 2019 09:44 An: Wolfgang Apolinarski wolfgang.apolinarski@ipfire.org Cc: development@lists.ipfire.org Betreff: Re: [PATCH 0/1] Parallelized build for several packages
On the most of this packages (eg. boot and cmake) i have disabled parallel build because it fails on machines with low ram (less than 2GB) or on arm. So such changes need tested on an 1GB machine, on all architectures and build two times after deleting ccache. Because the build need much lower resources if it has the file already prebuilt in cache.