rtems-bsp-builder Job Load Calculation

[Chris Johns]

On 2020-04-04 12:28, Joel Sherrill wrote:
> On Fri, Apr 3, 2020 at 6:23 PM Joel Sherrill <joel at rtems.org 
> <mailto:joel at rtems.org>> wrote:
>     On Fri, Apr 3, 2020 at 5:26 PM Chris Johns <chrisj at rtems.org
>     <mailto:chrisj at rtems.org>> wrote:
>         On 2020-04-04 07:57, Joel Sherrill wrote:
>          > How does rtems-bsp-builder calculate the amount of
>         parallelism to use?
> 
>         The option is documented here ...
> 
>         https://docs.rtems.org/branches/master/user/tools/bsp-builder.html#cmdoption-jobs
> 
>         and is ...
> 
>            --jobs
> 
>            The jobs options where the format is build-jobs/make-jobs.
>         The default
>            is 1/num-cores where num-cores is the operating system
>         reported number
>            of cores.
> 
> 
>     What's the definition of build job vs make job?
> 
>     Is make jobs equivalent to N in make -jN?
> 
>          > I ask because I have realized that I selected jobs=6/12 on a
>         very beefy
>          > machine with 12 cores and 32 GB RAM. The main test machine I
>         use is 8
>          > cores/32 GB RAM. But the slowest ones are old with 4 cores/4
>         GB RAM.
>          > They are hitting the swap space. All have SSDs.
>          >
>          > Should I drop the --jobs parameter? Will the script select a
>         decent load
>          > automatically? The machines are essentially dedicated to testing.
> 
>         You will need a --jobs option or the machine will not be fully
>         loaded.
>         You need to find a balance between the number of parallel BSPs
>         that are
>         built at once and the number of parallel build jobs each BSP
>         makes runs.
> 
>         The critical factor is the amount of bootstrap and configure
>         time. This
>         is a single core task and if you use the default a lot of the
>         build time
>         the machine is basically idle.
> 
> 
>     But if you go too high, the number of compiles will swamp a machine
>     without a huge amount of memory. So the 4 core/4 GB machines
>     probably would be something like --jobs=2/4. if that's two bsp builds
>     in parallel with a -j4.
> 
>     That generally gives you a 2x factor on processes which is close
>     to the old school rule of the -jN number should be between 1.5 and
>     2 times the number of cores as long as the RAM and disk subsystem
>     could handle it. That tends to overlap IO and computation.
> 
>     Maybe the default should be (roundup(cores)/2) instead of 1.
> 
>     Temporarily I may have to add a jobs option to my script and hand-tune
>     it to the machines I am testing on.
> 
> 
>         There is no way I know of to automatically find this balance for
>         all the
>         different types of host machines.
> 
> 
> Thinking a bit more, I wonder if there could be some named
> build/job ratios. 1/cores, (cores/2)/cores, etc.

As long and we keep the fact one is the number of jobs given to a build 
or `make` or whatever for a package. I am happy to review a patch.

Chris