Re: Behavior of --use-cpus-instead-of-cores

[Joe Sapp]

Would you consider including the patch in the original message in the
next release?  Limiting the number of jobs to the number of physical
sockets is unnecessarily conservative and it seems like it wasn't the
intention.

-- 
Thanks,
Joe

On Tue, Apr 17, 2018 at 2:38 PM, Joe Sapp <sappj@ieee.org> wrote:
> The option --use-cpus-instead-of-cores causes Parallel to count the
> number of physical sockets on the machine instead of what some might
> call the number of physical CPUs (what lscpu calls "cores" -- there is
> varied use of terminology out there).  Without this option, as
> expected, Parallel counts the number of CPUs, virtual or not (what
> lscpu calls "CPUs").  If these are the intended behaviors, disregard
> below.
>
> It seems the intention of this option was to only allow using up to
> the number of CPUs with independent resources (i.e., disregard virtual
> CPU technologies like hyperthreading).  I think then the number of
> CPUs should be determined as the number of physical cores on each
> socket times the number of sockets, like `lscpu -p | egrep -v '^#' |
> sort -u -t, -k 2,4 | wc -l` [1].  If so, it seems there is an easy fix
> for GNU/Linux; a patch to do so is attached.
>
> --
> Joe Sapp
>
> [1] https://stackoverflow.com/a/23378780/610375