Re: OpenMPI 5.0 to be 32-bit only ?
On 15/02/2023 00:16, Steve Langasek wrote:
On Mon, Feb 13, 2023 at 02:18:06PM +0000, Alastair McKinstry wrote:
On 13/02/2023 12:51, Adrian Bunk wrote:
There are a significant number of science libraries dependent on MPI.
We would need to do MPI-free builds of these libraries; I'm not sure how
much breaks as we do.
Would we, though? Or should we remove the 32-bit builds of those libraries
I think it's accurate that no one is using those scientific libraries in
production (which is, basically: doing lots of matrix math) on 32-bit
architectures, because all of the vector instructions you want for such work
are only available on 64-bit CPUs.
So the only application of those 32-bit binaries, really, is either a)
letting users of those 32-bit archs learn the tools on the hardware they
have available, so they can use them to advantage later on fit-for-purpose
hardware; or b) using them to build other software in Debian.
Is either of those a compelling reason to keep building those software
stacks for 32-bit? I would argue not. But neither is it obvious at what
point it's worth the effort to remove them, since this requires tracking the
reverse-dependency tree, working out which of those reverse-dependencies are
*not* scientific applications that should drop the build-dependency rather
than being removed, and so forth.
So it's a tradeoff between the maintenance work of keeping mpi working on
32-bit, and the one-time work of removing it.
Thats a more drastic solution, but also realistic. There are a number of
scientific libraries and packages that will never realistically be used;
they're only being built and tested for code quality checking, but add a
technical maintenance cost.
The counterpoint is if someone does a high-core-count 32-bit arch for
HPC; x32 could (have been) such an architecture, but its development
ph: +353 87 6847928 e: email@example.com, im: @sceal.ie:mckinstry