Installation Stalls Under Ubuntu 26.04

[CarloGraziani]

Hello.

The standard installation process, which works fine under Ubuntu 24.04, stalls for difficult-to-diagnose reasons under Ubuntu 26.04. I suspect that pip is at fault, but I can't be sure. Unfortunately I can't find a firedrake-install.log file in the directory where I started the install.

Petsc compilation is unproblematic, and all the tests pass. When I do the pip install in a virgin venv according to the instructions, at some point the pip-related processes go into sleep mode, forever.

I ran the installation with 'pip -v' to see what was going on. I saw the output in the attached file ('pip-v-output.txt')
pip-v-output.txt

There is a (warning?) diagnostic in the compilation of PETSc.o, but I have no idea whether it is significant.

At the time when I issued the SIGKILL, there were sleeping pip-related processes, which never woke up even when I waited for over 20 minutes. Here's some extracts of 'ps' output:
ps-output.txt

I reran the install (after a 'pip cache purge'), using 'pip -vv' to get more information. This results in a huge amount of output that I won't inflict here. The install stalls at the same place. However, after issuing SIGKILL, at least I get a stack trace, which is in the second attached file ('pip-vv-output.txt')
pip-vv-output.txt

When I did this a couple of times, the final line is always the same 'line: str = proc.stdout.readline()'. It looks as if that is where pip stalls: it's blocking on a readline() call.

Not sure what to make of that. It looks like a pip resolution issue. However, when I tried installing with uv ('uv pip...') I got the same stall. I also tried making pip use the legacy resolver, which made no difference.

Any suggestions as to what I could try?

[connorjward]

This may be related to the fact that building petsc4py is very demanding. I would make sure that you aren't running out of RAM. At the same time I recommend setting CFLAGS="-O0" as this has a dramatic impact on how demanding it is to compile things.

[CarloGraziani]

Thanks. I'm quite sure that running out of memory was not the issue here, since this machine is a newish laptop with 16GB, and since I was running 'top' to try and troubleshoot the problem at the time and did not see any evidence of OOM.

I'd be happy to try changing the CFLAGS option, but I'm not sure how to make pip do that. In any event, the process stalled inside pip, and not inside the compilation. It looks as if the mpicc command to create PETSc.cpython-314-x86_64-linux-gnu.so succeeded, and whatever emitted the line 'writing build/lib.linux-x86_64-cpython-314/petsc4py/lib/petsc.cfg' must have succeeded as well, since the process then stops inside of pip.

[connorjward]

You have to run export CFLAGS="-O0" before running pip install. It may also be useful to try running pip cache purge and using a new venv. You could also try running pip install -U pip to update pip itself.

And definitely continue to install with the -v flag. It can give some useful insights.

[CarloGraziani]

Thanks for the continued help. Running with CFLAGS="-O0" in the environment did not change the outcome. I can see the effect in the compilation line of PETSc.o, in that the -O0 occurs after -O3 (and therefore, I think, overrides it). However, in the compilation of PETSc.cpython-314-x86_64-linux-gnu.so, while -O0 follows a -O3 and a -O1, it is itself succeeded by a -O2, so presumably it has no effect. Is that -O2 due to a different flag that can be set in the environment?

Here's the output from 'pip -v':
pip-v-output.txt

[connorjward]

I just remembered that another user reported something similar on Slack some time ago. Can you try the fix that they used:

$ export HWLOC_COMPONENTS=-gl

(source open-mpi/hwloc#517)

[CarloGraziani]

Thank Connor, I can confirm that this works. That environment variable then needs to be set in startup scripts (e.g. .bashrc), because the installation only works if it is present. For example, the 'firedrake-check' also hangs forever without it, but completes successfully with it.