Suitesparse bug on Ubuntu 20.04

Since I updated Ubuntu from version 19.10 to 20.04 the tests umfpackamgtest, umfpackfastamgtest, umfpacktest and spqrtest are all failing due to a timeout error. The Suitesparse library is hereby the (only) Ubuntu-Repository version for Ubuntu 20.04 (and got updated together with Ubuntu from version 5.4.0 to 5.7.1).

In my concrete case this means that my code is getting compiled without error, but then it is just hanging when trying to run UMFPack and/or Cholmod (which where working perfectly fine on Ubuntu 19.10 with Suitesparse 5.4.0).

Did you do a full rebuild? Might it be that the ABI changed and you observe some infinite loop or similar?

Yes, I even ran the attached script to make a fresh installation and run the tests in a seperate folder. It still produces the same result.test-dune-istl-master.sh

Can you try:

1. run without -j4
2. explicitly add an optimization flag (e.g. -O2)

cmake does not enable any optimization by default. Perhaps there was some implicit -O flag pulled in via one of the dependencies(?!) and not anymore since the update.

If this also does not help, you should try to attach a debugger to the slow/long running process and try to find out where you are spending all the time.

I tried to reproduce this in an Ubuntu 20.04 Docker container, but the test passes (without adding any optimization flags). While discussing the issue with @simon.boisseree, he told me that umfpacktest also hangs, if run separately, so -j4 has no influence.

changed the description

Without -j4 and with -O2 nothing changes and if I use gdb to debug (here I just tested umfpacktest) it hangs again and I get the attached output after pressing CTRL-C and backtracing. gdb_output.txt

Maybe this bug is related:
https://bugs.launchpad.net/ubuntu/+source/suitesparse/+bug/1843277
It doesn't help much, as no details are included. I couldn't find the related Debian bug.

@simon.boisseree Following @gruenich's guess you may try running the test with

OPENBLAS_NUM_THREADS=1 OMP_THREAD_LIMIT=1 ./umfpacktest

That seems to be it, because with this option all tests are running through without any problem.

closed

reopened

I think that disabling multithreading is a workaround but not a solution. Since the real problem persists, I reopened the issue. It may also be helpful to others to have this documented, if more people run into the same problem once they update their Ubuntu distribution.

Digging deeper, I found a related issue/discussion in SuiteSpare's GitHub
https://github.com/DrTimothyAldenDavis/SuiteSparse/issues/1
It seems to be an issue of openBLAS. OpenBLAS' threading conflicts with the one from SuiteSparse, cf. https://github.com/xianyi/OpenBLAS/issues/2392

I tried it by installing intel-mkl and reinstalling libsuitesparse-dev and now everything seems to work again.

Maybe a small update: After I updated the openblas package in Ubuntu 20.04 (which was probably released a few days ago) everything works again without intel-mkl.

Did you do a full rebuild? Might it be that the ABI changed and you observe some infinite loop or similar?

Yes, I did a full rebuild, but since the update of openblas everything worked again either way.

So, can we close this? This is no Dune issue, we cannot fix it. If someone thinks it's important, we can make this a known issue.

From my side that's totally fine, I think it was just left open because it may have affected other users who updated to Ubuntu 20.04. But now this shouldn'd be the case either way.

Should be listed as known issue.

mentioned in commit 3bd109fa

mentioned in merge request !437 (merged)

mentioned in commit 3fb00693

mentioned in commit 48d87858

closed with commit 3fb00693

closed with merge request !437 (merged)