Firedrake meeting 2020 11 04

Date and time 2020-11-04 16:00UTC (16:00GMT)

Action Items

1. Pick Chair and Minuter.
2. ALL: (ongoing) triage the open issues and confirm if they are indeed still open (and perhaps provide labels)
3. ??: Build master on centos to catch errors
4. DS: Anaconda on Mac

Agenda

Present: David Ham, Lawrence Mitchell, Rob Kirby, Colin Cotter, Jack Betteridge, Connor Ward, Sophia Vorderwuelbecke, Paul Kelly, Stephan Kremar, Koki Sagiyama

Apologies: Reuben Nixon-Hill

JB: Numpy Nightmare

Issues:

• linking to old libgfortran -> link agaist the right openblas from homebrew
• numpy import time error (RK and CJC have observed on IPython).
• A lot of wheels are missing in Python3.9 -> Suggest that users use 3.8
• JB has observed some stochastic Xcode errors...

Good thing about firedrake:

• Platform independent install mechanism (except on super-computer)
• Incremental rebuild.

Build conda package?

• will need completely different system on supercomputer (bad).

Short term solution suggested by JB:

• Worked for RK
• JB: Have CJC check if this solution works.
• JB: Put the recipe of this solution on this minute:

Run

export DYLD_INSERT_LIBRARIES=/usr/local/opt/openblas/lib/libblas.dylib:/usr/local/opt/openblas/lib/liblapack.dylib

in a terminal before importing Firedrake, or add to Firedrake activate script.

Strange test failures (from last week)

Update

Questions:

• Does it always fail on the same build hardware but not other build hardware?
• Is SNES deterministic?
• Could the FPU on one of the builders be bad?

Summary:

• PYOP2_DEBUG=1 -> pass
• vector registers in Intel chips have a lower precision than the scalar ones?

CW submitted PR:

• increased snes_max_it
• see https://github.com/firedrakeproject/firedrake/pull/1898

CW/LM: PyOP2 documentation

It's very out of date. Should we remove it?

PyOP2 documentation, some are wrong, others are not.

Action:

CW: Push sphinx documentation when test (LM: auto-build with github action is better?)

CW/LM: Random test failures

See, e.g., https://github.com/firedrakeproject/firedrake/pull/1898

First observed around when:

• complex landed,
• default solver changed (mumps global variable clean up issue?).

Action:

LM: make superlu default and see if it magically fixes the issue.

??: set omp_num_threads

Merge PRs:

AOB

Date of next meeting

2020-11-11 16:00UTC (16:00GMT)