Adding sysimage generation to mpi tests

[amartinhuertas]

Solve Medium Priority Task in #39

IMPORTANT NOTE: This PR MUST BE reviewed after #47

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 0.00%. Comparing base (73de176) to head (120e0a4).

Additional details and impacted files

@@             Coverage Diff             @@
##           release-0.2     #48   +/-   ##
===========================================
  Coverage         0.00%   0.00%           
===========================================
  Files               33      33           
  Lines             3205    3205           
===========================================
  Misses            3205    3205           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[fverdugo]

I am not sure about this. Introduces considerable complexity and it has not a positive impact in the test times. In fact, by looking the gh actions it seems to even lead to slower times.

Perhaps separating the sequential and mpi tests is enough for now.

[amartinhuertas]

I am not sure about this. Introduces considerable complexity and it has not a positive impact in the test times. In fact, by looking the gh actions it seems to even lead to slower times.

Yes, by now it does not improve much. If we want to perform tests with different values of the number of processors (apart from 1 and 4), this will be very helfpul.

[amartinhuertas]

I am not sure about this. Introduces considerable complexity and it has not a positive impact in the test times. In fact, by looking the gh actions it seems to even lead to slower times.

Besides, we will be evaluating the ability to precompile GridapDistributed.jl. Such action for clusters is a must.

[fverdugo]

Besides, we will be evaluating the ability to precompile GridapDistributed.jl. Such action for clusters is a must.

yes, but precompilation needs to be done in the final app, because we cannot anticipate which other packages is going to call the user.

With PackageCompiler generating a system image is straight forward once you have the the final app. I think, it is not our job to provide a precompilaiton mechanism having PackageCompiler at our disposal.

In any case, perhaps it is a good idea to create a new repo GridapBenchmarks or whatever, where we can add both serial and parallel benchs. And we definitively want precompilation scripts in that repo.

[amartinhuertas]

I think it makes a lot of sense to test the workflow that you are going to face when dealing with a cluster. The repo of apps is going to be abandoned, and not tested frequently (we have already many other experiences of similar repos).

[fverdugo]

I think it makes a lot of sense to test the workflow that you are going to face when dealing with a cluster. The repo of apps is going to be abandoned, and not tested frequently (we have already many other experiences of similar repos).

This is a very good point! but as the sysimage is created now, it is not testing the actual workflow that one needs to follow in a cluster. To check the workflow, you need an app, i.e., you need a (or some) "main" function(s) inside a Package and then create a sys image for those functions.

We have several options:

1. Move the "main" functions of the tests inside the src dir and then create a system image by calling them on a single MPI task
2. Create a dummy "test app" in the test dir and move the "main" test functions to the src dir of the "test app" and then precompile the "test app".

I would go for 2 definitively, even though it has some more boilerplate code, since it actually mimics what a user needs to do in a cluster.

What do you think?

[amartinhuertas]

yes, but precompilation needs to be done in the final app, because we cannot anticipate which other packages is going to call the user.

I agree the solution that we currently have in this branch is not definitive, in the sense that it is not testing exactly what will happen with the actual app, with all dependencies etc.. But it may happen that even with the current solution we detect a failure of precompilation after a commit. This can give us some early useful information, that would be otherwise very hard to obtain if the issue happens when you deploy the software in the cluster.

[fverdugo]

and BTW the precompilation scrips should be in the "test app" (i.e. the final app) as usual

[amartinhuertas]

What do you think?

Ok. Let us go for 2. Can we use the sysimage of the test app to accelerate the mpi tests or would you avoid that?

[amartinhuertas]

@fverdugo ... all tests are now passing. PR ready to review/merge.

The CI script generates the TestApp.so sysimage, and then leverages it to run the MPI tests with 1x and 4x MPI tasks. The generation of the image takes 21 min according to github CI. On the other hand, the parallel MPI tests runtime is missleading in Gitlub CI: https://github.com/gridap/GridapDistributed.jl/runs/3970159570?check_suite_focus=true I think it only measures the time spent in the runtests.jl script, not the time spent in the mpirun processes forked out of it.