Restart DLA-Future and the pika runtime less frequently in C API tests

[msimberg]

No description provided.

[msimberg]

cscs-ci run

[msimberg]

cscs-ci run

[msimberg]

cscs-ci run

[msimberg]

cscs-ci run

[msimberg]

This does help quite a bit. In the last commit (26abeba) I only start and stop the runtime once per google-test. In the previous commit (70a69d9) I skip starting and stopping for the innermost loop but keep the others (it's still a good stress test, though not very realistic).

My proposal would be to drop the last commit so we start and stop a bit less often, but still fairly often (this already gives quite a nice speedup).

[msimberg]

Already requesting reviews, or at least comments, on this. I would still need to update the other C API tests, I've only applied the change to the gen_eigensolver one so far. Waiting for comments before applying changes elsewhere.

[msimberg]

I've rebased this on master instead of #1225, since it's not a requirement.

[RMeli]

Starting/stopping once per test sounds good to me (26abeba). I think at some point we already used this approach to avoid a memory leak in pika (see #886 (comment)), but I don't recall why we decided to to start and stop the runtime at every test once the leak was fixed.

[msimberg]

cscs-ci run

[msimberg]

Starting/stopping once per test sounds good to me (26abeba).

I've updated both eigensolver tests to restart DLAF/pika once per grid (we need the grid for init).

I think at some point we already used this approach to avoid a memory leak in pika (see #886 (comment)), but I don't recall why we decided to to start and stop the runtime at every test once the leak was fixed.

I don't remember for sure either, but I think it was to have a regression/unit test for the kind of issues that that mode exposed (memory leaks and other issues from restarting everything a lot of times), but I think we didn't consider the test time implications of it.