Engine barrier

Following the results in the threadpool_task_plots branch, I've replaced the elaborate hand-crafted engine_barrier function by two pthread_barriers.

As a result, the runner threads should all start, and synchronize, faster.

Added 2 commits:

• 8783c2e1 - engine_launch is only ever called with all threads, remove the option to limit t…
• be037f3e - engine_launch is only ever called with all threads, remove the option to limit t…

Looking at some results (more in a non-public place), I am starting to wonder whether we should not launch all the threads when we have very little stuff to do.

@nnrw56 anything specific that needs testing with this branch ? Is there a test-case that you fear might break ?

All my tests run fine using 1 to n threads. Same with MPI.

is there any way to check that the tasks start faster on 16 cores?

otherwise, no expected side effects, so good to go!

Added 52 commits:

• be037f3e...49c6a402 - 47 commits from branch master
• b6392d82 - grossly simplify the engine_barrier function using pthread_barriers.
• a473067b - use the new barrier.
• 49837b88 - engine_launch is only ever called with all threads, remove the option to limit t…
• 091462a9 - engine_launch is only ever called with all threads, remove the option to limit t…
• 2ec5c08f - Merge branch 'engine_barrier' of gitlab.cosma.dur.ac.uk:swift/swiftsim into engine_barrier

So, running the master and this branch on the device whose name shall not be mentioned I see no noticeable difference in the total run time on 16 cores.

OK, so thread startup/stop time is not the issue. Thanks for checking!

Status changed to merged

mentioned in commit ddad736c