We have a small fixed (and between 2 and 16) number of worker threads
(essentially a thread pool). See the --jobs arguments to bismon
program. Look into implementation details in agenda_BM.c source
file.
There is an (implicit) agenda grouping tasklet objects, of various priority.
The priorities can be very_high, high, normal, low, very_low.
The agenda maintains a few (changing) hash sets of tasklet objects, one set per priority.
Each worker thread loops by taking (and removing) a tasklet from the
agenda, and runs that tasklet, and repeat. The running step of a
tasklet is expected to be short (a few milliseconds) and should be
bounded (see also taskletcputime_BM giving its CPU time, and
taskletelapsed_BM giving its elapsed time) and without significant
I/O (so some "CPU-intensive" computation). A tasklet is run by sending
it the run_tasklet selector (without additional arguments), with the
tasklet being locked. That run would usually (but not always) add some
tasklets into the agenda.
So the agenda API is with functions like
agenda_add_very_high_priority_tasklet_BM
.... agenda_add_normal_priority_tasklet_BM
... agenda_add_very_low_priority_tasklet_BM adding tasklets, and
agenda_remove_tasklet_BM to remove a tasklet.
The worker threads are running tasklets of very_high then high
... up to very_low priority first. Occasionally a lower priority
task could be run (once every 32 times, randomly). The order in which
tasklets of same priority are running is unspecified (and essentially
random).