asterisk

History

Gian Diego Javes a35939885b taskpool: count the in-flight task when selecting an executor The taskpool selectors measure load with ast_taskprocessor_size(), which only reports queued tasks. A taskprocessor that is executing a long running task has an empty queue, so it looks identical to an idle one and the least full selector hands it more work while other executors sit free. Add ast_taskprocessor_is_executing() and include the in-flight task in the load the selectors compare. This is not a new semantic. taskprocessor_push() already counts the in-flight task as outstanding work when it notifies the listener: /* The currently executing task counts as still in queue */ was_empty = tps->executing ? 0 : previous_size == 0; The taskpool selectors are the one place that still measures only the queue. This change applies the same rule there. The accessor reads tps->executing without the object lock, the way the CLI already does in main/taskprocessor.c, because taking the lock on every push cost roughly six times the throughput: taskpool push efficiency went from 324438 to 52966 tasks per second and the serializer variant from 217467 to 38313. Without the lock both are back to baseline. A stale read only means one push is routed as if the taskprocessor had just changed state, which the selector already tolerates. Also add a unit test that holds one executor of a four executor pool inside a long task and checks that a task pushed to a different serializer runs on one of the three free executors. It fails on every run without this change. Fixes: #2074 UserNote: Taskpool executors that are running a long task are no longer considered idle when work is distributed, so tasks are routed to free executors instead of queueing behind a busy one.		2 months ago
..
asterisk	taskpool: count the in-flight task when selecting an executor	2 months ago
solaris-compat	…
asterisk.h	asterisk: Add macro for curl user agent.	5 years ago
jitterbuf.h	progdocs: Fix grouping for latest Doxygen.	5 years ago