mirror of
https://github.com/status-im/nim-taskpools.git
synced 2026-08-27 17:31:08 +00:00
Esteban C Borsani
6897c03c83
Fixes #6 Adds an `injectQueue` which any thread can use to add tasks. After consuming local tasks, before trying to steal, tasks are drained from `injectQueue` into the local queue and processed; ~LIFO order is maintained. One flaw of this approach is if the tasks keep spawning tasks faster than they complete and so all workers are permanently busy, the tasks in inject-queue won't run. They only run if at some point one of the workers has no pending tasks. The injection-queue is drained every 61 processed local tasks to avoid this. On my machine current benchmarks show no difference. Added 2 benchs. The `taskpool_spc_external` can be used to compare VS the `taskpool_spc` queueing to local queue. The `iqs_latency/taskpool_iqs_latency.nim` shows latency of injected tasks when the local queues are permanently filled. there is a race cond between a thread adding a task and worker checking the injection queue + parking (if worker see no tasks in the queue, thread adds it, worker parks). But it also seems to exist for local queue + parking. It requires #54 to fix it properly (the sleepy / sleep ticket feature); also requires wakeAll to avoid the notification per worker.