8 releases
0.3.1 | May 7, 2019 |
---|---|
0.3.0 | May 7, 2019 |
0.2.2 | Feb 27, 2019 |
0.2.1 | Jan 16, 2019 |
0.1.2 | Jan 12, 2019 |
#16 in #task-scheduler
30KB
566 lines
Contains (Zip file, 17KB) sketch.odg
mpmc-scheduler
A Fair, Per-Channel Cancellable, multi-mpmc task scheduler running on top of tokio.
It bundles together multiple mpmc channels and schedules incoming work with fair rate limiting among the allowed maximum of workers.
Example
use mpmc_scheduler;
use tokio::runtime::Runtime;
use futures::future::Future;
let (controller, scheduler) = mpmc_scheduler::Scheduler::new(
4,
|v| {
println!("Processing {}", v);
v
},
Some(|r| println!("Finalizing {}", r)),
true
);
let mut runtime = Runtime::new().unwrap();
let tx = controller.channel(1,4);
runtime.spawn(scheduler);
for i in 0..4 {
tx.try_send(i);
}
drop(tx); // drop tx so scheduler & runtime shut down
runtime.shutdown_on_idle().wait().unwrap();
Details
You can think of it as a round-robin scheduler for rate limited workers which always run the same function.
o- -x
\ /
o--|--Scheduler --|--x
/ \
o- -x
In this image we have an n amount of Producers o
and m amount of Workers x
We want to handle all incoming work from o
in a fair manner. Such that if
one producers has 20 jobs and another 2, both are going to get handled equally in a round robin fashion.
Each channel queue can be cleared such that all to-be-scheduled jobs are droppped.
To allow also stopping currently running (expensive) operations, these can be split into two sections (functions).
The worker_fn
which can't be canceled and worker_fn_finalize
which is not called if a job is marked as canceled.
For example http requests whose result is stored into a database. If we abort before the store operation we can prevent all outstanding
worker operations of one channel plus the remaining jobs. We create fetch-http as the blocking and the db storing as the optional part.
Closed channels are detected and removed from the scheduler when iterating.
You can manually trigger a schedule tick by calling gc
on the controller.
Performance
If you have idle workers it takes ~ 1ms or less to process a job. Depending on your worker/producer ratio your mileage may vary. For example with Arcane Magic benchmarks it results in 56ms/job on a i7-6700HQ with 1 million jobs, 8 parallel producing channels & 8 Workers, 1024 bound per channel. Note that at most two roundtrips per schedule interval are done (so at most 16 jobs scheduled per interval) and we constantly have to re-send. This means that above numbers include iteration & polling start-stop fees.
Limitations
-
mpmc-scheduler can only be used with its own Producer channels due to missing traits for other channels. futures mpsc also doesn't work as they are not waking up the scheduler.
-
The channel bound has to be a power of two.
-
You can only define one work-handler function per
Scheduler
and it cannot be changed afterwards. You can work around this by passing along aBox<dyn Fn>
containing your dynamic function to be dispatched. See dynamic_dispatch example.
Thanks
Thanks to udoprog for the help with reducing the amount of generics & making it possible to store Controller & Scheduler in a non-generic way.
License
This project is licensed under the MIT license.
Contribution
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in mpmc-scheduler by you, shall be licensed as MIT, without any additional terms or conditions.
Dependencies
~0.6–6MB
~14K SLoC