|0.4.2||Nov 24, 2023|
|0.4.0||Oct 16, 2023|
|0.3.52||Dec 5, 2022|
|0.3.49||Nov 18, 2022|
|0.1.0||Oct 26, 2020|
#422 in Asynchronous
4,417 downloads per month
Used in 27 crates (21 directly)
A global, auto-scaling scheduler for [async-task] using work-balancing.
What? Another executor?
smolscale is a work-balancing executor based on [async-task], designed to be a drop-in replacement to
async-global-executor. It is designed based on the idea that work-stealing, the usual approach in async executors like
tokio, is not the right algorithm for scheduling huge amounts of tiny, interdependent work units, which are what message-passing futures end up being. Instead,
smolscale uses work-balancing, an approach also found in Erlang, where a global "balancer" thread periodically balances work between workers, but workers do not attempt to steal tasks from each other. This avoids the extremely frequent stealing attempts that work-stealing schedulers generate when applied to async tasks.
smolscale's approach especially excels in two circumstances:
- When the CPU cores are not fully loaded: Traditional work stealing optimizes for the case where most workers have work to do, which is only the case in fully-loaded scenarios. When workers often wake up and go back to sleep, however, a lot of CPU time is wasted stealing work.
smolscalewill instead drastically reduce CPU usage in these circumstances --- a
async-executorapp that takes 80% of CPU time may now take only 20%. Although this does not improve fully-loaded throughput, it significantly reduces power consumption and does increase throughput in circumstances where multiple thread pools compete for CPU time.
- When a lot of message-passing is happening: Message-passing workloads often involve tasks quickly waking up and going back to sleep. In a work-stealing scheduler, this again floods the scheduler with stealing requests.
smolscalecan significantly improve throughput, especially compared to executors like
async-executorthat do not special-case message passing.