A global, auto-scaling, preemptive scheduler using work-balancing.
What? Another executor?
smolscale is a work-balancing executor based on [async-task], designed to be a drop-in replacement to smol and async-global-executor. It is designed based on the thesis that work-stealing, the usual approach in async executors like async-executor and 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 --- aasync-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 likeasync-executorthat do not special-case message passing.
Furthermore, smolscale has a preemptive thread pool that ensures that tasks cannot block other tasks no matter what. This means that you can do things like run expensive computations or even do blocking I/O within a task without worrying about causing deadlocks. Even with "traditional" tasks that do not block, this approach can reduce worst-case latency. Preemption is heavily inspired by Stjepan Glavina's previous work on async-std.
smolscale also experimentally includes Nursery, a helper for structured concurrency on the smolscale global executor.
Show me the benchmarks!
Right now, smolscale uses a very naive implementation (for example, stealable local queues are implemented as SPSC queues with a spinlock on the consumer side, and worker parking is done naively through event-listener), and its performance is expected to drastically increase. However, at most tasks it is already much faster than async-global-executor (the de-facto standard "non-Tokio-world" executor, which powers async-std), sometimes an order of magnitude faster. Here are some unscientific benchmark results; percentages are compared to async-global-executor:
spawn_one time: [105.08 ns 105.21 ns 105.36 ns]
change: [-98.570% -98.549% -98.530%] (p = 0.00 < 0.05)
Performance has improved.
spawn_many time: [3.0585 ms 3.0598 ms 3.0613 ms]
change: [-87.576% -87.291% -86.948%] (p = 0.00 < 0.05)
Performance has improved.
yield_now time: [4.1676 ms 4.1917 ms 4.2166 ms]
change: [-50.455% -49.994% -49.412%] (p = 0.00 < 0.05)
Performance has improved.
//
ping_pong time: [8.5389 ms 8.6990 ms 8.8525 ms]
change: [+12.264% +14.548% +16.917%] (p = 0.00 < 0.05)
Performance has regressed.
Benchmarking spawn_executors_recursively:
spawn_executors_recursively
time: [180.26 ms 180.40 ms 180.56 ms]
change: [+497.14% +500.08% +502.97%] (p = 0.00 < 0.05)
Performance has regressed.
context_switch_quiet time: [100.67 us 102.05 us 103.07 us]
change: [-42.789% -41.170% -39.490%] (p = 0.00 < 0.05)
Performance has improved.
context_switch_busy time: [8.7637 ms 8.9012 ms 9.0561 ms]
change: [+3.3147% +5.5719% +7.6684%] (p = 0.00 < 0.05)
Performance has regressed.