Tokio seems oddly good compared to the others. Is that because it’s using a thread pool instead of creating something new for every task? I wonder how it would fare if the tasks were more intense, and how the time-to-completion compares between these approaches.
async Rust functions are compiled down to a finite state machine. A stackless coroutine I think it can be called. I suspect that C#/.NET does something similar though I am just speculating here. The difference between async-std and tokio would then be explained by a lower overhead runtime per task in tokio.
All the async and green thread runtimes are using thread pools by definition. Avoiding spawning as many OS threads as program's tasks is their main job.
This benchmark wasn't very rigorous, so maybe it has overlooked something, but the result seems plausible. Tokio is a good, mature runtime. Tokio-based frameworks are near top of TechEmpower benchmarks.
Rust's Futures were designed from the start with efficiency in mind. Unlike most async runtimes where chaining of multiple async calls creates new heap allocations, Rust flattens the entire call tree and all await points into a single struct representing a state machine. This allows the runtime to use exactly one heap allocation per task, even if the task is pretty complex.
