You can schedule coroutines over a threadpool. From the author's code, I'm not even sure this was accounted for in the benchmark. If they are using a threadpool and vanilla coroutines, without specifying where they are scheduled, yes, of course it won't start up a pool of threads for you and be as fast as a pool of threads. It would be terrible if it did that by default.
The comparison is nonsensical. It's not an either or kind of thing. If your co-routines are doing any blocking IO, you'd use the built in IO dispatcher, which is backed by a thread pool or you'd create your own dispatcher. Each co-routine gets its own thread if you use that; which can then do blocking IO and block the thread. Very common on the jvm if you are interacting with blocking IO things. I was working on some code that needed this yesterday.
The examples in this "benchmark" don't make any sense. It's not clear what the work is or why threads/co-routines are even needed. Handling UI events with threads is not a thing. But co-routines are very useful for that. Because they have very low overhead.
The rule of thumb should be:
- Is your task CPU bound? Put it in a thread pool.
- Is your task spending most of its time waiting (e.g. network call)? Use a coroutine