So, I have some experience with Kotlin coroutines, but let’s call my level, hm, not too deep. This article inspired me for doing some experiment:

Dispatchers.IO
This dispatcher is like Default but has a bigger parallelism limit. It uses the same thread pool under the hood but can dedicate more threads for coroutines execution. IO dispatcher, as mean its name, is useful for blocking input/output operations, but do not use it for CPU-heavy operations: large parallelism may lead to parallel work of a lot of threads, and switching between them also kills your performance.

  First, what are dispatchers which are mentioned above ? Here is link to Kotlin documentation. and CoroutineDispatcher abstract class
In short, dispatcher

determines what thread or threads the corresponding coroutine uses for its execution

Idea
  The question is: what will happen if I violate recommendation from above and use Dispatchers.IO for CPU-intensive calculations ?
  I’have created benchmark, which calculates factorial and I’m running it using var nCoroutines = 4000 3 times: with plain threads, with Dispatchers.Default, with Dispatchers.IO

Results

main summary:
Benchmark                                             (factorialSize)  Mode  Cnt        Score        Error  Units
DispatchersCPUBenchMark.coroutinesDispatchersDefault             2000  avgt    3     5255.348 ±  75191.476  us/op
DispatchersCPUBenchMark.coroutinesDispatchersDefault             4000  avgt    3     3813.471 ±  33133.584  us/op
DispatchersCPUBenchMark.coroutinesDispatchersIO                  2000  avgt    3     6437.317 ±  12881.315  us/op
DispatchersCPUBenchMark.coroutinesDispatchersIO                  4000  avgt    3     9983.775 ±  27035.014  us/op
DispatchersCPUBenchMark.threads                                  2000  avgt    3  1732705.383 ± 355342.356  us/op
DispatchersCPUBenchMark.threads                                  4000  avgt    3  6034969.050 ± 664852.667  us/op
  plain threads coroutinesDispatchersDefault coroutinesDispatchersIO
CPU, % 100 100 100
Memory, % 40 64 64
Runtime, millis 1641755.324 - 6028442.417 5617.26 - 7342.059 13205.202 - 5070.477

  Benchmark summary looks quite logical: threads show maximum running time and coroutinesDispatchersIO performs slower than coroutinesDispatchersDefault

  BUT…

  • performance monitor reported that plain threads take the least amount of memory while it’s declared that coroutines are light-weight.
  • when I increase load(from factorialSize = 2000 to factorialSize = 4000) running time for coroutinesDispatchersDefault goes down from 5355.348 millis to 3813.471 millis:
  • running times of different iterations for coroutinesDispatchersDefault vary a lot: 
    main: coroutines.DispatchersCPUBenchMark.coroutinesDispatchersDefault | factorialSize=2000
Warm-up 1: 2170.449 us/op
Iteration 1: 4014.905 us/op
Iteration 2: 1896.534 us/op
Iteration 3: 9854.606 us/op

      You can notice that Iteration 2 ended faster than Iteration 1 but Iteration 3 took even longer time. The same picture(different running time for particular iteration) is for coroutinesDispatchersIO but deviation is not so huge(see results in below).

    main: coroutines.DispatchersCPUBenchMark.coroutinesDispatchersIO | factorialSize=2000
Warm-up 1: 5923.528 us/op
Iteration 1: 5426.481 us/op
Iteration 2: 28599.197 us/op
Iteration 3: 5589.929 us/op

Conclusion
  While Kotlin documentation looks correct, benchmark highlighted few things, which look like an anomaly and deserve additional testing.
I’ll follow up them later…