I’m seeing some pretty disappointing CPU usage fro...

# coroutines

[kevinherron]

I’m seeing some pretty disappointing CPU usage from a low throughput coroutines workload (sending ~40-50 network requests per second) that is all non-blocking I/O, barely any CPU work. Using the async profiler shows the DefaultDispatcher threads are spending the majority of their time in the coroutines scheduler (mostly in Worker.doPark or WorkQueue.trySteal. Is this… expected?

[Luis Munoz]

I don't understand your question, you say you are disappointment in CPU usage and you say you you have tasks that do barely any CPU work. Seems like a contradiction to me. If the tasks are IO bound then I would expect all the threads to be parked doing nothing.

[kevinherron]

My point is that the coroutine scheduler itself is using much more CPU than I would expect

[marstran]

When your code is mainly waiting for IO, then the percentage of time spent in the scheduler will go up because nothing else are running.

[kevinherron]

@elizarov @Vsevolod Tolstopyatov [JB] is it possible this branch will improve things? https://github.com/Kotlin/kotlinx.coroutines/tree/scheduler-changes

[Vsevolod Tolstopyatov [JB]]

yes, that's its point :)

[kevinherron]

👍

[kevinherron]

i’d test instead of ask a stupid question but I can’t seem to build it right now because I don’t have Xcode installed

[kevinherron]

this jprofiler call tree seems crazy to me. 12 million invocations of findTaskWithCpuPermit to only 20,000 actual calls into suspended code (DispatchedTask.run)

[Vsevolod Tolstopyatov [JB]]

(sending from my phone, so can't give more detailed instructions) you can comment out all native targets from https://github.com/Kotlin/kotlinx.coroutines/blob/master/gradle/compile-native-multiplatform.gradle

[kevinherron]

(image not uploading)

[kevinherron]

thanks for the link

[kevinherron]

message has been deleted

[kevinherron]

@Vsevolod Tolstopyatov [JB] that branch makes a significant difference in CPU usages. Nice work 👍