For full text search please use the '?' prefix. e.g. ? Onboarding

Dispatchers.Default

Dispatchers.Default

  • CPU-intensive work (calculations, data processing)
  • Thread pool size equals CPU cores

Questions that come up

Should we limit max concurrency by one less thread for heavy workloads to leave some for UI?

From Q: Should you limit the concurrency of Dispatchers.Default? (A: No, for Linux multi core CPUs)
Go to text ā†’

Should we limit the concurrency of Dispatchers.Default by creating a limitedParallelism dispatcher?

// Like this:
val maxConcurrency = (Runtime.getRuntime().availableProcessors() - 1).coerceAtLeast(1)
val limitedDispatcher = dispatcher.limitedParallelism(maxConcurrency)

Why would we want to limitedParallelism? Is due to the following question: if you use up all threads will operating system (OS) create cooperative environment for other processes that are running?.

Experiment aimed to roughly answer this

Experiment: load up CPU and try to use other apps in the meantime.

When I tried this on 16-core/32-thread AMD Ryzen 9 7945HX on Linux it seems at least Linux OS was creating a cooperative environment between the threads to not show perceivable slow down in other processes while the CPU was shown to be running at 100% per htop. Hence, as of now I don't think it's worth artificially lowering the parallelism and rather let OS take care of scheduling threads.

img

From code-example loading up CPU for 30 seconds
Go to text ā†’
GT-Sandbox-Snapshot: example loading up all the CPU cores for 30 seconds doing computation in a loop

Code

package com.glassthought.sandbox

import kotlinx.coroutines.*
import kotlin.system.measureTimeMillis

fun main(): Unit = runBlocking(CoroutineName("main-runBlocking")) {
  val availableProcessors = Runtime.getRuntime().availableProcessors()

  println("Starting CPU stress test on $availableProcessors cores...")
  println("This will run for 30 seconds. System may become less responsive.")

  val duration = 30_000L // 30 seconds in milliseconds

  val elapsedTime = measureTimeMillis {
    // Create a list of jobs, one for each processor
    val jobs = List(availableProcessors) { coreIndex ->
      // Use Dispatchers.Default which has a thread pool sized to CPU cores
      // Each coroutine will pin to a thread and consume CPU
      launch(Dispatchers.Default + CoroutineName("cpu-loader-$coreIndex")) {
        println("Core $coreIndex: Starting intensive computation...")

        val startTime = System.currentTimeMillis()
        var counter = 0L
        var sum = 0.0

        // Tight loop that runs for 30 seconds
        while (System.currentTimeMillis() - startTime < duration) {
          // Perform some CPU-intensive operations
          counter++
          sum += Math.sqrt(counter.toDouble())

          // Add some more operations to ensure CPU usage
          if (counter % 1000000 == 0L) {
            // Occasionally perform more complex calculations
            for (i in 1..100) {
              sum += Math.sin(sum) * Math.cos(counter.toDouble())
            }
          }

          // Check for cancellation periodically (every million iterations)
          if (counter % 10000000 == 0L) {
            yield() // Allow coroutine cancellation if needed
          }
        }

        println("Core $coreIndex: Completed. Counter=$counter, Sum=$sum")
      }
    }

    // Wait for all jobs to complete
    jobs.joinAll()
  }

  println("\nCPU stress test completed in ${elapsedTime}ms")
  println("All $availableProcessors cores were loaded for approximately 30 seconds")
}

Command to reproduce:

gt.sandbox.checkout.commit 552c189a3d508ae510c9 \
&& cd "${GT_SANDBOX_REPO}" \
&& cmd.run.announce "./gradlew run --quiet"

Recorded output of command:

Picked up JAVA_TOOL_OPTIONS: -Dkotlinx.coroutines.debug
Picked up JAVA_TOOL_OPTIONS: -Dkotlinx.coroutines.debug
Starting CPU stress test on 32 cores...
This will run for 30 seconds. System may become less responsive.
Core 0: Starting intensive computation...
Core 1: Starting intensive computation...
Core 2: Starting intensive computation...
Core 3: Starting intensive computation...
Core 4: Starting intensive computation...
Core 5: Starting intensive computation...
Core 6: Starting intensive computation...
Core 7: Starting intensive computation...
Core 8: Starting intensive computation...
Core 9: Starting intensive computation...
Core 10: Starting intensive computation...
Core 11: Starting intensive computation...
Core 12: Starting intensive computation...
Core 13: Starting intensive computation...
Core 14: Starting intensive computation...
Core 15: Starting intensive computation...
Core 16: Starting intensive computation...
Core 17: Starting intensive computation...
Core 18: Starting intensive computation...
Core 19: Starting intensive computation...
Core 20: Starting intensive computation...
Core 21: Starting intensive computation...
Core 22: Starting intensive computation...
Core 23: Starting intensive computation...
Core 24: Starting intensive computation...
Core 25: Starting intensive computation...
Core 26: Starting intensive computation...
Core 27: Starting intensive computation...
Core 29: Starting intensive computation...
Core 28: Starting intensive computation...
Core 30: Starting intensive computation...
Core 31: Starting intensive computation...
Core 3: Completed. Counter=1209229201, Sum=2.8033135773977207E13
Core 11: Completed. Counter=1211769364, Sum=2.8121513705789645E13
Core 8: Completed. Counter=1199590272, Sum=2.7698620755711055E13
Core 2: Completed. Counter=1203021176, Sum=2.7817535389470645E13
Core 13: Completed. Counter=1190825648, Sum=2.7395612274237375E13
Core 31: Completed. Counter=1218795961, Sum=2.83664671475123E13
Core 21: Completed. Counter=1212393233, Sum=2.814323367875016E13
Core 19: Completed. Counter=1204196135, Sum=2.785829831628206E13
Core 12: Completed. Counter=1223621706, Sum=2.8535106632545105E13
Core 20: Completed. Counter=1223951172, Sum=2.854663222930776E13
Core 26: Completed. Counter=1209303550, Sum=2.803572122383962E13
Core 29: Completed. Counter=1225677066, Sum=2.8607033934887254E13
Core 0: Completed. Counter=1222882977, Sum=2.8509269567738004E13
Core 23: Completed. Counter=1209603905, Sum=2.8046166730119613E13
Core 24: Completed. Counter=1189998737, Sum=2.7367081901189492E13
Core 9: Completed. Counter=1215672477, Sum=2.8257492284361145E13
Core 25: Completed. Counter=996223257, Sum=2.0962532819551504E13
Core 30: Completed. Counter=1199697911, Sum=2.770234892705607E13
Core 5: Completed. Counter=1214359549, Sum=2.821172744882745E13
Core 28: Completed. Counter=1206475954, Sum=2.7937448955144168E13
Core 18: Completed. Counter=1213988582, Sum=2.8198801103505355E13
Core 4: Completed. Counter=1208161491, Sum=2.7996015452241723E13
Core 7: Completed. Counter=1173755242, Sum=2.6808657016430223E13
Core 22: Completed. Counter=1193089284, Sum=2.7473763708917047E13
Core 16: Completed. Counter=1216669588, Sum=2.8292265177977305E13
Core 6: Completed. Counter=1214497392, Sum=2.8216531091415805E13
Core 14: Completed. Counter=1202774004, Sum=2.7808962768959176E13
Core 27: Completed. Counter=1213167406, Sum=2.817019424997135E13
Core 17: Completed. Counter=1204046012, Sum=2.785308898096554E13
Core 1: Completed. Counter=1208010559, Sum=2.7990769428434086E13
Core 15: Completed. Counter=1215303167, Sum=2.8244616717065184E13
Core 10: Completed. Counter=1215761769, Sum=2.8260605640605434E13

CPU stress test completed in 30043ms
All 32 cores were loaded for approximately 30 seconds

Note: Would be interesting to try this out on Mac & Windows later.

Children
  1. Q: Should you limit the concurrency of Dispatchers.Default? (A: No, for Linux multi core CPUs)
Backlinks