Modern concurrency on the JVM: Coroutines and Loom
Yassine Benabbas- 13 Min To Read
- 11 Aug, 2026
Concurrent programming allows running multiple tasks simultaneously. Historically, it was mostly achieved through threads, which consist of a set of instructions that can be executed independently. However, they are harder to work with and have some limitations. Modern concurrent programming introduces concepts that make concurrent code easier to write and more efficient. Two of those concepts are structured concurrency and lightweight threads, which are implemented in the JVM ecosystem through Kotlin coroutines and Project Loom. Let’s explore these two approaches.
Introduction
Traditional (thread-based) concurrency has two notable issues: callback-style asynchronous flows can become hard to read at scale (also called callback hell), and extensive thread creation can consume significant system resources. The following example illustrates the extensive use of threads.
void main(String... args) {
Set<Long> uniqueThreadNames = ConcurrentHashMap.newKeySet();
for (int i = 0; i < 1_000; i++) {
new Thread(() -> {
try {
// Simulate IO request (database, HTTP call, ...)
Thread.sleep(1000);
uniqueThreadNames.add(Thread.currentThread().threadId());
IO.println(uniqueThreadNames.size());
} catch (Exception e) {
e.printStackTrace();
}
}).start();
}
}
The above code creates 1000 threads that each sleep for 1 second and print the current number of unique thread IDs.
In the example, we use a ConcurrentHashMap to store the unique thread IDs while remaining thread-safe.
The output is non-deterministic because threads run concurrently, but it should show growth toward 1000.
1
...
...
999
1000
This means up to 1000 unique threads can be created, which is expected since we create 1000 threads in the code.
By analyzing both the code and the output, we can note two above-mentioned problems.
The first one is that coordinating many thread completions often pushes code toward callback-based orchestration (or manual synchronization), which can become hard to read, predict and maintain.
The second problem is that system resources are not optimized: creating 1000 threads can be resource intensive, especially if the tasks are I/O-bound and spend much of their time waiting.
This can lead to high memory consumption and context switching overhead.
In addition to that, there is a limit to the number of threads that can be created by the OS.
We can verify this by increasing the number of threads to 1 million, which will throw an OutOfMemoryError or Too many threads error depending on the OS.
void main(String... args) {
for (int i = 0; i < 1_000_000; i++) {
new Thread(() -> {
try {
Thread.sleep(1000);
} catch (Exception e) {
e.printStackTrace();
}
}).start();
}
}
Running the above code will throw an error similar to the following:
[0.536s][warning][os,thread] Failed to start thread "Unknown thread" - pthread_create failed (EAGAIN) for attributes: stacksize: 2048k, guardsize: 16k, detached.
[0.536s][warning][os,thread] Failed to start the native thread for java.lang.Thread "Thread-4068"
Exception in thread "main" java.lang.OutOfMemoryError: unable to create native thread: possibly out of memory or process/resource limits reached
at java.base/java.lang.Thread.start0(Native Method)
at java.base/java.lang.Thread.start(Thread.java:1417)
at MillionThreads.main(MillionThreads.java:9)
Modern concurrency concepts solve these issues as we’ll see in the next sections. Let’s start by defining some concepts before delving into concrete implementations.
Modern concurrency concepts
There are two main concepts in modern concurrency that we will explore in this post: lightweight threads and structured concurrency.
Lightweight threads
They are threads that are managed by the runtime (like the JVM or Kotlin runtime) instead of the operating system. They still run on top of OS threads, also called platform threads or carrier threads (because they carry the lightweight threads). However, lightweight threads can reuse platform threads, which allows bypassing the limit of OS threads. In addition to that, the usage of system resources is optimized, since platform threads are more expensive to create and maintain than lightweight ones.
Lightweight threads are very efficient for tasks that spend most of their time waiting, such as I/O-bound (network, file I/O, etc.). This means that we can spawn a large number of lightweight threads that download files or communicate with printers without worrying about system resources, which is not the case with traditional threads. It is important to note that compute-intensive tasks are still bound to the raw CPU and GPU cores and power. Thus, running a large number of compute-intensive tasks efficiently in parallel is not possible on low-end hardware, even with lightweight threads.
The JVM has two implementations of lightweight threads: coroutines (introduced by Kotlin) and virtual threads (introduced by Project Loom).
Structured concurrency
Structured concurrency is a programming paradigm that aims to make concurrent code similar to a sequential one.
It is achieved by providing APIs that replace traditional callback-based code with constructs that enforce a sequential structure for concurrent tasks.
We can also see related structured-concurrency patterns in JavaScript, C# and Swift with the async/await model.
In the JVM ecosystem, structured concurrency is implemented by Kotlin coroutines and Project Loom.
Kotlin coroutines
A coroutine is a lightweight thread that is managed by the Kotlin runtime. Coroutines were initially proposed in Kotlin version 1.1 M01 , released in 2017. The first stable version was released with Kotlin 1.3 one year later.
Two concepts are essential to understand coroutines: suspending functions and CoroutineScope.
Coroutines run inside a CoroutineScope, which is a context that defines the lifecycle of the coroutines.
A suspending function is a function that is marked with the suspend keyword.
Any function that calls suspending functions must be marked as suspend as well (similar to the async keyword in other languages).
Regarding the lightweight aspect of coroutines, we can illustrate this by creating thousands of coroutines without worrying about system resources. Let’s illustrate this with a program that creates 1 million coroutines that each sleep for 1 second and then prints the number of unique threads used to execute them.
package org.example
import kotlinx.coroutines.*
import kotlin.time.Duration.Companion.seconds
import java.util.concurrent.ConcurrentHashMap
suspend fun main(){
// Thread safe set
val uniqueThreadNames = ConcurrentHashMap.newKeySet<Long>()
coroutineScope {
for (i in 1..1_000_000) {
launch {
delay(1.seconds)
uniqueThreadNames.add(Thread.currentThread().threadId())
}
}
}
println("Unique threads used: ${uniqueThreadNames.size}")
}
The output depends on runtime scheduling and machine characteristics (such as available cores and dispatcher parallelism). It should be something like this on an M1 Mac with 8 cores:
Unique threads used: 8
This means that the coroutines are efficiently scheduled on the available hardware cores, without the overhead of creating a large number of OS threads. In fact, we can even increase the number of coroutines to more than 1 million without any issue, which is not possible with traditional (OS) threads.
The second aspect of coroutines is structured concurrency, which allows writing concurrent code in a sequential style. Let’s see an example of how to create a coroutine scope that launches two coroutines.
package org.example
import kotlinx.coroutines.*
import kotlin.time.Duration.Companion.seconds
suspend fun main(){
coroutineScope {
// First coroutine
launch {
println("Start of coroutine 1")
delay(1.seconds)
println("End of coroutine 1")
}
// Second coroutine
launch { println("I am another coroutine") }
}
println("Coroutine scope completed")
}
The coroutine scope is created with the coroutineScope suspending function (defined with the suspend qualifier).
Since it is a suspending function, then the main function that calls it must be marked as suspend as well.
That’s why the main function is defined with suspend fun main().
The coroutine scope launches two coroutines with the launch function (launch creates a coroutine and runs it).
The first one prints a message, waits for 1 second and then prints another message.
The second one simply prints a message.
Can you guess the output of this code? Here is the answer:
Start of coroutine 1
I am another coroutine
End of coroutine 1
Coroutine scope completed
Since the first coroutine waits for 1 second, the second coroutine is executed while the first one is suspended.
What if we want to start the second one only after the first one completes?
That can be achieved with the join function, which waits for the completion of a coroutine.
package org.example
import kotlinx.coroutines.*
import kotlin.time.Duration.Companion.seconds
suspend fun main(){
coroutineScope {
// Keep a reference to the first coroutine
val coroutine = launch {
println("Start of coroutine 1")
delay(1.seconds)
println("End of coroutine 1")
}
// Wait for the first coroutine to complete
coroutine.join()
// Second coroutine
launch { println("I am another coroutine") }
}
println("Coroutine scope completed")
}
The output of this code is:
Start of coroutine 1
End of coroutine 1
I am another coroutine
Coroutine scope completed
By getting a reference to the first coroutine with val coroutine = launch { ... }, we call coroutine.join() to wait for its completion before starting the second one.
The above two examples show the essence of structured concurrency: the code seems sequential, but it is actually concurrent code. In practice, this removes callback hell while keeping the code easy to follow.
Let’s now see how the JDK implements modern concurrency.
Java’s virtual threads and structured concurrency
In addition to Kotlin coroutines, the JDK natively achieves modern concurrency through two APIs: virtual threads and structured concurrency.
Virtual threads are the JDK implementation of lightweight threads, and are introduced by Project Loom.
They share a similar API with platform threads because both are represented by instances of the Thread class.
They can be created using the static method Thread.ofVirtual().
The following code snippet illustrates the creation of a platform thread and a virtual thread.
///usr/bin/env jbang "$0" "$@" ; exit $?
//JAVA 25+
//PREVIEW
void main(String... args) {
// Platform (or OS) thread
Thread.ofPlatform().start(() -> {
IO.println(Thread.currentThread());
});
// Virtual thread
Thread.ofVirtual().start(() -> {
IO.println(Thread.currentThread());
});
try {
Thread.sleep(1000);
} catch (InterruptedException e) {
}
}
The output of this code is similar to the following:
Thread[#25,Thread-0,5,main]
VirtualThread[#27]/runnable@ForkJoinPool-1-worker-1
In the above logs, we can confirm that the platform thread is spawned from the main thread, while the virtual thread runs on a worker thread shown as ForkJoinPool-*.
This naming reflects the current runtime implementation and can vary depending on JDK version and workload.
The worker thread is the platform thread that runs the virtual thread.
Thanks to their nature, and similarly to Kotlin coroutines, we can create a large number of virtual threads without worrying about system resources. The following code snippet creates 1 million virtual threads and prints additional information about them.
void main(String... args) {
int taskCount = 1_000_000;
Set<String> uniqueWorkers = ConcurrentHashMap.newKeySet();
var latch = new java.util.concurrent.CountDownLatch(taskCount);
for (int i = 0; i < taskCount; i++) {
Thread.ofVirtual().start(() -> {
try {
Thread.sleep(1000);
var threadInfo = Thread.currentThread().toString();
// threadInfo will be something like VirtualThread[#(id)]/runnable@ForkJoinPool-(id)-worker-(id)
var workerStartIndex = threadInfo.indexOf("ForkJoinPool");
if (workerStartIndex >= 0) {
var workerName = threadInfo.substring(workerStartIndex);
uniqueWorkers.add(workerName);
}
} catch (InterruptedException e) {
Thread.currentThread().interrupt();
} catch (Exception e) {
e.printStackTrace();
} finally {
latch.countDown();
}
});
}
try {
// Wait for all virtual threads to finish
latch.await();
} catch (InterruptedException e) {
Thread.currentThread().interrupt();
throw new RuntimeException(e);
}
IO.println("Unique worker threads observed: " + uniqueWorkers.size());
uniqueWorkers.stream().sorted().forEach(IO::println);
}
Let’s explain some parts.
The line var threadInfo = Thread.currentThread().toString();, generates a string that looks like this: VirtualThread[#(id)]/runnable@ForkJoinPool-(id)-worker-(id).
The next lines extract the ForkJoinPool id and the worker id, where the worker thread is the platform thread that runs the virtual thread.
Because we wait for all virtual threads to complete with a CountDownLatch, the final worker-thread count reflects a complete execution rather than a time-bounded snapshot.
The output of the above code is similar to the following:
Unique worker threads observed: 8
ForkJoinPool-1-worker-8
ForkJoinPool-1-worker-6
ForkJoinPool-1-worker-7
ForkJoinPool-1-worker-1
ForkJoinPool-1-worker-4
ForkJoinPool-1-worker-5
ForkJoinPool-1-worker-2
ForkJoinPool-1-worker-3
We can divide the log output into two parts.
The first part prints 1 million lines in this format: VirtualThread[#(id)]/runnable@ForkJoinPool-(id)-worker-(id).
The second part consists of eight lines with this format: ForkJoinPool-1-worker-(1 to 8).
In this run, all virtual threads appear to be scheduled on the same ForkJoinPool instance.
The second part, with only eight lines, shows that virtual threads were scheduled on 8 unique worker threads in this execution.
The exact number of workers may vary by runtime and workload, but the key point is that many virtual threads can be multiplexed over a much smaller number of platform threads.
The other aspect of Java’s modern concurrency is structured concurrency.
The class that provides this feature is StructuredTaskScope
.
It is available in Java 21 as a preview feature (it is still the case in Java 25).
It returns an object, usually created with a try-with-resources block, that we’ll call a scope.
That scope is used to launch concurrent tasks with the fork method, and to wait for their completion with the join method.
Chaining the join method with the fork method allows creating a sequential structure for concurrent tasks, which is the essence of structured concurrency.
The following code snippet illustrates the use of structured concurrency.
It launches two tasks concurrently and then launches a third one after the first two complete.
///usr/bin/env jbang "$0" "$@" ; exit $?
//JAVA 25+
//PREVIEW
void main(String... args) {
try (var scope = new StructuredTaskScope<>()) {
var task1 = scope.fork(() -> {
IO.println("Task 1");
//long running task such as a network call or a database query
Thread.sleep(1000);
return 1;
});
var task2 = scope.fork(() -> {
IO.println("Task 2");
Thread.sleep(1000);
return 2;
});
scope.join();
IO.println("Sum: " + (task1.get() + task2.get()));
var task3 = scope.fork(() -> {
IO.println("Task 3 runs after task 1 and task 2");
return 3;
});
scope.join();
IO.println("Sum: " + (task1.get() + task2.get() + task3.get()));
} catch (InterruptedException e) {
e.printStackTrace();
}
}
Running the above code will produce an output similar to the following:
Task 1
Task 2
Sum: 3
Task 3 runs after task 1 and task 2
Sum: 6
We can see that the first two tasks are launched concurrently, and their results are printed after their completion. Then, the third task is launched after the completion of the first two tasks, which shows the sequential structure of the concurrent code. This, again, is the essence of structured concurrency: concurrent code written in a sequential way.
Coroutines vs Java modern concurrency
While coroutines and Java’s modern concurrency are two implementations of the same concepts (lightweight threads and structured concurrency), they have some differences and synergies that are worth mentioning.
Coroutines have been stable since 2017, while Java’s modern concurrency is still experimental as a whole (virtual threads are stable, but structured concurrency is not). This means that coroutines are currently more mature and widely adopted. In addition to that, coroutines are multiplatform in Kotlin. In practice, available dispatchers and runtime behavior can differ by target (JVM, Android, JS, and Native).
Since virtual threads use the Thread API, they can be used in Java and Kotlin, while coroutines can only be used in Kotlin.
Also, virtual threads can be used by frameworks behind the scenes, since it is sometimes just a matter of changing the thread factory.
This means that developers might not be aware of them or need to change their code to use them, while coroutines require explicit usage of the API.
A synergy between the two APIs is possible when Kotlin runs on JVM 21+.
In fact, coroutines can be dispatched on Java virtual threads by backing a coroutine dispatcher with Executors.newVirtualThreadPerTaskExecutor().asCoroutineDispatcher().
This can be useful for blocking I/O operations, but it is best to benchmark your workload before replacing Dispatchers.IO.
Because Kotlin runs on the JVM, it can call Java’s structured concurrency API directly, which is useful when working with existing Java code or libraries.
Conclusion
Modern concurrency is a powerful tool that allows writing concurrent code that is efficient and easy to read.
We have seen two ways to run structured concurrent code that optimizes I/O performance and readability.
On the JDK/JVM side, we have virtual threads and structured concurrency with StructuredTaskScope.
In Kotlin, coroutines are a high-level API that provides lightweight threads and structured concurrency.
Both APIs are efficient and easy to use, and can be used in synergy when running Kotlin on the JVM.
In conclusion, which one should you choose? The short answer is to use the one that is available in your language and framework. So, if you are using Java, then you can use Java’s modern concurrency. And, if you are using Kotlin, then you can use Kotlin coroutines.