jstclair · November 18, 2020 10:29
diff --git a/Program.cs b/Program.cs
 using System;
 using System.Diagnostics;
 using System.Threading.Tasks;
 using System.Threading.Channels;
 using System.Linq;
 using System.Linq.Expressions;
 using System.Threading;

 namespace ChannelsDemo
 {
    /// <summary>
    /// Simple "getting-started" demo of System.Threading.Channels
    /// </summary>
    /// <remarks>
    /// There is no error handling demonstrated here. Typically, you would surround `WaitTo(Read|Write)Async` with `try {} catch (ChannelClosedException) {}`
    /// but that would obscure some of the code here.
    /// </remarks>
    class Program
    {
        // NOTE: You can adjust these to see how various combinations work

        // NOTE: You might want to put some artificial delays in OutputGreeting (via `await Task.Delay(TimeSpan.FromSeconds(2))` 
        private const int MaxItemsToProduce = 5_000;

        private static readonly int MaxReaders = Environment.ProcessorCount; // or, hard-coded value, like 1, 2, or 10
        private static readonly int MaxBoundedSize = 2 * MaxReaders; // or, hard-coded value, like 1, 2, or 400

        // NOTE: in all these examples, the time is dominated by the slowness of the Console. Try disabling this and comparing elapsed times.
        private const bool WriteToConsole = true;
        // NOTE: in scenarios where you want to simulate CPU-bound work, you can disable writing to console and set this instead. Calls Thread.SpinWait(n)
        private const int CpuSpinTime = 0;

        // NOTE: when experimenting with different strategies, this can help you visualize where tasks are in the channel
        private const bool UseTracing = false;

        /* Additional Note:
         * Performance is constrained when running a Debug build and/or in VS. Try building a Release build and running in a separate console.
         *
         * Examples from my machine using MaxItemsToProduce = 5000 and the MultipleCompetingTransformsAndConsumers example
         *
         * [WriteToConsole = true]
         * Debug/F5:        1_183 items/sec
         * Debug/Cmd:       1_350 items/sec
         * Release/F5:      1_204 items/sec
         * Release/Cmd:     1_371 items/sec
         *
         * [WriteToConsole = false]
         * Debug/F5:       49_341 items/sec
         * Debug/Cmd:     204_788 items/sec
         * Release/F5:     44_856 items/sec
         * Release/Cmd:   204_611 items/sec
         *
         * NOTE: There is also overhead for such a small number of items. Increasing MaxItemsToProduce to 500_000:
         *
         * [WriteToConsole = false]
         * Release/F5:   845_265 items/sec (MultipleCompetingTransformsAndConsumers)
         * Release/F5: 1_638_367 items/sec (Simple)
         */

        static async Task Main(string[] args)
        {
            Console.WriteLine("Welcome to the Channels Demo app");

            // NOTE: 2 types of channels - unbounded and bounded. Bounded channels have a limit - writes will block until reads have drained the channel.
            // Unbounded channels will never block reads or writes, but assume you have capacity to store all items in memory.

            // Our first channel is just a list of ids that we want channel consumers to read, so we can easily hold hundreds of thousands in memory.
            // But for observing the flow of processing (i.e., with UseTracing = true), you might want to look at a Bounded channel
            var idChannel = Channel.CreateUnbounded<string>();
            //var idChannel = Channel.CreateBounded<string>(MaxBoundedSize);

            // Our second channel is doing work, and we want to constrain the amount of simultaneous work (both for memory and resource consumption).
            // But you can test various scenarios where you use an unbounded channel here as well.
            var greetChannel = Channel.CreateBounded<string>(MaxBoundedSize);
            //var greetChannel = Channel.CreateUnbounded<string>();
            
            var sw = Stopwatch.StartNew();

            // NOTE: because the examples close the channel, you can only run a single example at a time
            await Simple(idChannel, greetChannel);

            //await MultipleConsumers(idChannel, greetChannel);

            //await MultipleTransformsAndConsumers(idChannel, greetChannel);

            //var otherChannel = Channel.CreateBounded<string>(MaxBoundedSize);
            //await MultipleCompetingTransformsAndConsumers(idChannel, greetChannel, otherChannel);

            sw.Stop();

            Console.WriteLine($"{MaxItemsToProduce:N0} consumed in {sw.Elapsed.TotalSeconds:F1} seconds [{(MaxItemsToProduce/sw.Elapsed.TotalSeconds):N0} items/second]");
        }
        
        // Example task of producing data and writing it to a ChannelWriter
        private static async Task ProduceIds(ChannelWriter<string> writer, int max)
        {
            var ids = Enumerable.Range(1, max);

            foreach (var id in ids)
            {
                Trace($"{nameof(ProduceIds)}: writing id {id}");
                // NOTE: this works just fine...
                await writer.WriteAsync($"{id}");

                // NOTE: This is an optimization for avoiding the await when the channel isn't blocking, or has a large capacity
                // In the current case, since we have an unbounded writer, we would never need to await.
                //if (writer.TryWrite($"{id}") == false)
                //{
                //    await writer.WriteAsync($"{id}");
                //}

                // NOTE: you can signal an error by calling `Complete(Exception)` - this will be an orderly shutdown.
                //if (id == 1337)
                //{
                //    Trace("Closing the channel for 1337");
                //    var e = new Exception($"You should never get a 1337");
                //    writer.Complete(e);
                //    // In the simple `writer.WriteAsync` above, if you `Complete` the channel, you will attempt to write an additional item to it.
                //}
            }
            Trace($"{nameof(ProduceIds)}: completed writing");
            // when we are done, signal the channel that no more data is coming
            writer.Complete();
        }

        // Example task of reading data from one channel (ChannelReader<T>) and writing it to another channel (ChannelWriter<T>)
        private static async Task TransformIdAndPublishGreeting(ChannelReader<string> reader,
            ChannelWriter<string> writer, string greeting, bool autoClose = false)
        {
            while (await reader.WaitToReadAsync())
            {
                Trace($"{nameof(TransformIdAndPublishGreeting)}: entered WaitToReadAsync");
                // NOTE: this will also work just fine..
                while (reader.TryRead(out var id))
                {
                    await writer.WriteAsync(FormatGreeting(greeting, id));
                }

                // NOTE: but again, like above, we can avoid the unnecessary writing await as long as possible
                //while (reader.TryRead(out var id))
                //{
                //    var msg = FormatGreeting(greeting, id);
                //    if (writer.TryWrite(msg) == false)
                //    {
                //        Trace($"{nameof(TransformIdAndPublishGreeting)}: TryWrite: had to take slow path");
                //        await writer.WriteAsync(msg);
                //    }
                //}
                Trace($"{nameof(TransformIdAndPublishGreeting)}: exited WaitToReadAsync");
            }

            // NOTE: if we have a single instance of this task, we can close the writer when we're done;
            // but if we have multiple instances of this task, that would fail (because `Complete()` can only 
            // be called once on a channel.
            if (autoClose)
            {
                Trace($"{nameof(TransformIdAndPublishGreeting)}: completing writer");
                writer.Complete();
            }
        }

        private static string FormatGreeting(string greeting, string id) => $"{greeting} {id}";

        // Example task of reading from a channel
        private static async Task OutputGreeting(ChannelReader<string> reader, string optionalPrefix = "")
        {
            while (await reader.WaitToReadAsync())
            {
                Trace($"{nameof(OutputGreeting)}{optionalPrefix}: entered WaitToReadAsync");

                while (reader.TryRead(out var greeting))
                {
                    if (WriteToConsole)
                        Console.WriteLine($"{optionalPrefix}{greeting}");
                    // NOTE: this is here to heat up your CPUs
                    else if (CpuSpinTime >= 0)
                        Thread.SpinWait(CpuSpinTime);
                }

                Trace($"{nameof(OutputGreeting)}{optionalPrefix}: exited WaitToReadAsync");
            }
        }

        /* NOTE: Simple example
         * In this example, we are using a single task for each process - this is closest to a traditional pipe architecture, with the advantage that
         * we natively handle async actions.
         */
        private static async Task Simple(Channel<string> idChannel, Channel<string> greetChannel)
        {
            var produceTask = Task.Run(() => ProduceIds(idChannel.Writer, MaxItemsToProduce));

            var transformTask = Task.Run(() => TransformIdAndPublishGreeting(idChannel.Reader, greetChannel.Writer, "Hello", autoClose: true));

            var greetTask = Task.Run(() => OutputGreeting(greetChannel.Reader));

            // NOTE: we await each task - in addition, we want to await any reader channels. Only readers have a `Task Completion`; writers have a `.Complete()` method.
            await Task.WhenAll(idChannel.Reader.Completion, produceTask, transformTask, greetChannel.Reader.Completion, greetTask);
        }

        /* NOTE: Multiple Consumers example
         * In this example, we are using a single task for producing and transforming, but multiple greeting processes. Since the greeting only reads,
         * there's minimal changes required.
         */
        private static async Task MultipleConsumers(Channel<string> idChannel, Channel<string> greetChannel)
        {
            var produceTask = Task.Run(() => ProduceIds(idChannel.Writer, MaxItemsToProduce));

            var transformTask = Task.Run(() => TransformIdAndPublishGreeting(idChannel.Reader, greetChannel.Writer, "Hello", autoClose: true));

            var greetTasks = Enumerable.Range(0, MaxReaders)
                .Select(i => Task.Run(() => OutputGreeting(greetChannel.Reader, $"{i}")))
                .ToArray(); // NOTE: *Very* important that we materialize the running tasks, since `.Select` is lazy

            await Task.WhenAll(new[] { idChannel.Reader.Completion, produceTask, transformTask, greetChannel.Reader.Completion }.Concat(greetTasks));
        }

        /* NOTE: Multiple Transforms And Consumers example
         * In this example, we are using a single task for producing, but multiple transform and greeting processes. Since the transform task was
         * responsible for signalling that it was done writing in the single-task scenario, we now need to handle this explicitly.
         */
        private static async Task MultipleTransformsAndConsumers(Channel<string> idChannel, Channel<string> greetChannel)
        {
            var produceTask = Task.Run(() => ProduceIds(idChannel.Writer, MaxItemsToProduce));

            var transformTasks = Enumerable.Range(0, MaxReaders * 2)
                .Select(i => Task.Run(() => TransformIdAndPublishGreeting(idChannel.Reader, greetChannel.Writer, $"'{i}' Hello", autoClose: false)))
                .ToArray();

            var greetTasks = Enumerable.Range(0, MaxReaders * 2)
                .Select(i => Task.Run(() => OutputGreeting(greetChannel.Reader, $"{i}")))
                .ToArray();

            // NOTE: once the transform tasks have completed writing to their channel, we can signal the greet tasks that no more data is coming.
            await Task.WhenAll(new [] {idChannel.Reader.Completion, produceTask }.Concat(transformTasks));
            greetChannel.Writer.Complete();
            
            await Task.WhenAll(new[] { greetChannel.Reader.Completion }.Concat(greetTasks));
        }

        /* NOTE: Multiple Competing Transforms And Consumers example
         * In this example, we are using a single task for producing, but multiple transform and greeting processes. In addition, the transform task
         * is now writing to multiple channels. In real code, this might be producing a read model and updating a search index. Like the MultipleTransformsAndConsumers
         * example, we are responsible for explicitly signalling that the transform tasks are completed writing.
         */
        private static async Task MultipleCompetingTransformsAndConsumers(Channel<string> idChannel, Channel<string> greetChannel, Channel<string> otherChannel)
        {
            // NOTE: in this example, tweaking the running tasks is important for throughput. In my case, I have 12 processors.
            // If I leave these all set to `MaxReaders`, the best result is:
            //   205_000 items/sec
            // If I adjust them to MaxReaders/3 (so that consumers roughly equal CPUs), the best result is:
            //   218_702 items/sec
            // NOTE: this will also depend on the work each of your tasks is doing. If some are slow, you may wish to create more of them...
            // NOTE: the produceTask will also burn a CPU while it's working (for large # of items), so you may wish to calculate based on MaxReaders - 1

            var produceTask = Task.Run(() => ProduceIds(idChannel.Writer, MaxItemsToProduce));

            var readersPerTask = MaxReaders; // MaxReaders / 3;

            var transformTasks = Enumerable.Range(0, readersPerTask)
                .Select(i => Task.Run(() => TransformIdAndPublishGreeting(idChannel.Reader, greetChannel.Writer, otherChannel.Writer, $"'{i}' Hello")))
                .ToArray();

            // NOTE: We want to have multiple consumers, but we can use the same underlying function; we will use the optional prefix to help
            // distinguish between the two tasks in the console. Each set of greet tasks reads from it's own channel (so slowing down one will only
            // affect those tasks. If you want to see this, you can add a `if (optionalPrefix.StartsWith("OTHER") Thread.SpinWait(CpuSpinTime);` in 
            // the `OutputGreeting` method.
            var greetTasks = Enumerable.Range(0, readersPerTask)
                .Select(i => Task.Run(() => OutputGreeting(greetChannel.Reader, $"GREET [{i:00}] ")))
                .ToArray();

            var otherTasks = Enumerable.Range(0, readersPerTask)
                .Select(i => Task.Run(() => OutputGreeting(otherChannel.Reader, $"OTHER: [{i:00}] ")))
                .ToArray();

            // NOTE: like in MultipleTransformsAndConsumers, we signal consumers that there's no more data to come. However, in this case, we need to
            // signal to both channels.
            await Task.WhenAll(new[] { idChannel.Reader.Completion, produceTask }.Concat(transformTasks));
            greetChannel.Writer.Complete();
            otherChannel.Writer.Complete();

            await Task.WhenAll(new[] { greetChannel.Reader.Completion, otherChannel.Reader.Completion }.Concat(greetTasks).Concat(otherTasks));
        }

        // NOTE: Overload of TransformIdAndPublishGreeting that takes a multiple channel writers (for the MultipleCompetingTransformsAndConsumers example)
        private static async Task TransformIdAndPublishGreeting(ChannelReader<string> reader,
            ChannelWriter<string> writer, ChannelWriter<string> otherWriter, string greeting)
        {
            while (await reader.WaitToReadAsync())
            {
                while (reader.TryRead(out var id))
                {
                    await writer.WriteAsync($"{greeting}, {id}");
                    await otherWriter.WriteAsync($"{greeting}, {id}");
                }
            }
        }

        private static void Trace(string msg)
        {
            if (UseTracing) Console.WriteLine($"     {msg}");
        }
    }
 }
	using System;
	using System.Diagnostics;
	using System.Threading.Tasks;
	using System.Threading.Channels;
	using System.Linq;
	using System.Linq.Expressions;
	using System.Threading;

	namespace ChannelsDemo
	{
	/// <summary>
	/// Simple "getting-started" demo of System.Threading.Channels
	/// </summary>
	/// <remarks>
	/// There is no error handling demonstrated here. Typically, you would surround `WaitTo(Read\|Write)Async` with `try {} catch (ChannelClosedException) {}`
	/// but that would obscure some of the code here.
	/// </remarks>
	class Program
	{
	// NOTE: You can adjust these to see how various combinations work

	// NOTE: You might want to put some artificial delays in OutputGreeting (via `await Task.Delay(TimeSpan.FromSeconds(2))`
	private const int MaxItemsToProduce = 5_000;

	private static readonly int MaxReaders = Environment.ProcessorCount; // or, hard-coded value, like 1, 2, or 10
	private static readonly int MaxBoundedSize = 2 * MaxReaders; // or, hard-coded value, like 1, 2, or 400

	// NOTE: in all these examples, the time is dominated by the slowness of the Console. Try disabling this and comparing elapsed times.
	private const bool WriteToConsole = true;
	// NOTE: in scenarios where you want to simulate CPU-bound work, you can disable writing to console and set this instead. Calls Thread.SpinWait(n)
	private const int CpuSpinTime = 0;

	// NOTE: when experimenting with different strategies, this can help you visualize where tasks are in the channel
	private const bool UseTracing = false;

	/* Additional Note:
	* Performance is constrained when running a Debug build and/or in VS. Try building a Release build and running in a separate console.
	*
	* Examples from my machine using MaxItemsToProduce = 5000 and the MultipleCompetingTransformsAndConsumers example
	*
	* [WriteToConsole = true]
	* Debug/F5: 1_183 items/sec
	* Debug/Cmd: 1_350 items/sec
	* Release/F5: 1_204 items/sec
	* Release/Cmd: 1_371 items/sec
	*
	* [WriteToConsole = false]
	* Debug/F5: 49_341 items/sec
	* Debug/Cmd: 204_788 items/sec
	* Release/F5: 44_856 items/sec
	* Release/Cmd: 204_611 items/sec
	*
	* NOTE: There is also overhead for such a small number of items. Increasing MaxItemsToProduce to 500_000:
	*
	* [WriteToConsole = false]
	* Release/F5: 845_265 items/sec (MultipleCompetingTransformsAndConsumers)
	* Release/F5: 1_638_367 items/sec (Simple)
	*/

	static async Task Main(string[] args)
	{
	Console.WriteLine("Welcome to the Channels Demo app");

	// NOTE: 2 types of channels - unbounded and bounded. Bounded channels have a limit - writes will block until reads have drained the channel.
	// Unbounded channels will never block reads or writes, but assume you have capacity to store all items in memory.

	// Our first channel is just a list of ids that we want channel consumers to read, so we can easily hold hundreds of thousands in memory.
	// But for observing the flow of processing (i.e., with UseTracing = true), you might want to look at a Bounded channel
	var idChannel = Channel.CreateUnbounded<string>();
	//var idChannel = Channel.CreateBounded<string>(MaxBoundedSize);

	// Our second channel is doing work, and we want to constrain the amount of simultaneous work (both for memory and resource consumption).
	// But you can test various scenarios where you use an unbounded channel here as well.
	var greetChannel = Channel.CreateBounded<string>(MaxBoundedSize);
	//var greetChannel = Channel.CreateUnbounded<string>();

	var sw = Stopwatch.StartNew();

	// NOTE: because the examples close the channel, you can only run a single example at a time
	await Simple(idChannel, greetChannel);

	//await MultipleConsumers(idChannel, greetChannel);

	//await MultipleTransformsAndConsumers(idChannel, greetChannel);

	//var otherChannel = Channel.CreateBounded<string>(MaxBoundedSize);
	//await MultipleCompetingTransformsAndConsumers(idChannel, greetChannel, otherChannel);

	sw.Stop();

	Console.WriteLine($"{MaxItemsToProduce:N0} consumed in {sw.Elapsed.TotalSeconds:F1} seconds [{(MaxItemsToProduce/sw.Elapsed.TotalSeconds):N0} items/second]");
	}

	// Example task of producing data and writing it to a ChannelWriter
	private static async Task ProduceIds(ChannelWriter<string> writer, int max)
	{
	var ids = Enumerable.Range(1, max);

	foreach (var id in ids)
	{
	Trace($"{nameof(ProduceIds)}: writing id {id}");
	// NOTE: this works just fine...
	await writer.WriteAsync($"{id}");

	// NOTE: This is an optimization for avoiding the await when the channel isn't blocking, or has a large capacity
	// In the current case, since we have an unbounded writer, we would never need to await.
	//if (writer.TryWrite($"{id}") == false)
	//{
	// await writer.WriteAsync($"{id}");
	//}

	// NOTE: you can signal an error by calling `Complete(Exception)` - this will be an orderly shutdown.
	//if (id == 1337)
	//{
	// Trace("Closing the channel for 1337");
	// var e = new Exception($"You should never get a 1337");
	// writer.Complete(e);
	// // In the simple `writer.WriteAsync` above, if you `Complete` the channel, you will attempt to write an additional item to it.
	//}
	}
	Trace($"{nameof(ProduceIds)}: completed writing");
	// when we are done, signal the channel that no more data is coming
	writer.Complete();
	}

	// Example task of reading data from one channel (ChannelReader<T>) and writing it to another channel (ChannelWriter<T>)
	private static async Task TransformIdAndPublishGreeting(ChannelReader<string> reader,
	ChannelWriter<string> writer, string greeting, bool autoClose = false)
	{
	while (await reader.WaitToReadAsync())
	{
	Trace($"{nameof(TransformIdAndPublishGreeting)}: entered WaitToReadAsync");
	// NOTE: this will also work just fine..
	while (reader.TryRead(out var id))
	{
	await writer.WriteAsync(FormatGreeting(greeting, id));
	}

	// NOTE: but again, like above, we can avoid the unnecessary writing await as long as possible
	//while (reader.TryRead(out var id))
	//{
	// var msg = FormatGreeting(greeting, id);
	// if (writer.TryWrite(msg) == false)
	// {
	// Trace($"{nameof(TransformIdAndPublishGreeting)}: TryWrite: had to take slow path");
	// await writer.WriteAsync(msg);
	// }
	//}
	Trace($"{nameof(TransformIdAndPublishGreeting)}: exited WaitToReadAsync");
	}

	// NOTE: if we have a single instance of this task, we can close the writer when we're done;
	// but if we have multiple instances of this task, that would fail (because `Complete()` can only
	// be called once on a channel.
	if (autoClose)
	{
	Trace($"{nameof(TransformIdAndPublishGreeting)}: completing writer");
	writer.Complete();
	}
	}

	private static string FormatGreeting(string greeting, string id) => $"{greeting} {id}";

	// Example task of reading from a channel
	private static async Task OutputGreeting(ChannelReader<string> reader, string optionalPrefix = "")
	{
	while (await reader.WaitToReadAsync())
	{
	Trace($"{nameof(OutputGreeting)}{optionalPrefix}: entered WaitToReadAsync");

	while (reader.TryRead(out var greeting))
	{
	if (WriteToConsole)
	Console.WriteLine($"{optionalPrefix}{greeting}");
	// NOTE: this is here to heat up your CPUs
	else if (CpuSpinTime >= 0)
	Thread.SpinWait(CpuSpinTime);
	}

	Trace($"{nameof(OutputGreeting)}{optionalPrefix}: exited WaitToReadAsync");
	}
	}

	/* NOTE: Simple example
	* In this example, we are using a single task for each process - this is closest to a traditional pipe architecture, with the advantage that
	* we natively handle async actions.
	*/
	private static async Task Simple(Channel<string> idChannel, Channel<string> greetChannel)
	{
	var produceTask = Task.Run(() => ProduceIds(idChannel.Writer, MaxItemsToProduce));

	var transformTask = Task.Run(() => TransformIdAndPublishGreeting(idChannel.Reader, greetChannel.Writer, "Hello", autoClose: true));

	var greetTask = Task.Run(() => OutputGreeting(greetChannel.Reader));

	// NOTE: we await each task - in addition, we want to await any reader channels. Only readers have a `Task Completion`; writers have a `.Complete()` method.
	await Task.WhenAll(idChannel.Reader.Completion, produceTask, transformTask, greetChannel.Reader.Completion, greetTask);
	}

	/* NOTE: Multiple Consumers example
	* In this example, we are using a single task for producing and transforming, but multiple greeting processes. Since the greeting only reads,
	* there's minimal changes required.
	*/
	private static async Task MultipleConsumers(Channel<string> idChannel, Channel<string> greetChannel)
	{
	var produceTask = Task.Run(() => ProduceIds(idChannel.Writer, MaxItemsToProduce));

	var transformTask = Task.Run(() => TransformIdAndPublishGreeting(idChannel.Reader, greetChannel.Writer, "Hello", autoClose: true));

	var greetTasks = Enumerable.Range(0, MaxReaders)
	.Select(i => Task.Run(() => OutputGreeting(greetChannel.Reader, $"{i}")))
	.ToArray(); // NOTE: Very important that we materialize the running tasks, since `.Select` is lazy

	await Task.WhenAll(new[] { idChannel.Reader.Completion, produceTask, transformTask, greetChannel.Reader.Completion }.Concat(greetTasks));
	}

	/* NOTE: Multiple Transforms And Consumers example
	* In this example, we are using a single task for producing, but multiple transform and greeting processes. Since the transform task was
	* responsible for signalling that it was done writing in the single-task scenario, we now need to handle this explicitly.
	*/
	private static async Task MultipleTransformsAndConsumers(Channel<string> idChannel, Channel<string> greetChannel)
	{
	var produceTask = Task.Run(() => ProduceIds(idChannel.Writer, MaxItemsToProduce));

	var transformTasks = Enumerable.Range(0, MaxReaders * 2)
	.Select(i => Task.Run(() => TransformIdAndPublishGreeting(idChannel.Reader, greetChannel.Writer, $"'{i}' Hello", autoClose: false)))
	.ToArray();

	var greetTasks = Enumerable.Range(0, MaxReaders * 2)
	.Select(i => Task.Run(() => OutputGreeting(greetChannel.Reader, $"{i}")))
	.ToArray();

	// NOTE: once the transform tasks have completed writing to their channel, we can signal the greet tasks that no more data is coming.
	await Task.WhenAll(new [] {idChannel.Reader.Completion, produceTask }.Concat(transformTasks));
	greetChannel.Writer.Complete();

	await Task.WhenAll(new[] { greetChannel.Reader.Completion }.Concat(greetTasks));
	}

	/* NOTE: Multiple Competing Transforms And Consumers example
	* In this example, we are using a single task for producing, but multiple transform and greeting processes. In addition, the transform task
	* is now writing to multiple channels. In real code, this might be producing a read model and updating a search index. Like the MultipleTransformsAndConsumers
	* example, we are responsible for explicitly signalling that the transform tasks are completed writing.
	*/
	private static async Task MultipleCompetingTransformsAndConsumers(Channel<string> idChannel, Channel<string> greetChannel, Channel<string> otherChannel)
	{
	// NOTE: in this example, tweaking the running tasks is important for throughput. In my case, I have 12 processors.
	// If I leave these all set to `MaxReaders`, the best result is:
	// 205_000 items/sec
	// If I adjust them to MaxReaders/3 (so that consumers roughly equal CPUs), the best result is:
	// 218_702 items/sec
	// NOTE: this will also depend on the work each of your tasks is doing. If some are slow, you may wish to create more of them...
	// NOTE: the produceTask will also burn a CPU while it's working (for large # of items), so you may wish to calculate based on MaxReaders - 1

	var produceTask = Task.Run(() => ProduceIds(idChannel.Writer, MaxItemsToProduce));

	var readersPerTask = MaxReaders; // MaxReaders / 3;

	var transformTasks = Enumerable.Range(0, readersPerTask)
	.Select(i => Task.Run(() => TransformIdAndPublishGreeting(idChannel.Reader, greetChannel.Writer, otherChannel.Writer, $"'{i}' Hello")))
	.ToArray();

	// NOTE: We want to have multiple consumers, but we can use the same underlying function; we will use the optional prefix to help
	// distinguish between the two tasks in the console. Each set of greet tasks reads from it's own channel (so slowing down one will only
	// affect those tasks. If you want to see this, you can add a `if (optionalPrefix.StartsWith("OTHER") Thread.SpinWait(CpuSpinTime);` in
	// the `OutputGreeting` method.
	var greetTasks = Enumerable.Range(0, readersPerTask)
	.Select(i => Task.Run(() => OutputGreeting(greetChannel.Reader, $"GREET [{i:00}] ")))
	.ToArray();

	var otherTasks = Enumerable.Range(0, readersPerTask)
	.Select(i => Task.Run(() => OutputGreeting(otherChannel.Reader, $"OTHER: [{i:00}] ")))
	.ToArray();

	// NOTE: like in MultipleTransformsAndConsumers, we signal consumers that there's no more data to come. However, in this case, we need to
	// signal to both channels.
	await Task.WhenAll(new[] { idChannel.Reader.Completion, produceTask }.Concat(transformTasks));
	greetChannel.Writer.Complete();
	otherChannel.Writer.Complete();

	await Task.WhenAll(new[] { greetChannel.Reader.Completion, otherChannel.Reader.Completion }.Concat(greetTasks).Concat(otherTasks));
	}

	// NOTE: Overload of TransformIdAndPublishGreeting that takes a multiple channel writers (for the MultipleCompetingTransformsAndConsumers example)
	private static async Task TransformIdAndPublishGreeting(ChannelReader<string> reader,
	ChannelWriter<string> writer, ChannelWriter<string> otherWriter, string greeting)
	{
	while (await reader.WaitToReadAsync())
	{
	while (reader.TryRead(out var id))
	{
	await writer.WriteAsync($"{greeting}, {id}");
	await otherWriter.WriteAsync($"{greeting}, {id}");
	}
	}
	}

	private static void Trace(string msg)
	{
	if (UseTracing) Console.WriteLine($" {msg}");
	}
	}
	}