# Golang channel in C#

I have tried to mimic Golang channels in C# and its performance is pretty good compared to golang itself. On my machine, each channel operation of Golang takes ~75 nano-sec and each Chan<T> (in C#) operation takes ~90 nano-sec.

Please let me know if this code can be improved in any way.

class Chan<T>
{

T[] buffer;
long tail = -1;
long closed = 0;

public Chan() { this.size = 0; }
public Chan(int size)
{
if (size < 0) throw new ArgumentOutOfRangeException();

this.size = size;
this.buffer = new T[this.size];
}

public bool To(T t)
{
{
long localClosed = 0L;

if (tail - head == buffer.Length) SpinWait.SpinUntil(() => (localClosed = Interlocked.Read(ref closed)) > 0 || tail - head < buffer.Length);
if (localClosed > 0) return false;

var newTail = Interlocked.Increment(ref tail);
buffer[newTail % buffer.Length] = t;

return true;
}
}

object tailLock = new object();
public bool From(out T val)
{
lock (tailLock)
{
long localClosed = 0L;

if (tail - head == 0) SpinWait.SpinUntil(() => (localClosed = Interlocked.Read(ref closed)) > 0 || tail - head > 0);
if (localClosed > 0)
{
val = default(T);
return false;
}

return true;
}
}

public void Close()
{
Interlocked.Increment(ref closed);
}
}

• Have you tried TPL Dataflow? I think BufferBlock is quite similar to Go channel. – svick Oct 9 '13 at 23:55
• No, I have not tried it; Thanks for mentioning it. But as far as I know BufferBlock does not provide the concept of 'closing' a channel. – Kaveh Shahbazian Oct 10 '13 at 0:27
• @KavehShahbazian It does, that's what the Complete() method is for. Though it does it differently than your code: after completing a block, you can't add new items to it, but the items that are already in it can still be processed. – svick Oct 10 '13 at 9:44
• @svick I'll look into it (I can not find it in .NET framework so I guess there should be a NuGet for it). – Kaveh Shahbazian Oct 10 '13 at 9:55
• Link for TPL Dataflow – rwong Oct 11 '13 at 7:18

Without the select operator, you haven't got Go channels - you've just got buffered queues, which are much easier to implement but much less useful.

Also, it's important to allow channels with a zero size buffer - in that case the sender should synchronise with the receiver.

• Ouch! I totally forgot select! – Kaveh Shahbazian Oct 10 '13 at 11:35
• It may be worth to mention that the default constructor of Chan should set the buffer size to zero, not one. The default buffer size of a chan in Go is zero. – rightfold Apr 9 '14 at 14:05
1. Chan sounds like the abbreviated name for Channel and apparently it's a channel. So I'd use Channel or maybe even GoChannel.
2. The most commonly used naming convention I have seen for private members is to prefix them with an underscore. This way you can see at the first glance whether it's a local variable or a class member. This also means you can get rid of this. most of the time.
3. I really prefer to spell out the access modifier even if private is default but YMMV.
4. Method names should describe actions or operations (because they operate on data and sometimes modify the state of the object). To and From are not actions or operations. That being said: The channel seems to have fixed size queue semantics (FIFO) so I'd consider calling the operations Enqueue and Dequeue which would make it immediately clear how the data is being processed. (The semantic of the underlying data structure should not be exposed. Don't know what I was thinking there.) Rather use the Send and Receive semantics from the go definition.
5. Given the previous point it could be useful to have a Peek method to check what will come next.
6. _head and _tail are longs and access is not guaranteed to be atomic so you should use Interlocked.Read to obtain them.
7. Also the implementation is actually broken. Assume two threads A and B, A calls Send() and B calls Receive(), first execution _head == _tail == -1:

• A: execute Interlocked.Increment(_tail) (_tail is now 0)
• B: _tail - _head > 0 is true (0 - -1 == 1), leaves spinlock
• B: execute Interlocked.Increment(_head) (_head is now 0)
• B: read _buffer[_head]
• A: write _buffer[_tail]
• B has read from buffer before element was written.
• This problem can be easily reproduced with this test case (almost every iteration results in dupes):

[TestCase]
public void TestSPSC()
{
int numItems = 10000;
int numIterations = 100;

for (int i = 0; i < numIterations; ++i)
{
var channel = new Channel<int>(100);
var writer = Task.Factory.StartNew(() => { foreach (var num in Enumerable.Range(1, numItems)) { channel.Send(num); } channel.Close(); });
var numbers = new List<int>(numItems);
for (int idx = 1; idx <= numItems; ++idx)
{
int num;
}
return numbers.OrderBy(x => x).ToList();
});
var dupes = reader.Result.GroupBy(x => x).Where(g => g.Count() > 1).ToList();
if (dupes.Count > 0)
{
Console.WriteLine("{0}: {1} DUPES!", i, dupes.Count);
}
}
}


I changed the implementation to use .NET's BlockingCollection<T> wrapped around a ConcurrentQueue<T>:

public class Channel<T>
{
private BlockingCollection<T> _buffer;

public Channel() : this(1) { }
public Channel(int size)
{
_buffer = new BlockingCollection<T>(new ConcurrentQueue<T>(), size);
}

public bool Send(T t)
{
try
{
}
catch (InvalidOperationException)
{
// will be thrown when the collection gets closed
return false;
}
return true;
}

{
try
{
val = _buffer.Take();
}
catch (InvalidOperationException)
{
// will be thrown when the collection is empty and got closed
val = default(T);
return false;
}
return true;
}

public void Close()
{
}

public IEnumerable<T> Range()
{
T val;
{
yield return val;
}
}
}


The code is much easier to read and has probably less bugs than your self implemented one. It's also fast. I can pump 10,000,000 items (I tested with int) through a channel (buffer size 100) with single producer single consumer in 5sec. That's 0.5ns per item.

    [TestCase]
public void TestSPSC_Performance()
{
int numItems = 10000000;
int numIterations = 10;

var stopWatch = new Stopwatch();
stopWatch.Start();
for (int i = 0; i < numIterations; ++i)
{
var channel = new Channel<int>(100);
var writer = Task.Factory.StartNew(() => { foreach (var num in Enumerable.Range(1, numItems)) { channel.Send(num); } channel.Close(); });
var reader = Task.Factory.StartNew<List<int>>(() => { var res = new List<int>(numItems); foreach (var num in channel.Range()) { res.Add(num); } return res; });
}
stopWatch.Stop();

var elapsedMs = stopWatch.Elapsed.TotalMilliseconds;
Console.WriteLine("SPSC N = {0}: {1:.00}ms/iteration, {2:.00}ns/item (tx+rx)", numItems,  elapsedMs / numIterations, elapsedMs * 1000.0 / numItems / numIterations);
}

• I was trying to mimic golang chan. But in C# land you are right about naming; I've done some modifications based on your answer. – Kaveh Shahbazian Oct 10 '13 at 10:16
• A Peek method is not in general a good idea because it's ok for several receivers to use the channel at once, in which case Peek cannot necessarily provide an accurate prediction of what you will read, because another process might have read the value first. That's why Go's channels do not provide that primitive. – rog Oct 10 '13 at 11:35
• @rog, sure but if you only have 1 consumer then it might be – ChrisWue Oct 10 '13 at 18:48
• If you've got only one consumer and the buffer size is greater than zero and you must avoid actually reading the value. Better to leave it out IMHO - it's a misleading operation. – rog Oct 11 '13 at 12:35
• @KavehShahbazian: Found a problem with your implementation. I got interested in this and started implementing Select and wrote a bunch of unit tests for the class. – ChrisWue Oct 15 '13 at 7:17

I would not use SpinWait, since this basically runs a small loop that checks the condition over and over again. This means a lot of CPU cycles are wasted. I would suggest a signalling construct like the ManualResetEventSlim class.

Note: the SpinWait class is only preferred when you know in advance that the wait times will be very small (smaller than time it takes to do a thread context switch). You can configure the ManulResetEventSlim class to spin for a short time and then fall back to a kernel-based wait operation, by setting the SpinCount property. This is usefull if you expect a very short wait time, but don't want to waste a too manu CPU cylces when it turns out that you have to wait longer.