By default, ThreadPool has 25 worker threads per CPU and 1000 IO
threads for each process. I am wondering how CLR managed to allocate
IO asynchronous tasks to IO thread and other asynchonous tasks to
worker thread.

With the help of Reflector, I found that default implementation of
Stream.Read just use BeginInovke to get a IAsynResult.

   ReadDelegate delegate2 = new ReadDelegate(this.Read);
   ... ...
   return delegate2.BeginInvoke(buffer, offset, count, callback,
state);

Since subclass of Stream may not involves I/O, e.g. MemoryStream. This
is understandable.

According to reflected code of FileStream.BeginRead, it is indeed
implemented with I/O completion port. That is also fine.

However, for stream returned by HttpResponse.GetResponseStream is
System.Net.ConnectStream. And it use System.Net.Connection to do
BeginRead. Connection has no its own implementation. So, it is
actually use Stream.BeginRead. As a result, ConnnectStream.BeginRead
use BeginInvoke. In my understanding, task passed to BeginInvoke will
be dispatched to worker thread.

I make a small program to launch 50 Http requests and read their
response streams with BeginRead. There is no threads burst. It seems
that completion port i/o is still used. I set breakpoint at callback
to BeginRead, the stack trace shows that it is called from completion
port as below

    System.dll!System.Net.LazyAsyncResult.Complete(System.IntPtr
userToken) + 0x7f bytes
    System.dll!System.Net.ContextAwareResult.CompleteCallback(object
state) + 0x1a bytes
    mscorlib.dll!
System.Threading.ExecutionContext.Run(System.Threading.ExecutionContext
executionContext, System.Threading.ContextCallback callback, object
state) + 0x81 bytes
    System.dll!System.Net.ContextAwareResult.Complete(System.IntPtr
userToken) + 0xa7 bytes
    System.dll!System.Net.LazyAsyncResult.ProtectedInvokeCallback(object
result, System.IntPtr userToken) + 0x8b bytes
    System.dll!
System.Net.Sockets.BaseOverlappedAsyncResult.CompletionPortCallback(uint
errorCode, uint numBytes, System.Threading.NativeOverlapped*
nativeOverlapped) + 0x116 bytes
    mscorlib.dll!
System.Threading._IOCompletionCallback.PerformIOCompletionCallback(uint
errorCode, uint numBytes, System.Threading.NativeOverlapped* pOVERLAP)
+ 0x68 bytes

I cannot understand this. System.Net.ConnectStream is using
BeginInvoke to launch asynchronous task; and the task is done with IO
completion port. Then, CLR must have a way to distinguish IO task and
other task, but how?