implementing a Fast .NET Lock-Free Inter-Process using SharedMemory MMF

Question

I am new to multitasking and IPC and I am trying to construct an approach for fast inter process comunication using shared memory( at first I was researching the IPC term, having in mind wcf sockets and named pipes only to eventually discover about MMF).

Now that I have successfully implemented a small test using shared memory between two processes via using Lock and EventWaitHandle signalling, I am up to an approach that implements non blocking / no-wait pattern. Now, I am trying to combine Thread.MemoryBarrier() & reading a signalling Sector from the MemoryMapedFile .

The problem is unidentified! First round goes through and second was last sited in Bermuda triangle ... out of the scope of the debugger...

say process a is sending a burst of requests for a showMsg() to process b .

                                         //offset positions in mmf
MemoryMappedViewAccessor MmfAcc; const int opReady= 0, opCompleteRead = 4, .....



ReadTrd()
{
  //[0,3] - Reader is stationed
  //[4,7] - Read Complete successfully
  //[8,11] - Data-size 
  //[12,15] - Reader-exiting
  "format" the signals Section (write zeroes). 

   for(;;){if (WrTrd-StepMMF1  Confimed) break;}
  MmfAcc- read DataSize val @offset[8]
  MmfAcc- read Data val @offset[50]
  MmfAcc Write exit to offset....
  ....heavy use of  Thread.MemoryBarrier(); sets !!! (all over the place, on every shared variable...)

}

writeTrd()
{
  heavy use of  Thread.MemoryBarrier() !!!
  //[15-19] - wr is stationed
  //[20-23] - wr Complete successfully
  //[24-27] - wrExiting
  "format" the signals Section . 
   for(;;){if Reader-StepMMF1 is Confim break;}
   MmfAcc- DataSize to offset[8]
   write Data To offset[50] using the method below
   for(;;){if Read StepMMF2 is Confim break;}
 } 

As I was first using the named piped solution, the Mmf approach(despite Lock and Event Wait Handle) was great performance gain compared to the namedpipe approach, but could I even go further using the above approach somehow.. ?

I could just clone this pattern like striping Raid ...

Reader1 + Reader2 & WriteThred1  + WriteThread2

so I tried it and got stuck at that point.

Is this valid approach using the full-memoryfence & sharedmemory for signalling?

If so, all is left is to see why the second iteration failed , the performance difference.

EDIT - added the logic behind the extra threads test

This is the "Bridge" I am using to manipulate writer threads (same approach for readers.

public void Write(byte[] parCurData)
{
    if (ReadPosition < 0 || WritePosition < 0)
        throw new ArgumentException();
    this.statusSet.Add("ReadWrite:-> " + ReadPosition + "-" + WritePosition);
    // var s = (FsMomitorIPCCrier)data;

    ////////lock (this.dataToSend)
    ////////{
    Thread.MemoryBarrier();
        LiveDataCount_CurIndex = dataQue.Where(i => i != null).Count();
    this.dataQue[LiveDataCount_CurIndex] = parCurData;

    Console.WriteLine("^^^^^" + Thread.CurrentThread.Name + " has Entered WritingThreads BRIDGE");
    Console.WriteLine("^^^^^[transactionsQue] = {1}{0}^^^^^[dataQue.LiveDataASIndex = {2}{0}^^^^^[Current Requests Count = {3}{0}", "\r\n", Wtransactions, LiveDataCount_CurIndex, ++dataDelReqCount);

    //this.itsTimeForWTrd2 = false;

    if (Wtransactions != 0 && Wtransactions > ThrededSafeQ_Initial_Capcity - 1)
    if (this.dataQueISFluded) this.DataQXpand();


    if (itsTimeForWTrd2)
    {

        bool firstWt = true;
        while (writerThread2Running)
        {
            if (!firstWt) continue;
            Console.WriteLine("SECOND WRITERThread [2] is In The CoffeeCorner");
                    firstWt=false;
        }

        this.dataDelivery2 = this.dataQue[LiveDataCount_CurIndex];
        Console.WriteLine("Activating SECOND WRITERThread [2]");
        itsTimeForWTrd2 = false;

        writerThread2Running = true;
        //writerThread1Running = true;
        writerThread2 = new System.Threading.Thread(WriterThread2);
        writerThread2.IsBackground = true;
        writerThread2.Name = this.DepoThreadName + "=[WRITER2]";
        writerThread2.Start();

    }
    else
    {
        bool firstWt = true;
        while (writerThread1Running)
        {
            if (!firstWt)continue;
                Console.WriteLine("WRITERThread [1] is In The CoffeeCorner");
            firstWt=false;
        }
        Console.WriteLine("Activating WRITERThread [1]");
        this.dataDelivery1 = this.dataQue[LiveDataCount_CurIndex]; 

        writerThread1Running = true;
        writerThread1 = new System.Threading.Thread(WriterThread1);
        writerThread1.IsBackground = true;
        writerThread1.Name = this.DepoThreadName+"=[WRITER1]";
        writerThread1.Start();
        itsTimeForWTrd2 = true;

    }
    Thread.MemoryBarrier();
}

Using the write handle to read & write the actual data (similar code for the write)

public unsafe byte[] UsReadBytes(int offset, int num)
{
    byte[] arr = new byte[num];
    byte* ptr = (byte*)0;
    this.accessor.SafeMemoryMappedViewHandle.AcquirePointer(ref ptr);
    Marshal.Copy(IntPtr.Add(new IntPtr(ptr), offset), arr, 0, num);
    this.accessor.SafeMemoryMappedViewHandle.ReleasePointer();
    return arr;
}

As I said, I have researched this issue of synchronization of the data and shared memory via non blocking / no-wait / etc. Semaphores Locks so I am trying to remove any kind of overhead during the process of each transaction of data into the shared memory mapped file. I am here to ask what could be the problem eliminating the Lock And The EventWaitHandle and replacing it with the logic of memory fences and signaling through the mmf?

Answer 1

If you are planning to use this for an specific purpose other than R&D, the easiest way would be to use a library that already provides that. One way to think about this is, Lock Free = Message Passing. Two possible approaches are: For a minimalistic message-passing implementation, ZeroMQ IPC, which provides great .Net support and superb IPC performance ( http://www.codeproject.com/Articles/488207/ZeroMQ-via-Csharp-Introduction ) and for a more complete Actor-Model implementation (including support for IPC), look at Akka.net ( http://getakka.net/docs/#networking ).

If the purpose is more R&D in nature (meaning, you want to write your own implementation, which is cool), I would still suggest to look at the source of these products (especially, Akka.net, since it's written in C#), for implementation ideas about Message-passing and Actor-based IPC.