I have a home-made bluetooth device measuring ECG at 500Hz: every 2 ms the device sends 9 bytes of data (header, ECG measurment, footer). So this is roughly a 9*500=4.5kbytes/s data stream.
I have a C++ Windows program able to connect the device and retrieve the data stream (displaying it with Qt/qwt). In this case, I use Windows control panel to bond the device and I connect it via a virtual COM port using boost serial_port interface. This works perfectly and I'm receiving my data stream in real time: I get a measurment point every 2ms or so.
I ported the whole program on Android via QtCreator 3.0.1 (Qt 5.2.1). It appears that virtual COM ports cannot be accessed by boost (probably SDK permissions won't allow that) so I wrote a piece of Java code to open and manage the Bluetooth connection. So my app remains C++/Qt but only the layer connecting and reading data from the device was reworked in Java (opening the connexion with createInsecureRfcommSocketToServiceRecord):
Java code to read the data:
public int readData( byte[] buffer )
{
    if( mInputStream == null )
    {
        traceErrorString("No connection, can't receive data");
    }
    else
    {
        try
        {
            final boolean verbose = false;
            int available = mInputStream.available();
            if ( verbose )
            {
                Calendar c = Calendar.getInstance();
                Date date = new Date();
                c.setTime(date);
                c.get(Calendar.MILLISECOND);
                SimpleDateFormat sdf = new SimpleDateFormat("HH:mm:ss");
                String currentTime = sdf.format(date);
                traceDebugString( currentTime + ":" + c.get(Calendar.MILLISECOND) + " - " + available + " bytes available, requested " + buffer.length );
            }
            if ( available >= buffer.length )
                return mInputStream.read( buffer ); // only call read if we know it's not blocking
            else
                return 0;
        }
        catch (IOException e)
        {
            traceDebugString( "Failed to read data...disconnected?" );
        }
    }
    return -1;
}
Called from C++ like that:
bool ReceiveData( JNIEnv* env,
                  char* data,
                  size_t length,
                  bool& haserror )
{
    bool result = false;
    jbyteArray array = env->NewByteArray(length);
    jint res = env->CallIntMethod(j_object, s_patchIfReceiveDataID, array );
    if ( static_cast<size_t>(res) == length )
    {
        env->GetByteArrayRegion(array, 0, length, reinterpret_cast<jbyte*>(data));
        result = true;
    }
    else if ( res == -1 )
    {
        haserror = true;
    }
    else
    {
        // not enough data in the stream buffer
        haserror = false;
    }
    return result;
}
bool readThread( size_t blockSize )
{
    BTGETANDCHECKENV // retrieving environment
    char* buf = new char[blockSize];
    bool haserror = false;
    while ( !haserror )
    {
        if ( !ReceiveData( env, buf, blockSize, haserror ) )
        {
            // could not read data
            if ( haserror )
            {
                // will stop this thread soon
            }
            else
            {
                boost::this_thread::sleep( boost::posix_time::milliseconds( 10 ) );
            }
        }
    }
    delete [] buf;
    return true;
}
This works pretty well...for the five first seconds I'm gettings values in a sort of real time, then:
- Sometimes it freezes for ever, meaning the mInputStream.available() value remains lower than requested.
 - Sometimes it freezes only for a second or so and then it continues but data are received by blocks of ~1 second. Meaning mInputStream.available() can move from 0 to more than 3000 between two calls (elapsed by 10ms). Actually, I see the same during the 5 firsts seconds, but the buffer availability never exceeds 150 bytes, after 5 seconds, it can go up to 3000 bytes.
 
Here is what the log can look like when verbose is set to true:
14:59:30:756 - 0 bytes available, requested 3
14:59:30:767 - 0 bytes available, requested 3
14:59:30:778 - 0 bytes available, requested 3
14:59:30:789 - 1728 bytes available, requested 3
14:59:30:790 - 1725 bytes available, requested 6
14:59:30:792 - 1719 bytes available, requested 3
My ECG device definitely did not send 1728 bytes in 11ms!!
I know my device sends 9 bytes every 2ms (otherwise, it would not work on my PC application). Looks like Java does some unexpected buffering and does not make available 9 bytes every 2ms.... It's also strange things appear to work fine for only 5 seconds at the beginning.
Note that I tried using read() without checking available() (blocking version) but experienced exactly the same behaviour.
So I'm wondering what I'm doing wrong...
- Is there a way to force a Java input stream to update itself?
 - Is there a way to ask Java to proceed it's pending events (like we have QApplication::processEvents)?
 - Is there any global settings to specify buffer sizes for streams (I did not find any at BluetoothDevice/BluetoothSocket level)
 - On PC, when opening the virtual COM port, I have to specify baudrate, stop bit, handshaking and stuff like that. On Android I just open the Rfcomm socket with no option, could this be the problem (then ECG device and smartphone would not be synced...)?
 
Any help or idea would be welcomed!
Edit: I'm experiencing that on a Nexus 5 phone, Android 4.4.2 I just tested the same apk package on different devices:
- a Galaxy S4 with Android 4.4.2: Same problem.
 - a Galaxy S3 with custom CyanogenMod 11 Android 4.4.2: data streaming seems perfect, no freezing after 5sec and data are arriving in real-time....so looks like the whole system is able to achieve what I want, but looks like Android default setup makes things too slow....dunno if there could be a setting to be changed at the OS level to fix this issue.
 
Edit: As I got no answer :-( I tried to do the same thing using a pure Java program (no C++, no Qt). Had the same problem: Real-time Bluetooth SPP data streaming on Android only works for 5 seconds