Understanding I2C transactions / PCA9685 issue with "turning off PWM"

[Sciss]

Hi there. I opened this as https://github.com/Pi4J/pi4j/issues/540 , but perhaps it's better for discussion. Basically I am trying to disable the PWM for certain channels, when the servos that make up a robot arm I am using, should not longer apply force - the arm is virtually rested, and turning off the servos would make it completely reset in a particular position, while keeping the servos running would have them constantly work against gravity to sustain the arm.

Now, I am using the PCA9685 code that is part of Pi4j 1.3. That of course talks to the I2CDevice and I2CBus. The code to turn off in the library looks like this:

    public void setAlwaysOff(Pin pin) {
...
        final int channel = pin.getAddress();
        try {
            device.write(PCA9685A_LED0_ON_L  + 4 * channel, (byte) 0x00);
            device.write(PCA9685A_LED0_ON_H  + 4 * channel, (byte) 0x00);
            device.write(PCA9685A_LED0_OFF_L + 4 * channel, (byte) 0x00);
            device.write(PCA9685A_LED0_OFF_H + 4 * channel, (byte) 0x10); // set bit 4 to high
        } catch (IOException e) {
            throw new RuntimeException("Error while trying to set channel [" + channel + "] always OFF.", e);
        }
...
    }

The effect I am seeing is that in some cases, perhaps one out of ten, or out of fifteen, some of the servos (arm consists of six channels) make a rapid movement to their minimum position instead of just releasing their force, potentially causing damage.

I would like to understand better what causes this, and how I can avoid it. Looking at the above code, the first thing that strikes me is that there seems to be no transaction in which this happens, so unless I am missing something, those four write commands happen one after the other and asynchronous to whatever the chip is doing. If I understand the specs correctly, the only command that is actually needed is to set the 12th bit, i.e. the 4th bit of the high byte of PCA9685A_LED0_OFF_H to 1. Could it be that the preceding writes are causing the problem, so that sometimes, an oscillator cycle completes between the PCA9685A_LED0_OFF_L and the PCA9685A_LED0_OFF_H write, and thus there is a single cycle with minimum duty cycle, causing the servo to jump to minimum position? Or is the problem located elsewhere? Would it help to use the array-write function instead of four individual byte writes? I don't really understand how a linux device file (RandomAccessFile) is used, and if that comes with any form of atomic updates.

I also notice both from the Pi4j code and the specs, that the chip understands a SLEEP mode. Should I perhaps use that instead? As far as I understand, it will also disconnect the internal oscillator, presumably halting any new levels sent to the servos.

Thanks for help!