The Synchronous Blog

The point of this post is to show that it is wrong to rely on timing issues in preemptive multithreading for activites that require a synchronized behavior. I compare the same program specification in three different implementations.

The example program is supposed to blink two leds with different frequencies (400ms and 1000ms) on an Arduino board. They must blink together every 2 seconds.

The first two implementations use different RTOSes with preemptive multithreading. The third implementation uses the synchronous language Céu.

(UPDATE) The complete source files can be found here.

The first implementation uses the ChibiOS RTOS. I omitted all code related to creating and starting the threads (which run with the same priority).

Follows the code and video for the implementation in ChibiOS:

static msg_t Thread1(void *arg) {
    while (TRUE) {
        digitalWrite(11, HIGH);
        chThdSleepMilliseconds(400);
        digitalWrite(11, LOW);
        chThdSleepMilliseconds(400);
    }
}
static msg_t Thread2(void *arg) {
    while (TRUE) {
        digitalWrite(12, HIGH);
        chThdSleepMilliseconds(1000);
        digitalWrite(12, LOW);
        chThdSleepMilliseconds(1000);
    }
}

You can see that around 1:00 the leds loose synchronism among them, and also with the metronome.

The second implementation uses the DuinOS RTOS. I also omitted all code related to creating and starting the threads (which run with the same priority).

In this example the leds were well synchronized, so I included another task that uses the serial port with a different frequency.

Follows the code and video for the implementation in DuinOS:

taskLoop (T1) {
    digitalWrite(11, HIGH);
    delay(400);
    digitalWrite(11, LOW);
    delay(400);
}

taskLoop (T2) {
    digitalWrite(12, HIGH);
    delay(1000);
    digitalWrite(12, LOW);
    delay(1000);
}

int c = 0;
taskLoop (T3) {
    delay(77);
    Serial.print(c++);
}

Since the beginning you can see that the leds loose synchronism among them, and also with the metronome.

The third implementation uses the synchronous language Céu.

In this example the leds were well synchronized, even with the third activity that uses the serial port.

Follows the code and video for the implementation in Céu:

par do
    loop do
        _digitalWrite(11, _HIGH);
        await 400ms;
        _digitalWrite(11, _LOW);
        await 400ms;
    end
with
    loop do
        _digitalWrite(12, _HIGH);
        await 1s;
        _digitalWrite(12, _LOW);
        await 1s;
    end
with
    int c = 0;
    loop do
        await 77ms;
        c = c + 1;
        _Serial.print(c);
    end
end

Conclusion:

The execution model of preemptive multithreading does not ensure implicit synchronization among threads.

There’s nothing wrong with the RTOSes and the example implementations: the behavior shown in the videos is perfectly valid.

The problem is that usually the very first examples for these systems use blinking leds (supposely synchronized) to show how easy is to write multithreaded code. This is not true!

Preemptive multithreading should not be used as is to write this kind of highly synchronized applications. Adding semaphores or other synchronization primitives to these codes won’t help alone, they require a single thread to handle timers that is responsible to dispatching others.

I used timers in the example, but any kind of high frequency input would also behave nondeterministically in multithreaded systems.

In synchronous languages like Céu, the execution model enforces that all activities are synchronized all the time.

Click here to get the PDF with the full paper presented at SBLP entitled “LuaGravity: a Reactive Language based on Implicit Invocation”.