Hello,

I've recently profiled the latency of LTTng tracepoints on arm platforms,

using the follow sample program:

----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

static inline uint64_t get_time_nsec(void)

{

	struct timespec ts;

	if (caa_unlikely(clock_gettime(CLOCK_MONOTONIC, &ts))) {

		ts.tv_sec = 0;

		ts.tv_nsec = 0;

	}

	return ((uint64_t) ts.tv_sec * 1000000000ULL) + ts.tv_nsec;

}

int main(int argc, char *argv[])

{

    unsigned int i;

    int tp_num = 0;

    uint64_t total_time = 0;

    uint64_t now, nowz;

    if (argc > 1) {

        sscanf (argv[1],"%d",&tp_num);

    }

    for (i = 0; i < tp_num; i++) {

        now = get_time_nsec();

        lttng_ust_tracepoint(hello_world, my_first_tracepoint,

                             i, "some_str");

        nowz = get_time_nsec();

        total_time += (nowz - now);

    }

    if (tp_num) {

        printf("---------------------------Average TP time is %"PRIu64"---------------------------\n", total_time / tp_num);      

    }

}

----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

I observed a big average latency variance on different platforms when tracing a high number (many thousands to millions) of tracepoints:

[platform 1] with CPU info running a linux kernel based on Buildroot (4.19.273 aarch64 GNU/Linux):

 Saw an average latency of 2-3usec

[platform 2] with CPU info running a linux kernel based on Amazon Linux (4.14.294-220.533.amzn2.aarch64 aarch64 GNU/Linux):

Saw an average latency of ~0.5usec

Are there any suggestions to root cause the high latency and potentially improve it on platform 1?

Thanks and best regards,

Anas.