RCCE: measuring round-trip latency between one core and all the other cores

pingpong-all.c

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <RCCE.h>
#include "timer.h"

#define PRINT(...) { printf(__VA_ARGS__); fflush(stdout); }
#define CORE(i)	if (RCCE_ue() == (i))

#define MSG_SIZE 32
#define ROUNDS 1000

extern double RC_REFCLOCKGHZ;

int RCCE_APP(int argc, char *argv[])
{
	mytimer_t timer;
	char buf[MSG_SIZE];
	int ME = 0, YOU, i, j;

	dup2(STDOUT_FILENO, STDERR_FILENO);

	RCCE_init(&argc, &argv);

	timer_new(&timer, RC_REFCLOCKGHZ);

	for (i = 0; i < RCCE_num_ues(); i++) {
		RCCE_barrier(&RCCE_COMM_WORLD);
		YOU = i;
		if (ME == YOU)
			continue;
		CORE(ME) {
			timer_reset(&timer, RC_REFCLOCKGHZ);
			timer_start(&timer);	
			for (j = 0; j < ROUNDS; j++) {
				RCCE_send(buf, MSG_SIZE, YOU);
				RCCE_recv(buf, MSG_SIZE, YOU);
			}
			timer_end(&timer);
		}

		CORE(YOU) {
			for (j = 0; j < ROUNDS; j++) {
				RCCE_recv(buf, MSG_SIZE, ME);
				RCCE_send(buf, MSG_SIZE, ME);
			}
		} 

		CORE(ME) PRINT("Round-trip latency %d <--> %d: %.2lf\n", ME, YOU, 
			       timer_elapsed(&timer, timer_us) / ROUNDS);
	}

	RCCE_finalize();

	return 0;
}