diff --git a/test/uti/mpi/delay.c b/test/uti/mpi/delay.c
new file mode 100644
index 00000000..0ae14d49
--- /dev/null
+++ b/test/uti/mpi/delay.c
@@ -0,0 +1,70 @@
+#define _GNU_SOURCE         /* See feature_test_macros(7) */
+#include <stdio.h>
+#include <unistd.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <sys/syscall.h>   /* For SYS_xxx definitions */
+#include <sched.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <string.h>
+#include <time.h>
+#include "delay.h"
+
+#define N_INIT 10000000
+double nspw; /* nsec per work */
+
+void ndelay_init(int verbose)
+{
+	struct timeval start, end;
+
+	//clock_gettime(TIMER_KIND, &start);
+	gettimeofday(&start, NULL);
+
+#pragma omp parallel
+	{
+		asmloop(N_INIT);
+	}
+
+	//clock_gettime(TIMER_KIND, &end);
+	gettimeofday(&end, NULL);
+
+	nspw = DIFFUSEC(end, start) * 1000 / (double)N_INIT;
+	if (verbose) {
+		pr_debug("nspw=%f\n", nspw);
+	}
+}
+
+void ndelay(long delay_nsec)
+{
+	if (delay_nsec < 0) {
+		printf("delay_nsec < 0\n");
+		return;
+	}
+#pragma omp parallel
+	{
+		asmloop(delay_nsec / nspw);
+	}
+}
+
+static double cycpw; /* cyc per work */
+
+void cdlay_init(void)
+{
+	unsigned long start, end;
+
+	start = rdtsc_light();
+#define N_INIT 10000000
+	asmloop(N_INIT);
+	end = rdtsc_light();
+	cycpw = (end - start) / (double)N_INIT;
+}
+
+void cdelay(long delay_cyc)
+{
+	if (delay_cyc < 0) {
+		return;
+	}
+	asmloop(delay_cyc / cycpw);
+}
diff --git a/test/uti/mpi/delay.h b/test/uti/mpi/delay.h
new file mode 100644
index 00000000..8d79ac88
--- /dev/null
+++ b/test/uti/mpi/delay.h
@@ -0,0 +1,40 @@
+#ifndef __DELAY_H_INCLUDED__
+#define __DELAY_H_INCLUDED__
+
+static inline uint64_t rdtsc_light(void)
+{
+	uint64_t x;
+
+	/* rdtscp don't jump over earlier instructions */
+	__asm__ __volatile__("rdtscp;"
+			     "shl $32, %%rdx;"
+			     "or %%rdx, %%rax" :
+			     "=a"(x) :
+			     :
+			     "%rcx", "%rdx", "memory");
+	return x;
+}
+
+static inline void asmloop(unsigned long n)
+{
+	int j;
+
+	for (j = 0; j < n; j++) {
+		asm volatile(
+			     "movq $0, %%rcx\n\t"
+			     "1:\t"
+			     "addq $1, %%rcx\n\t"
+			     "cmpq $99, %%rcx\n\t"
+			     "jle 1b\n\t"
+			     :
+			     :
+			     : "rcx", "cc");
+	}
+}
+
+void ndelay_init(void);
+void ndelay(long delay_nsec);
+void cdelay_init(void);
+void cdelay(long delay_cyc);
+
+#endif
diff --git a/test/uti/mpi/fwq.c b/test/uti/mpi/fwq.c
new file mode 100644
index 00000000..88247e52
--- /dev/null
+++ b/test/uti/mpi/fwq.c
@@ -0,0 +1,38 @@
+#define _GNU_SOURCE         /* See feature_test_macros(7) */
+#include <stdio.h>
+#include <unistd.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <sys/syscall.h>   /* For SYS_xxx definitions */
+#include <sched.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <string.h>
+#include <time.h>
+#include "fwq.h"
+
+static double nspw; /* nsec per work */
+
+void fwq_init(void)
+{
+	struct timespec start, end;
+	unsigned long nsec;
+	int i;
+
+	clock_gettime(CLOCK_THREAD_CPUTIME_ID, &start);
+#define N_INIT 10000000
+	bulk_fsw(N_INIT);
+	clock_gettime(CLOCK_THREAD_CPUTIME_ID, &end);
+	nsec = DIFFNSEC(end, start);
+	nspw = nsec / (double)N_INIT;
+}
+
+void fwq(long delay_nsec)
+{
+	if (delay_nsec < 0) {
+		return;
+		//printf("%s: delay_nsec < 0\n", __func__);
+	}
+	bulk_fsw(delay_nsec / nspw);
+}