diff --git a/test/qlmpi/rusage/Makefile b/test/qlmpi/rusage/Makefile
index aee00914..079a6738 100644
--- a/test/qlmpi/rusage/Makefile
+++ b/test/qlmpi/rusage/Makefile
@@ -29,7 +29,6 @@ rusage010: rusage010.o
 rusage010.o: rusage010.c
 	$(CC) $(CCFLAGSMCK) $(CPPFLAGSMCK) -c $<
 
-
 rusage008: rusage008.o
 	$(CC) -o $@ $^ $(LDFLAGSMCK)
 
@@ -48,6 +47,17 @@ rusage011: rusage011.o
 rusage011.o: rusage011.c
 	$(CC) $(CCFLAGSMCK) $(CPPFLAGSMCK) -c $<
 
+rusage002: rusage002.o
+	$(CC) -o $@ $^ $(LDFLAGSMCK) -lrt
+
+rusage002.o: rusage002.c
+	$(CC) $(CCFLAGSMCK) $(CPPFLAGSMCK) -c $<
+
+rusage003: rusage003.o
+	$(CC) -o $@ $^ $(LDFLAGSMCK) -lrt
+
+rusage003.o: rusage003.c
+	$(CC) $(CCFLAGSMCK) $(CPPFLAGSMCK) -c $<
 
 clean:
 	rm -f core $(EXES) $(OBJS) $(EXESMCK) $(OBJSMCK)
diff --git a/test/qlmpi/rusage/README b/test/qlmpi/rusage/README
index 12b3b442..173f6bc4 100644
--- a/test/qlmpi/rusage/README
+++ b/test/qlmpi/rusage/README
@@ -2,14 +2,15 @@
 Test matrix
 ===========
 
+rusage002:
+  do_mmap(),/dev/shm with --mpol_shm_premap,pre-page->get_page(),st->set_pte()->munmap()->clear_range() [OK]
+
+rusage003:
+  do_mmap(),/dev/shm without --mpol_shm_premap,pre-page->get_page(),st->set_pte()->munmap()->clear_range() [OK]
+
 rusage010:
   app->mmap() 2M,anon,pre-page ->set_range()->munmap()->free_process_memory_range()->clear_range()[OK]
 
-rusage005: device file (ib ping-pong)
-  devobj()->get_page()->pf->munmap()->clear_range() [OK]
-  remote page fault->cow->clear_range() [OK]
-  ld-linux.so->mmap private->cow->clear_range() [OK]
-
 rusage008: sharing file-map page
 fork()->filemap->pf->clear_range() [OK]
 
@@ -19,3 +20,8 @@ fork()->shmat()->pf->clear_range() [OK]
 rusage011: sharing xpmem page
 fork()->xpmem_attach()->pf->clear_range() [OK]
 
+device file (ib ping-pong) in verbs directory
+  devobj()->get_page()->pf->munmap()->clear_range() [OK]
+  remote page fault->cow->clear_range() [OK]
+  ld-linux.so->mmap private->cow->clear_range() [OK]
+
diff --git a/test/qlmpi/rusage/enable_swap.patch b/test/qlmpi/rusage/enable_swap.patch
new file mode 100644
index 00000000..b8c17edb
--- /dev/null
+++ b/test/qlmpi/rusage/enable_swap.patch
@@ -0,0 +1,14 @@
+diff --git a/executer/user/qlmpilib.c b/executer/user/qlmpilib.c
+index 7fcbcb5..9058ec8 100644
+--- a/executer/user/qlmpilib.c
++++ b/executer/user/qlmpilib.c
+@@ -158,7 +158,8 @@ int ql_client(int *argc,char ***argv)
+ 	syscall(803);
+ 	rc = PMI_Barrier();
+ 	
+-	rc = swapout(swap_file, buffer, BUF_SIZE, 1);
++	//rc = swapout(swap_file, buffer, BUF_SIZE, 1);
++	rc = swapout(swap_file, buffer, BUF_SIZE, 0);
+ 
+ #ifdef QL_DEBUG
+ 	printf(" swapout rc=%d\n",rc);
diff --git a/test/qlmpi/rusage/run.sh b/test/qlmpi/rusage/run.sh
index 1befba69..9128c335 100755
--- a/test/qlmpi/rusage/run.sh
+++ b/test/qlmpi/rusage/run.sh
@@ -31,6 +31,11 @@ if [ "${pid}" != "" ]; then
 fi
 
 case ${testname} in
+	rusage002)
+	mcexecopt="--mpol-shm-premap"
+	;;
+	rusage003)
+	;;
     rusage010)
 	testopt="1"
 	;;
diff --git a/test/qlmpi/rusage/rusage002.c b/test/qlmpi/rusage/rusage002.c
new file mode 100644
index 00000000..1033e0f6
--- /dev/null
+++ b/test/qlmpi/rusage/rusage002.c
@@ -0,0 +1,99 @@
+#include <unistd.h>
+#include <stdio.h>
+#include <sys/mman.h>
+#include <string.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <errno.h>
+#include "qltest.h"
+
+#define DEBUG
+
+#ifdef DEBUG
+#define dprintf(...)                                            \
+    do {                                                        \
+	char msg[1024];                                         \
+	sprintf(msg, __VA_ARGS__);                              \
+	fprintf(stderr, "%s,%s", __FUNCTION__, msg);            \
+    } while (0);
+#define eprintf(...)                                            \
+    do {                                                        \
+	char msg[1024];                                         \
+	sprintf(msg, __VA_ARGS__);                              \
+	fprintf(stderr, "%s,%s", __FUNCTION__, msg);            \
+    } while (0);
+#else
+#define dprintf(...) do {  } while (0)
+#define eprintf(...) do {  } while (0)
+#endif
+
+#define CHKANDJUMP(cond, err, ...)                                      \
+    do {                                                                \
+		if(cond) {                                                      \
+			eprintf(__VA_ARGS__);                                       \
+			ret = err;                                                  \
+			goto fn_fail;                                               \
+		}                                                               \
+    } while(0)
+
+int sz_mem[] = {
+	4 * (1ULL<<10),
+	2 * (1ULL<<20),
+	1 * (1ULL<<30),
+	134217728};
+
+#define SZ_INDEX 0
+
+int main(int argc, char** argv) {
+	void* mem;
+	int ret = 0;
+	int fd;
+	char fn[256] = "/dev/shm/Intel_MPI";
+#define TEST_VAL 0x1234
+	int swap_rc = 0;
+	char buffer[BUF_SIZE];
+	
+	fd = open(fn, O_CREAT | O_TRUNC | O_RDWR, S_IRUSR | S_IWUSR);
+	CHKANDJUMP(fd == -1, 255, "shm_open failed,str=%s\n", strerror(errno));
+
+	ret = ftruncate(fd, sz_mem[SZ_INDEX]);
+	CHKANDJUMP(ret != 0, 255, "ftruncate failed\n");
+
+	mem = mmap(0, sz_mem[SZ_INDEX], PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
+	CHKANDJUMP(mem == MAP_FAILED, 255, "mmap failed\n");
+	memset(mem, 0, sz_mem[SZ_INDEX]);
+
+	// before swap
+	*((unsigned long*)mem) = TEST_VAL;
+	unsigned long val = *((unsigned long*)mem);
+	if (val == TEST_VAL) {
+		printf("[OK] before swap, val:0x%lx\n", val);
+	} else {
+		printf("[NG] before swap, val is not correct, val:0x%lx\n", val);
+	}
+
+	swap_rc = do_swap("/tmp/rusage002.swp", buffer);
+	if (swap_rc < 0) {
+		printf("[NG] swap is failed\n");
+	}
+
+	// after swap
+	val = *((unsigned long*)mem);
+	if (val == TEST_VAL) {
+		printf("[OK] after swap,  val:0x%lx\n", val);
+	} else {
+		printf("[NG] after swap,  val is not correct, val:0x%lx\n", val);
+	}
+
+	munmap(mem, sz_mem[SZ_INDEX]);
+	ret = close(fd);
+	CHKANDJUMP(ret != 0, 255, "close failed\n");
+	ret = unlink(fn);
+	CHKANDJUMP(ret != 0, 255, "shm_unlink failed\n");
+	
+ fn_exit:
+	return ret;
+ fn_fail:
+	goto fn_exit;
+}
diff --git a/test/qlmpi/rusage/rusage003.c b/test/qlmpi/rusage/rusage003.c
new file mode 120000
index 00000000..239a35e2
--- /dev/null
+++ b/test/qlmpi/rusage/rusage003.c
@@ -0,0 +1 @@
+rusage002.c
\ No newline at end of file
diff --git a/test/qlmpi/rusage/rusage008.c b/test/qlmpi/rusage/rusage008.c
index 8c395255..bdf4abd9 100644
--- a/test/qlmpi/rusage/rusage008.c
+++ b/test/qlmpi/rusage/rusage008.c
@@ -54,6 +54,7 @@ int main(int argc, char** argv) {
 	int status;
 	int fd;
 // for swap_test
+#define TEST_VAL 0x1234
 	int swap_rc = 0;
 	char buffer[BUF_SIZE];
 	
@@ -65,14 +66,16 @@ int main(int argc, char** argv) {
 		
 		mem = mmap(0, sz_mem[SZ_INDEX], PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
 		CHKANDJUMP(mem == MAP_FAILED, 255, "mmap failed\n");
-		
-		unsigned long val = *((unsigned long*)mem);
+		memset(mem, 0, sz_mem[SZ_INDEX]);
 
 // for swap_test
 		swap_rc = do_swap("/tmp/rusage008_c.swp", buffer);
 		if (swap_rc < 0) {
 			printf("[NG] swap in child is failed\n");
 		}
+
+		*((unsigned long*)mem) = TEST_VAL;
+
 		_exit(123);
 	} else {
 		fd = open("./file", O_RDWR);
@@ -81,16 +84,33 @@ int main(int argc, char** argv) {
 		mem = mmap(0, sz_mem[SZ_INDEX], PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
 		CHKANDJUMP(mem == MAP_FAILED, 255, "mmap failed\n");
 		
-		unsigned long val = *((unsigned long*)mem);
 
 		ret = waitpid(pid, &status, 0);
 		CHKANDJUMP(ret == -1, 255, "waitpid failed\n");
+		printf("child exit status=%d\n", WEXITSTATUS(status));
+
 // for swap_test
+		// before swap
+		unsigned long val = *((unsigned long*)mem);
+		if (val == TEST_VAL) {
+			printf("[OK] before swap, val:0x%lx\n", val);
+		} else {
+			printf("[NG] before swap, val is not 0x%lx, val is 0x%lx\n", TEST_VAL, val);
+		}
+
 		swap_rc = do_swap("/tmp/rusage008_p.swp", buffer);
 		if (swap_rc < 0) {
 			printf("[NG] swap in parent is failed\n");
 		}
-		printf("exit status=%d\n", WEXITSTATUS(status));
+
+		// after swap
+		val = *((unsigned long*)mem);
+		if (val == TEST_VAL) {
+			printf("[OK] after swap,  val:0x%lx\n", val);
+		} else {
+			printf("[NG] after swap,  val is not 0x%lx, val is 0x%lx\n", TEST_VAL, val);
+		}
+
 	}
 	
  fn_exit:
diff --git a/test/qlmpi/rusage/rusage009.c b/test/qlmpi/rusage/rusage009.c
index 2f352d92..340829dc 100644
--- a/test/qlmpi/rusage/rusage009.c
+++ b/test/qlmpi/rusage/rusage009.c
@@ -56,6 +56,7 @@ int main(int argc, char** argv) {
 	key_t key = ftok(argv[0], 0);
 	int shmid;
 // for swap_test
+#define TEST_VAL 0x1234
 	int swap_rc = 0;
 	char buffer[BUF_SIZE];
 	
@@ -67,13 +68,14 @@ int main(int argc, char** argv) {
 	if(pid == 0) {
 		mem = shmat(shmid, NULL, 0);
 		CHKANDJUMP(mem == (void*)-1, 255, "shmat failed: %s\n", strerror(errno));
+		memset(mem, 0, sz_mem[SZ_INDEX]);
 		
 // for swap_test
 		swap_rc = do_swap("/tmp/rusage009_c.swp", buffer);
 		if (swap_rc < 0) {
 			printf("[NG] swap in child is failed\n");
 		}
-		*((unsigned long*)mem) = 0x1234;
+		*((unsigned long*)mem) = TEST_VAL;
 
 		ret = shmdt(mem);
 		CHKANDJUMP(ret == -1, 255, "shmdt failed\n");
@@ -87,12 +89,26 @@ int main(int argc, char** argv) {
 		CHKANDJUMP(ret == -1, 255, "waitpid failed\n");
 
 // for swap_test
+		// before swap
+		unsigned long val = *((unsigned long*)mem);
+		if (val == TEST_VAL) {
+			printf("[OK] before swap, val:0x%lx\n", val);
+		} else {
+			printf("[NG] before swap, val is not 0x%lx, val is 0x%lx\n", TEST_VAL, val);
+		}
+
 		swap_rc = do_swap("/tmp/rusage009_p.swp", buffer);
 		if (swap_rc < 0) {
 			printf("[NG] swap in parent is failed\n");
 		}
 
-		printf("%lx\n", *((unsigned long*)mem));
+		// after swap
+		val = *((unsigned long*)mem);
+		if (val == TEST_VAL) {
+			printf("[OK] after swap,  val:0x%lx\n", val);
+		} else {
+			printf("[NG] after swap,  val is not 0x%lx, val is 0x%lx\n", TEST_VAL, val);
+		}
 
 #if 0
 		struct shmid_ds buf;
diff --git a/test/qlmpi/rusage/rusage010.c b/test/qlmpi/rusage/rusage010.c
index 25b8bd78..2d1e601c 100644
--- a/test/qlmpi/rusage/rusage010.c
+++ b/test/qlmpi/rusage/rusage010.c
@@ -49,6 +49,8 @@ int main(int argc, char** argv) {
 	void* anon[NUM_AREAS];
 	int ret = 0;
 // for qlmpi test
+#define TEST_VAL 0x1234
+	void* mem;
 	int swap_rc = 0;
 	char buffer[BUF_SIZE];
 
@@ -61,9 +63,27 @@ int main(int argc, char** argv) {
 		memset(anon[i], 0, sz_anon[sz_index]);
 	}
 // for qlmpi test
+	// before swap
+	mem = anon[0];
+	*((unsigned long*)mem) = TEST_VAL;
+	unsigned long val = *((unsigned long*)mem);
+	if (val == TEST_VAL) {
+		printf("[OK] before swap, val:0x%lx\n", val);
+	} else {
+		printf("[NG] before swap, val is not correct, val:0x%lx\n", val);
+	}
+
 	swap_rc = do_swap("/tmp/rusage010.swp", buffer);
 	if (swap_rc < 0) {
-		printf("[NG] swap is failed.\n");
+		printf("[NG] swap in parent is failed\n");
+	}
+
+	// after swap
+	val = *((unsigned long*)mem);
+	if (val == TEST_VAL) {
+		printf("[OK] after swap,  val:0x%lx\n", val);
+	} else {
+		printf("[NG] after swap,  val is not correct, val:0x%lx\n", val);
 	}
 
 	for(i = 0; i < NUM_AREAS; i++) {
diff --git a/test/qlmpi/rusage/rusage011.c b/test/qlmpi/rusage/rusage011.c
index db614435..b1d87a60 100644
--- a/test/qlmpi/rusage/rusage011.c
+++ b/test/qlmpi/rusage/rusage011.c
@@ -58,6 +58,7 @@ int main(int argc, char** argv) {
 	int shmid;
 	xpmem_segid_t segid;
 // for swap_test
+#define TEST_VAL 0x1234
 	int swap_rc = 0;
 	char buffer[BUF_SIZE];
 
@@ -88,13 +89,14 @@ int main(int argc, char** argv) {
 		struct xpmem_addr addr = { .apid = apid, .offset = 0 };
 		void* attach = xpmem_attach(addr, sz_mem[SZ_INDEX], NULL);
 		CHKANDJUMP(attach == (void*)-1, 255, "xpmem_attach failed: %s\n", strerror(errno));
-		
-		*((unsigned long*)attach) = 0x1234;
+		memset(attach, 0, sz_mem[SZ_INDEX]);
+
 // for swap_test
 		swap_rc = do_swap("/tmp/rusage011_c.swp", buffer);
 		if (swap_rc < 0) {
 			printf("[NG] swap in child is failed\n");
 		}
+		*((unsigned long*)attach) = TEST_VAL;
 
 		ret = xpmem_detach(attach);
 		CHKANDJUMP(ret == -1, 255, "xpmem_detach failed\n");
@@ -116,12 +118,26 @@ int main(int argc, char** argv) {
 		CHKANDJUMP(ret == -1, 255, "waitpid failed\n");
 
 // for swap_test
+		// before swap
+		unsigned long val = *((unsigned long*)mem);
+		if (val == TEST_VAL) {
+			printf("[OK] before swap, val:0x%lx\n", val);
+		} else {
+			printf("[NG] before swap, val is not correct, val:0x%lx\n", val);
+		}
+
 		swap_rc = do_swap("/tmp/rusage011_p.swp", buffer);
 		if (swap_rc < 0) {
 			printf("[NG] swap in parent is failed\n");
 		}
 
-		printf("%lx\n", *((unsigned long*)mem));
+		// after swap
+		val = *((unsigned long*)mem);
+		if (val == TEST_VAL) {
+			printf("[OK] after swap,  val:0x%lx\n", val);
+		} else {
+			printf("[NG] after swap,  val is not 0x%lx, val is 0x%lx\n", TEST_VAL, val);
+		}
 
 		struct shmid_ds buf;
 		ret = shmctl(shmid, IPC_RMID, &buf);
diff --git a/test/qlmpi/rusage/verbs/README b/test/qlmpi/rusage/verbs/README
new file mode 100644
index 00000000..ec2cd625
--- /dev/null
+++ b/test/qlmpi/rusage/verbs/README
@@ -0,0 +1,33 @@
+=============
+Test program for swap in ib_pingpong
+=============
+Rank 0 process is sender, another is reciever.
+Sender send "MagicNumber + LoopCount" to reciever.
+Reciever check the number, and print OK or NG.
+Then, they swapout and wait for resume.
+
+When they are resumed, increment LoopCount and send/recieve again.
+
+=============
+How to run
+=============
+compile mpi_rdma_wr
+ $ make
+
+run by ql_mpiexec_start
+ $ ql_mpiexec_start -machinefile ./mfile ./mpi_rdma_wr -s `hostname` -p 9999
+will be printed
+ [OK] recv_val: 0x55aa55aa
+
+rerun by ql_mpiexec_start
+ $ ql_mpiexec_start -machinefile ./mfile ./mpi_rdma_wr -s `hostname` -p 9999
+
+if swap is disable, it will be printed
+ [OK] recv_val: 0x55aa55ab
+else (swap is enable)
+ [NG] recv_val is not correct, expected:0x55aa55ab  recv:0x55aa55aa
+   or 
+ occur abort
+
+finalize by ql_mpiexec_finalize
+ $ ql_mpiexec_finalize -machinefile ./mfile ./mpi_rdma_wr -s `hostname` -p 9999
diff --git a/test/qlmpi/rusage/verbs/makefile b/test/qlmpi/rusage/verbs/makefile
index 5371cb46..db998dc1 100644
--- a/test/qlmpi/rusage/verbs/makefile
+++ b/test/qlmpi/rusage/verbs/makefile
@@ -1,14 +1,19 @@
 VPATH = 
 
+MCK_DIR=/home/satoken/ppos
+MPICH_DIR=/usr/lib64/mpich-3.2
+
 CC = icc
+MPICC = mpicc
 CFLAGS = -Wall -O0 -g -DDEBUG -DERROR
 LD = $(CC)
 LFLAGS = -libverbs
+LDFLAGS = -L$(MPICH_DIR)/lib -L$(MCK_DIR)/lib
 
-SRCS = list.c curtime.c printmem.c debug.c post.c qp.c read_config.c resource.c rdma_wr.c 
+SRCS = list.c curtime.c printmem.c debug.c post.c qp.c read_config.c resource.c mpi_rdma_wr.c
 DSRCS = $(SRCS:.c=.d)
 OBJS = $(SRCS:.c=.o)
-EXECS = rdma_wr 
+EXECS = mpi_rdma_wr
 MODULES = list.o curtime.o printmem.o debug.o post.o qp.o read_config.o resource.o sock.o
 
 CWD := $(abspath $(dir $(lastword $(MAKEFILE_LIST))))
@@ -18,11 +23,17 @@ all: $(EXECS)
 rdma_wr: rdma_wr.o $(MODULES)
 	$(LD) -o $@ $^ $(LFLAGS)
 
+mpi_rdma_wr: mpi_rdma_wr.o $(MODULES)
+	$(LD) $(LDFLAGS) -lqlmpi -lmpi -o $@ $^ $(LFLAGS) 
+
+mpi_rdma_wr.o: mpi_rdma_wr.c
+	$(MPICC) -I$(MCK_DIR)/include $(CFLAGS) -c $<
+
 %.o: %.c
-	$(CC) $(CFLAGS) -c $<
+	$(MPICC) $(CFLAGS) $(LDFLAGS) -c $<
 
 %.d: %.c
-	$(CC) -MM $< > $*.d
+	$(MPICC) $(LDFLAGS) -MM $< > $*.d
 
 clean:
 	rm -f $(EXECS) $(OBJS) $(DSRCS)
diff --git a/test/qlmpi/rusage/verbs/mfile b/test/qlmpi/rusage/verbs/mfile
new file mode 100644
index 00000000..7aa47c68
--- /dev/null
+++ b/test/qlmpi/rusage/verbs/mfile
@@ -0,0 +1 @@
+wallaby15:2
diff --git a/test/qlmpi/rusage/verbs/mpi_rdma_wr.c b/test/qlmpi/rusage/verbs/mpi_rdma_wr.c
new file mode 100644
index 00000000..74c5fa7d
--- /dev/null
+++ b/test/qlmpi/rusage/verbs/mpi_rdma_wr.c
@@ -0,0 +1,295 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <memory.h>
+#include <sys/mman.h>
+#include <unistd.h>
+#include "ibcomm.h"
+#include "debug.h"
+#include "mtype.h"
+#include "mcons.h"
+#include "mm_ib_test.h"
+
+#include <mpi.h>
+#include <qlmpilib.h>
+
+//#define DEBUG_RDMA_WR
+#ifdef DEBUG_RDMA_WR
+#define dprintf printf
+#else
+#define dprintf(...)
+#endif
+
+#define TEST_SEND_BUF_NUM 3
+#define TEST_RDMA_FLG_SIZE (sizeof(unsigned short))
+#define NTRIAL 1 /* 120 */
+#define PPOLLS 1 /* sweet spot is around 10 */
+#define NSKIPS (PPOLLS*0)
+#define PPOLLR 1 /* sweet spot is around 10 */
+#define NSKIPR (PPOLLR*0)
+
+#define IBCOM_MAGIC 0x55aa55aa
+
+typedef struct tailmagic_t {
+    uint32_t magic;
+} tailmagic_t;
+
+enum rdma_buf_flg{
+	RDMA_BUF_RESET_FLG = 0,
+	RDMA_BUF_WRITE_FLG = 1,
+};
+
+static unsigned long rdtsc() {
+    unsigned long x;
+    __asm__ __volatile__("xorl %%eax, %%eax; cpuid;" : : : "%rax", "%rbx", "%rcx", "%rdx"); /* rdtsc cannot be executed earlier than this */
+    __asm__ __volatile__("rdtsc; shl $32, %%rdx; or %%rdx, %%rax" : "=a"(x) : : "memory"); /* rdtsc cannot be executed earlier than here */
+    __asm__ __volatile__("xorl %%eax, %%eax; cpuid;" : : : "%rax", "%rbx", "%rcx", "%rdx"); /* following instructions cannot be executed earlier than this */
+    return x;
+}
+
+volatile int k;
+int main(int argc, char **argv) {
+	config_t config;
+	unsigned long i, j;
+    int ibcom_errno = 0;
+	char sync_res;
+    unsigned long tscs, tsce;
+	resource_t res;
+	pdinfo_t pdinfo;
+	qpinfo_t qpinfo;
+	mrinfo_t *loc_mr_list = NULL;
+	int entry;
+    int ibv_errno;
+	int rc, my_rank, loop_cnt = 0;
+	uint32_t send_val = 0;
+
+	if (read_config(&config, argc, argv)) {
+		goto fn_exit;
+	}
+
+	config.use_rdma = 1;
+
+    unsigned long buf_size;
+    char*  str_env = getenv("BUF_SIZE");
+    buf_size = str_env ? atol(str_env) : 4096/*48,1073741824ULL * 1 + 4*/;
+
+	if(buf_size == 0) { printf("set buf_size"); goto fn_fail; }
+
+	if(resource_create(config, &res) || pd_create(&res, &pdinfo)) { printf("qp_create failed\n"); goto fn_fail; }
+    
+    ibv_errno = qp_create(&res, &pdinfo, &qpinfo);
+    IBCOM_ERR_CHKANDJUMP(ibv_errno, -1, printf("qp_create failed\n"));
+
+	/* create MR buffers */
+
+	// rdma-write-to buffer
+#if 1
+	void *rdma_buf = mmap(0, buf_size * NTRIAL, PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
+    memset(rdma_buf, 0, buf_size * NTRIAL);
+#else
+	void *rdma_buf = calloc(buf_size * NTRIAL, sizeof(char));
+#endif
+    if(!rdma_buf) { printf("mmap failed\n"); goto fn_fail; }
+	if(mr_create(&res, &pdinfo, buf_size * NTRIAL, rdma_buf, &res.rdma_mr)) { printf("mr_create failed\n"); goto fn_fail; }
+
+// for mpi
+	MPI_Init(&argc, &argv);
+	MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
+
+// switch server/client by rank
+	if (my_rank == 0) {
+		config.server_name = NULL;
+		config.server_flg = 1;
+	}
+
+	// local data buffers
+	loc_mr_list = calloc(sizeof(mrinfo_t) * NTRIAL, sizeof(char));
+	for (i = 0; i < NTRIAL; i++) {
+		void *loc_buf = mmap(0, buf_size, PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
+        if(loc_buf == MAP_FAILED) { printf("mmap failed\n"); goto fn_fail; }
+		if(config.server_flg) {
+            for(j = 0; j < buf_size; j++) {
+                *((unsigned char*)loc_buf + j) = (char)i;
+            }
+            *((uint32_t *)(loc_buf + buf_size - sizeof(uint32_t))) = 0 + IBCOM_MAGIC;
+		}
+
+        dprintf("magic addr=%lx\n", (unsigned long)(loc_buf + buf_size - TEST_RDMA_FLG_SIZE));
+
+		if(mr_create(&res, &pdinfo, buf_size, loc_buf, &loc_mr_list[i])) { printf("mr_create failed\n"); goto fn_fail;	}
+    }
+
+    if(!config.server_flg) { dprintf("res->rdma_mr.mr->addr=%lx\n", (unsigned long)res.rdma_mr.mr->addr); }
+	/* exchange gid, lid, qpn, raddr, rkey */
+	if(connect_qp(config, &res, &qpinfo)) { printf("connect_qp failed\n"); goto fn_fail; }
+	debug_print_qp_conn_info(res, qpinfo, &config);
+    printf("connect_qp done\n"); fflush(stdout);
+
+    if(config.server_flg) { dprintf("qpinfo->remote_conn_info[0].addr=%lx\n", qpinfo.remote_conn_info[0].addr); }
+
+	/* make qp RTS */
+	if(init_qp(config, &qpinfo) || rtr_qp(config, &qpinfo) || rts_qp(config, &qpinfo)) { printf("rts failed\n"); goto fn_fail; }
+    printf("rts done\n"); fflush(stdout);
+
+    /* barrier */
+    for(i = 0; i < (config.server_flg ? config.nremote : 1); i++) {
+        if(sock_sync_data(qpinfo.sock[i], 1, "R", &sync_res)) { perror("sock_sync_data"); }
+    }
+    printf("barrier done\n"); fflush(stdout);
+
+ql_loop:
+// set send-value
+	send_val = loop_cnt + IBCOM_MAGIC;
+	//*((uint32_t *)(loc_mr_list[0].buf + buf_size - sizeof(uint32_t))) = send_val;
+	*((uint32_t *)(loc_mr_list[0].buf)) = send_val;
+
+// send/recieve
+	if(config.server_flg) { /* sender side */
+        //usleep(500000);
+        if(NTRIAL % PPOLLS != 0) { printf("set NTRIAL multiple of PPOLLS\n"); goto fn_fail; }
+        if(NTRIAL <= NSKIPS) { printf("set NTRIAL > NSKIP\n"); goto fn_fail; }
+
+		for (i = 0; i < NTRIAL; i++) {
+            if(i == NSKIPS) { tscs = rdtsc(); }
+
+            post_send_req2(&qpinfo, &loc_mr_list[0], IBV_WR_RDMA_WRITE, &qpinfo.remote_conn_info[0], 0, i);
+
+		}
+        tsce = rdtsc(); printf("send_time,%.0f\n", (tsce-tscs)/(double)(NTRIAL-NSKIPS));
+		printf("send_val: %x\n", send_val);
+		fflush(stdout);
+#if 1
+        int nfound = 0;
+        k = 0;
+        while(1) {
+            int result;
+            struct ibv_wc cqe[NTRIAL];
+            result = ibv_poll_cq(qpinfo.scq, NTRIAL, &cqe[0]);
+            if(result < 0) { printf("ibv_poll_cq"); goto fn_fail; }
+            if(result > 0) {
+                for(j = 0; j < result; j++) { 
+                    if(cqe[j].status != IBV_WC_SUCCESS) { printf("cqe status,%s\n", ibv_wc_status_str(cqe[j].status)); goto fn_fail; }
+                }
+                //debug_print_mem((addr_t)loc_mr_list[entry].buf, buf_size);
+                nfound += result;
+                if(nfound >= NTRIAL) { break; }
+            }
+            k++;
+        }
+#endif
+	} else { /* receiver side */
+		//volatile uint32_t *recv_val = NULL;
+        if(NSKIPR % PPOLLR !=0) { printf("set NSKIP multiple of PPOLL\n"); goto fn_fail; }
+		for (i = 0; i < NTRIAL; i++) {
+            if(i == NSKIPR) { tscs = rdtsc(); }
+
+			// poll on magic
+            dprintf("res.rdma_mr.buf=%lx\n", (unsigned long)res.rdma_mr.buf);
+            dprintf("poll addr=%lx\n", (unsigned long)(rdma_buf + buf_size * i + buf_size - sizeof(uint32_t)));
+            //k = 0;
+            //volatile uint32_t *ptr = (volatile uint32_t *)(rdma_buf + buf_size * i + buf_size - sizeof(uint32_t));
+            volatile uint32_t *ptr = (volatile uint32_t *)(rdma_buf);
+            //*recv_val = *(volatile uint32_t *)(rdma_buf + buf_size * i + buf_size - sizeof(uint32_t));
+			printf("*** ptr:%d, MAGIC:%x\n", *ptr, loop_cnt + IBCOM_MAGIC);
+            while(*ptr != loop_cnt + IBCOM_MAGIC) {
+			//printf("++++ send_val:%x, recv_val:%x\n", send_val, *recv_val);
+            //while(*recv_val != send_val) {
+                //k++; if(i >= NSKIPR && k % 65536 == 65535) { printf("i=%d,poll value=%x\n", i, *((uint32_t *)(rdma_buf + buf_size * i + buf_size - sizeof(uint32_t)))); }
+                __asm__ __volatile__("pause");
+				// TIMEOUT
+				if (rdtsc() - tscs > 10000000) {
+					printf("*** recv wait is TIMEOUT\n");
+					break;
+				}
+            }
+			if (*ptr == loop_cnt + IBCOM_MAGIC) {
+				printf("[OK] recv_val: 0x%x\n", *ptr);
+			} else {
+				printf("[NG] recv_val is not correct, expected:0x%x  recv:0x%x\n",
+				       loop_cnt + IBCOM_MAGIC, *ptr);
+			}
+				
+		}
+        tsce = rdtsc(); printf("recv_time,%.0f\n", (tsce-tscs)/(double)(NTRIAL-NSKIPR));
+		//printf("recv_valu:%x\n", *recv_val);
+	}
+
+	rc = ql_client(&argc, &argv);
+	if (rc == QL_CONTINUE) {
+		loop_cnt++;
+		printf("ql: go back loop\n");
+		goto ql_loop;
+	}
+	else {
+		printf("ql_client returns %d, go to exit\n", rc);
+	}
+
+#if 0
+// 2nd send/recieve
+	if(config.server_flg) { /* sender side */
+		// other value
+		*((uint32_t *)(loc_mr_list[0].buf + buf_size - sizeof(uint32_t))) = 10 + IBCOM_MAGIC;
+        //usleep(500000);
+        if(NTRIAL % PPOLLS != 0) { printf("set NTRIAL multiple of PPOLLS\n"); goto fn_fail; }
+        if(NTRIAL <= NSKIPS) { printf("set NTRIAL > NSKIP\n"); goto fn_fail; }
+
+		for (i = 0; i < NTRIAL; i++) {
+            if(i == NSKIPS) { tscs = rdtsc(); }
+
+            post_send_req2(&qpinfo, &loc_mr_list[0], IBV_WR_RDMA_WRITE, &qpinfo.remote_conn_info[0], 0, i);
+
+		}
+#if 1
+        int nfound = 0;
+        k = 0;
+        while(1) {
+            int result;
+            struct ibv_wc cqe[NTRIAL];
+            result = ibv_poll_cq(qpinfo.scq, NTRIAL, &cqe[0]);
+            if(result < 0) { printf("ibv_poll_cq"); goto fn_fail; }
+            if(result > 0) {
+                for(j = 0; j < result; j++) { 
+                    if(cqe[j].status != IBV_WC_SUCCESS) { printf("cqe status,%s\n", ibv_wc_status_str(cqe[j].status)); goto fn_fail; }
+                }
+                //debug_print_mem((addr_t)loc_mr_list[entry].buf, buf_size);
+                nfound += result;
+                if(nfound >= NTRIAL) { break; }
+            }
+            k++;
+        }
+		printf("2nd send is OK!!\n");
+#endif
+	} else { /* receiver side */
+        if(NSKIPR % PPOLLR !=0) { printf("set NSKIP multiple of PPOLL\n"); goto fn_fail; }
+		for (i = 0; i < NTRIAL; i++) {
+            if(i == NSKIPR) { tscs = rdtsc(); }
+
+			// poll on magic
+            dprintf("res.rdma_mr.buf=%lx\n", (unsigned long)res.rdma_mr.buf);
+            dprintf("poll addr=%lx\n", (unsigned long)(rdma_buf + buf_size * i + buf_size - sizeof(uint32_t)));
+            //k = 0;
+            volatile uint32_t *ptr = (volatile uint32_t *)(rdma_buf + buf_size * i + buf_size - sizeof(uint32_t));
+            while(*ptr != 10 + IBCOM_MAGIC) {
+                //k++; if(i >= NSKIPR && k % 65536 == 65535) { printf("i=%d,poll value=%x\n", i, *((uint32_t *)(rdma_buf + buf_size * i + buf_size - sizeof(uint32_t)))); }
+                __asm__ __volatile__("pause");
+            }
+			//debug_print_mem((addr_t)res.rdma_mr.buf, buf_size);
+		}
+		printf("2nd revieve is OK!!\n");
+	}
+#endif
+
+	fn_exit:
+	/*Can free all resources*/
+#if 0
+	if (resource_destroy(&config, &res)) {
+		fprintf(stderr, "resource destroy failed\n");
+	} else {
+		dprintf("destroy all successfully..\n");
+	}
+	if(loc_mr_list) { free(loc_mr_list); }
+#endif
+	MPI_Finalize();
+	return ibcom_errno;
+ fn_fail:
+    goto fn_exit;
+}
diff --git a/test/qlmpi/rusage/verbs/qp.c b/test/qlmpi/rusage/verbs/qp.c
index f766b956..f77881ef 100644
--- a/test/qlmpi/rusage/verbs/qp.c
+++ b/test/qlmpi/rusage/verbs/qp.c
@@ -8,7 +8,7 @@
 #include "sock.h"
 #include "debug.h"
 
-#define DEBUG_QP
+//#define DEBUG_QP
 #ifdef DEBUG_QP
 #define dprintf printf
 #else
diff --git a/test/qlmpi/rusage/verbs/sock.c b/test/qlmpi/rusage/verbs/sock.c
index 847859e8..92c4e089 100644
--- a/test/qlmpi/rusage/verbs/sock.c
+++ b/test/qlmpi/rusage/verbs/sock.c
@@ -109,6 +109,7 @@ int sock_connect(char *server_name, int port, int *listenfd){
 
 		// client mode
 		}else{
+			sleep(1);
             inet_ntop(rp->ai_family, rp->ai_addr->sa_data, addrstr, 100);
             void *ptr;
             switch(rp->ai_family) {