diff --git a/arch/x86_64/kernel/perfctr.c b/arch/x86_64/kernel/perfctr.c
index ff05656b..4f33939d 100644
--- a/arch/x86_64/kernel/perfctr.c
+++ b/arch/x86_64/kernel/perfctr.c
@@ -10,9 +10,12 @@
 #include <ihk/perfctr.h>
 #include <march.h>
 #include <errno.h>
+#include <cls.h>
 #include <ihk/debug.h>
+#include <ihk/cpu.h>
 #include <registers.h>
 #include <mc_perf_event.h>
+#include <config.h>
 
 extern unsigned int *x86_march_perfmap;
 extern int running_on_kvm(void);
@@ -57,6 +60,10 @@ void x86_init_perfctr(void)
 	uint64_t ecx;
 	uint64_t edx;
 
+#ifndef ENABLE_PERF
+	return;
+#endif //ENABLE_PERF
+
 	/* Do not do it on KVM */
 	if (running_on_kvm()) return;
 
@@ -93,7 +100,7 @@ void x86_init_perfctr(void)
 	for(i = 0; i < X86_IA32_NUM_PERF_COUNTERS; i++) {
 		wrmsr(MSR_IA32_PERFEVTSEL0 + i, 0);
 	}
-		
+
 	/* Enable PMC Control */
 	value = rdmsr(MSR_PERF_GLOBAL_CTRL);
 	value |= X86_IA32_PERF_COUNTERS_MASK;
@@ -254,6 +261,41 @@ int ihk_mc_perfctr_init(int counter, enum ihk_perfctr_type type, int mode)
 	return set_perfctr_x86_direct(counter, mode, x86_march_perfmap[type]);
 }
 
+int ihk_mc_perfctr_set_extra(struct mc_perf_event *event)
+{
+	struct thread *thread = cpu_local_var(current);
+
+	// allocate extra_reg
+	if (thread->extra_reg_alloc_map & (1UL << event->extra_reg.idx)) {
+		if (event->extra_reg.idx == EXTRA_REG_RSP_0) {
+			event->extra_reg.idx = EXTRA_REG_RSP_1;
+		}
+		else if (event->extra_reg.idx == EXTRA_REG_RSP_1) {
+			event->extra_reg.idx = EXTRA_REG_RSP_0;
+		}
+
+		if (thread->extra_reg_alloc_map & (1UL << event->extra_reg.idx)) {
+			// extra_regs are full
+			return -1;
+		}
+	}
+
+	if (event->extra_reg.idx == EXTRA_REG_RSP_0) {
+		event->hw_config &= ~0xffUL;
+		event->hw_config |= ihk_mc_get_extra_reg_event(EXTRA_REG_RSP_0);
+		event->extra_reg.reg = MSR_OFFCORE_RSP_0;
+	}
+	else if (event->extra_reg.idx == EXTRA_REG_RSP_1) {
+		event->hw_config &= ~0xffUL;
+		event->hw_config |= ihk_mc_get_extra_reg_event(EXTRA_REG_RSP_1);
+		event->extra_reg.reg = MSR_OFFCORE_RSP_1;
+	}
+		
+	thread->extra_reg_alloc_map |= (1UL << event->extra_reg.idx);
+	wrmsr(event->extra_reg.reg, event->extra_reg.config);
+	return 0;
+}
+
 #ifdef HAVE_MARCH_PERFCTR_START
 extern void x86_march_perfctr_start(unsigned long counter_mask);
 #endif
diff --git a/config.h.in b/config.h.in
index e3721b6c..1c6bbdf1 100644
--- a/config.h.in
+++ b/config.h.in
@@ -9,12 +9,15 @@
 /* whether memdump feature is enabled */
 #undef ENABLE_MEMDUMP
 
-/* whether mcoverlayfs is enabled */
+/* whether qlmpi is enabled */
 #undef ENABLE_QLMPI
 
 /* whether rusage is enabled */
 #undef ENABLE_RUSAGE
 
+/* whether perf is enabled */
+#undef ENABLE_PERF
+
 /* Define to 1 if you have the <inttypes.h> header file. */
 #undef HAVE_INTTYPES_H
 
diff --git a/configure b/configure
index a689604b..a5f44367 100755
--- a/configure
+++ b/configure
@@ -708,6 +708,7 @@ enable_dcfa
 enable_memdump
 enable_mcoverlayfs
 enable_rusage
+enable_perf
 enable_qlmpi
 with_uname_r
 '
@@ -1333,6 +1334,7 @@ Optional Features:
   --enable-memdump        enable dumping memory and analyzing a dump
   --enable-mcoverlayfs    enable mcoverlayfs implementation
   --enable-rusage         enable rusage implementation
+  --enable-perf           enable perf_event implementation
   --enable-qlmpi          enable qlmpi implementation
 
 Optional Packages:
@@ -3568,6 +3570,14 @@ else
 fi
 
 
+# Check whether --enable-perf was given.
+if test "${enable_perf+set}" = set; then :
+  enableval=$enable_perf; ENABLE_PERF=$enableval
+else
+  ENABLE_PERF=yes
+fi
+
+
 # Check whether --enable-qlmpi was given.
 if test "${enable_qlmpi+set}" = set; then :
   enableval=$enable_qlmpi; ENABLE_QLMPI=$enableval
@@ -4964,6 +4974,17 @@ else
 $as_echo "$as_me: rusage is disabled" >&6;}
 fi
 
+if test "x$ENABLE_PERF" = "xyes" ; then
+
+$as_echo "#define ENABLE_PERF 1" >>confdefs.h
+
+	{ $as_echo "$as_me:${as_lineno-$LINENO}: perf is enabled" >&5
+$as_echo "$as_me: perf is enabled" >&6;}
+else
+	{ $as_echo "$as_me:${as_lineno-$LINENO}: perf is disabled" >&5
+$as_echo "$as_me: perf is disabled" >&6;}
+fi
+
 if test "x$MCKERNEL_INCDIR" != "x" ; then
 
 cat >>confdefs.h <<_ACEOF
diff --git a/configure.ac b/configure.ac
index b1a7d2d8..669ea1c3 100644
--- a/configure.ac
+++ b/configure.ac
@@ -134,6 +134,12 @@ AC_ARG_ENABLE([rusage],
 	      [ENABLE_RUSAGE=$enableval],
 	      [ENABLE_RUSAGE=yes])
 
+AC_ARG_ENABLE([perf],
+	      AC_HELP_STRING([--enable-perf],
+			     [enable perf_event implementation]),
+	      [ENABLE_PERF=$enableval],
+	      [ENABLE_PERF=yes])
+
 AC_ARG_ENABLE([qlmpi],
 	      AC_HELP_STRING([--enable-qlmpi],
 			     [enable qlmpi implementation]),
@@ -454,7 +460,7 @@ else
 fi
 
 if test "x$ENABLE_QLMPI" = "xyes" ; then
-	AC_DEFINE([ENABLE_QLMPI],[1],[whether mcoverlayfs is enabled])
+	AC_DEFINE([ENABLE_QLMPI],[1],[whether qlmpi is enabled])
 	AC_MSG_NOTICE([qlmpi is enabled])
 else
 	AC_MSG_NOTICE([qlmpi is disabled])
@@ -478,6 +484,13 @@ else
 	AC_MSG_NOTICE([rusage is disabled])
 fi
 
+if test "x$ENABLE_PERF" = "xyes" ; then
+	AC_DEFINE([ENABLE_PERF],[1],[whether perf is enabled])
+	AC_MSG_NOTICE([perf is enabled])
+else
+	AC_MSG_NOTICE([perf is disabled])
+fi
+
 if test "x$MCKERNEL_INCDIR" != "x" ; then
    AC_DEFINE_UNQUOTED(MCKERNEL_INCDIR,"$MCKERNEL_INCDIR",[McKernel specific headers])
 fi
diff --git a/kernel/include/process.h b/kernel/include/process.h
index 1b7dbe56..5f81d269 100644
--- a/kernel/include/process.h
+++ b/kernel/include/process.h
@@ -26,6 +26,7 @@
 #include <syscall.h>
 #include <bitops.h>
 #include <profile.h>
+#include <config.h>
 
 #define VR_NONE            0x0
 #define VR_STACK           0x1
@@ -691,6 +692,10 @@ struct thread {
 	int mod_clone;
 	struct uti_attr *mod_clone_arg;
 	int parent_cpuid;
+
+	// for performance counter
+	unsigned long pmc_alloc_map;
+	unsigned long extra_reg_alloc_map;
 };
 
 #define VM_RANGE_CACHE_SIZE	4
diff --git a/kernel/syscall.c b/kernel/syscall.c
index b4c5735c..d41fe087 100644
--- a/kernel/syscall.c
+++ b/kernel/syscall.c
@@ -3436,18 +3436,18 @@ SYSCALL_DECLARE(signalfd4)
 }
 
 int 
-perf_counter_alloc(struct mc_perf_event *event)
+perf_counter_alloc(struct thread *thread)
 {
-	int ret = 0;
-	struct perf_event_attr *attr = &event->attr;
-	struct mc_perf_event *leader = event->group_leader;
+	int ret = -1;
+	int i = 0;
 
-	ret = ihk_mc_perfctr_alloc_counter(&attr->type, &attr->config, leader->pmc_status); 
-
-	if(ret >= 0) {
-		leader->pmc_status |= 1UL << ret;
+	// find avail generic counter
+    for(i = 0; i < X86_IA32_NUM_PERF_COUNTERS; i++) {
+		if(!(thread->pmc_alloc_map & (1 << i))) {
+			ret = i;
+			break;
+		}
 	}
-	event->counter_id = ret;
 
 	return ret;
 }
@@ -3467,7 +3467,13 @@ perf_counter_start(struct mc_perf_event *event)
 	}
 
 	if(event->counter_id >= 0 && event->counter_id < X86_IA32_NUM_PERF_COUNTERS) {
-		ret = ihk_mc_perfctr_init_raw(event->counter_id, attr->config, mode);
+		if (event->extra_reg.reg) {
+			if (ihk_mc_perfctr_set_extra(event)) {
+				ret = -1;
+				goto out;
+			}
+		}
+		ret = ihk_mc_perfctr_init_raw(event->counter_id, event->hw_config, mode);
 		ihk_mc_perfctr_start(1UL << event->counter_id);
 	}
 	else if(event->counter_id >= X86_IA32_BASE_FIXED_PERF_COUNTERS &&
@@ -3478,7 +3484,8 @@ perf_counter_start(struct mc_perf_event *event)
 	else {
 		ret = -1;
 	}
-		
+
+out:
 	return ret;
 }
 
@@ -3569,18 +3576,18 @@ perf_event_read_group(struct mc_perf_event *event, unsigned long read_format, ch
 static int
 perf_event_read_one(struct mc_perf_event *event, unsigned long read_format, char *buf)
 {
-        unsigned long values[4];
-        int n = 0;
+	unsigned long values[4];
+	int n = 0;
 	int size = 0;
 
-        values[n++] = perf_event_read_value(event);
+	values[n++] = perf_event_read_value(event);
 
 	size = n * sizeof(unsigned long);
 
 	if (copy_to_user(buf, values, size))
 		return -EFAULT;
 
-        return size;
+	return size;
 }
 
 static long
@@ -3597,7 +3604,6 @@ perf_read(struct mckfd *sfd, ihk_mc_user_context_t *ctx)
 		ret = perf_event_read_one(event, read_format, buf);
 	}
 	return ret;
-
 }
 
 void 
@@ -3723,12 +3729,12 @@ perf_ioctl(struct mckfd *sfd, ihk_mc_user_context_t *ctx)
 			process_unlock(proc, &lock);
 		}
 		*/
-                break;
+		break;
         case PERF_EVENT_IOC_RESET:
 		// TODO: reset other process
 		ihk_mc_perfctr_set(counter_id, event->attr.sample_freq * -1);
 		event->count = 0L;
-                break;
+		break;
         case PERF_EVENT_IOC_REFRESH:
 		// TODO: refresh other process
 		
@@ -3753,7 +3759,13 @@ static int
 perf_close(struct mckfd *sfd, ihk_mc_user_context_t *ctx)
 {
 	struct mc_perf_event *event = (struct mc_perf_event*)sfd->data;
+	struct thread *thread = cpu_local_var(current);
 
+	thread->pmc_alloc_map &= ~(1UL << event->counter_id);
+	if (event->extra_reg.reg) {
+		thread->extra_reg_alloc_map &= ~(1UL << event->extra_reg.idx);
+	}
+		
 	kfree(event);
 
 	return 0;
@@ -3805,6 +3817,66 @@ perf_mmap(struct mckfd *sfd, ihk_mc_user_context_t *ctx)
 	return rc;
 }
 
+struct mc_perf_event*
+mc_perf_event_alloc(struct perf_event_attr *attr)
+{
+	unsigned long val = 0, extra_config = 0;
+	struct mc_perf_event *event;
+	int ereg_id;
+
+	if (!attr) {
+		return NULL;
+	}
+
+	event = kmalloc(sizeof(struct mc_perf_event), IHK_MC_AP_NOWAIT);
+	if (!event) {
+		return NULL;
+	}
+	memset(event, 0, sizeof(struct mc_perf_event));
+
+	INIT_LIST_HEAD(&event->group_entry);
+	INIT_LIST_HEAD(&event->sibling_list);
+	event->attr = *attr;
+
+	event->sample_freq = attr->sample_freq;
+	event->nr_siblings = 0;
+	event->count = 0L;
+	event->child_count_total = 0;
+	event->parent = NULL;
+
+	switch(attr->type) {
+	case PERF_TYPE_HARDWARE :
+		val = ihk_mc_hw_event_map(attr->config);
+		break;
+	case PERF_TYPE_HW_CACHE :
+		val = ihk_mc_hw_cache_event_map(attr->config);
+		extra_config = ihk_mc_hw_cache_extra_reg_map(attr->config);
+		break;
+	case PERF_TYPE_RAW :
+		val = attr->config;
+		break;
+
+	default:
+		// Unexpected type
+		return NULL;
+	}
+
+	if (val == 0) {
+		return NULL;
+	}
+	
+	event->hw_config = val;
+	event->hw_config_ext = extra_config;
+
+	ereg_id = ihk_mc_get_extra_reg_id(event->hw_config, event->hw_config_ext);
+	if (ereg_id >= 0) {
+		event->extra_reg.config = event->hw_config_ext;
+		event->extra_reg.reg = ihk_mc_get_extra_reg_msr(ereg_id);
+		event->extra_reg.idx = ihk_mc_get_extra_reg_idx(ereg_id);
+	}
+	return event;
+}
+
 SYSCALL_DECLARE(perf_event_open)
 {
 	struct syscall_request request IHK_DMA_ALIGN;
@@ -3812,6 +3884,7 @@ SYSCALL_DECLARE(perf_event_open)
 	struct process *proc = thread->proc;
 	struct mckfd *sfd, *cfd;
 	int fd;
+	int counter_idx;
 	long irqstate;
 	struct perf_event_attr *attr = (void *)ihk_mc_syscall_arg0(ctx);
 	int pid = ihk_mc_syscall_arg1(ctx);
@@ -3822,47 +3895,54 @@ SYSCALL_DECLARE(perf_event_open)
 
 	int not_supported_flag = 0;
 
+#ifndef ENABLE_PERF
+	return -ENOSYS;
+#endif // ENABLE_PERF
+
 	// check Not supported 
-	if(cpu > 0) {
+	if (cpu > 0) {
 		not_supported_flag = 1;	
 	}
-	if(flags > 0) {
+	if (flags > 0) {
 		not_supported_flag = 1;	
 	}
-	if(attr->read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
+
+	if ((attr->type != PERF_TYPE_RAW) && 
+	    (attr->type != PERF_TYPE_HARDWARE) &&
+	    (attr->type != PERF_TYPE_HW_CACHE)) {
 		not_supported_flag = 1;
 	}
-	if(attr->read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
-		not_supported_flag = 1;
-	}
-	if(attr->read_format & PERF_FORMAT_ID) {
+	if (attr->read_format & 
+	    (PERF_FORMAT_TOTAL_TIME_ENABLED |
+	     PERF_FORMAT_TOTAL_TIME_RUNNING |
+	     PERF_FORMAT_ID)) {
 		not_supported_flag = 1;
 	}
 
-	if(not_supported_flag) {
+	if (not_supported_flag) {
 		return -1;
 	}
 
-	// process of perf_event_open
-	event = kmalloc(sizeof(struct mc_perf_event), IHK_MC_AP_NOWAIT);
-	if(!event)
-		return -ENOMEM;
-	event->attr = (struct perf_event_attr)*attr;
+	event = mc_perf_event_alloc((struct perf_event_attr*)attr);
+	if (!event) {
+		return -1;
+	}
 
-	event->sample_freq = attr->sample_freq;
-	event->nr_siblings = 0;
-	event->count = 0L;
-	event->child_count_total = 0;
-	event->parent = NULL;
 	event->pid = pid;
-	INIT_LIST_HEAD(&event->group_entry);
-	INIT_LIST_HEAD(&event->sibling_list);
-	if(group_fd == -1) {
+
+	counter_idx = perf_counter_alloc(thread);
+	if (counter_idx < 0) {
+		return -1;
+	}
+	event->counter_id = counter_idx;
+
+	if (group_fd == -1) {
 		event->group_leader = event; 
 		event->pmc_status = 0x0UL;
-	} else {
-		for(cfd = proc->mckfd; cfd; cfd = cfd->next) {
-			if(cfd->fd == group_fd) {
+	}
+	else {
+		for (cfd = proc->mckfd; cfd; cfd = cfd->next) {
+			if (cfd->fd == group_fd) {
 				event->group_leader = (struct mc_perf_event*)cfd->data;
 				list_add_tail(&event->group_entry, &event->group_leader->sibling_list);
 				event->group_leader->nr_siblings++;
@@ -3871,10 +3951,7 @@ SYSCALL_DECLARE(perf_event_open)
 		}
 	}
 
-	if(perf_counter_alloc(event) < 0)
-		return -1;
-	if(event->counter_id < 0)
-		return -1;
+	event->group_leader->pmc_status |= (1UL << counter_idx);
 
 	request.number = __NR_perf_event_open;
 	request.args[0] = 0;
@@ -3883,6 +3960,8 @@ SYSCALL_DECLARE(perf_event_open)
 		return fd;
 	} 
 
+	thread->pmc_alloc_map |= 1UL << counter_idx;
+
 	sfd = kmalloc(sizeof(struct mckfd), IHK_MC_AP_NOWAIT);
 	if(!sfd)
 		return -ENOMEM;
diff --git a/lib/include/ihk/cpu.h b/lib/include/ihk/cpu.h
index 7a6e2431..4e7792f7 100644
--- a/lib/include/ihk/cpu.h
+++ b/lib/include/ihk/cpu.h
@@ -77,6 +77,15 @@ void ihk_mc_init_context(ihk_mc_kernel_context_t *new_ctx,
                          void *stack_pointer,
                          void (*next_function)(void));
 
+int ihk_mc_get_extra_reg_id(unsigned long hw_config, unsigned long hw_config_ext);
+unsigned int ihk_mc_get_nr_extra_regs();
+int ihk_mc_get_extra_reg_idx(int id);
+unsigned int ihk_mc_get_extra_reg_msr(int id);
+unsigned long ihk_mc_get_extra_reg_event(int id);
+unsigned long ihk_mc_hw_event_map(unsigned long hw_event);
+unsigned long ihk_mc_hw_cache_event_map(unsigned long hw_cache_event);
+unsigned long ihk_mc_hw_cache_extra_reg_map(unsigned long hw_cache_event);
+
 /* returns the 'prev' argument of the call that caused the switch to the context returned. */
 void *ihk_mc_switch_context(ihk_mc_kernel_context_t *old_ctx,
                            ihk_mc_kernel_context_t *new_ctx,
diff --git a/lib/include/ihk/perfctr.h b/lib/include/ihk/perfctr.h
index f63618ec..be000355 100644
--- a/lib/include/ihk/perfctr.h
+++ b/lib/include/ihk/perfctr.h
@@ -18,6 +18,8 @@
 #include <types.h>
 #endif /*POSTK_DEBUG_TEMP_FIX_29*/
 
+#include <mc_perf_event.h>
+
 #define PERFCTR_USER_MODE   0x01
 #define PERFCTR_KERNEL_MODE 0x02
 
@@ -60,6 +62,7 @@ int ihk_mc_perfctr_init_raw(int counter, uint64_t config, int mode);
 int ihk_mc_perfctr_init(int counter, enum ihk_perfctr_type type, int mode);
 int ihk_mc_perfctr_init_raw(int counter, unsigned int code, int mode);
 #endif/*POSTK_DEBUG_TEMP_FIX_29*/
+int ihk_mc_perfctr_set_extra(struct mc_perf_event *event);
 #ifdef POSTK_DEBUG_TEMP_FIX_30
 int ihk_mc_perfctr_start(int counter);
 int ihk_mc_perfctr_stop(int counter);
diff --git a/lib/include/mc_perf_event.h b/lib/include/mc_perf_event.h
index b70d5ddf..da62519c 100644
--- a/lib/include/mc_perf_event.h
+++ b/lib/include/mc_perf_event.h
@@ -2,9 +2,9 @@
 #ifndef MC_PERF_EVENT_H
 #define MC_PERF_EVENT_H
 
-#ifdef POSTK_DEBUG_TEMP_FIX_32
+//#ifdef POSTK_DEBUG_TEMP_FIX_32
 #include <list.h>
-#endif /*POSTK_DEBUG_TEMP_FIX_32*/
+//#endif /*POSTK_DEBUG_TEMP_FIX_32*/
 #include <march.h>
 
 struct perf_event_attr;
@@ -105,6 +105,40 @@ enum perf_hw_id {
         PERF_COUNT_HW_MAX,                      /* non-ABI */
 };
 
+/*
+ * Generalized hardware cache events:
+ *
+ *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
+ *       { read, write, prefetch } x
+ *       { accesses, misses }
+ */
+enum perf_hw_cache_id {
+	PERF_COUNT_HW_CACHE_L1D         = 0,
+	PERF_COUNT_HW_CACHE_L1I         = 1,
+	PERF_COUNT_HW_CACHE_LL          = 2,
+	PERF_COUNT_HW_CACHE_DTLB        = 3,
+	PERF_COUNT_HW_CACHE_ITLB        = 4,
+	PERF_COUNT_HW_CACHE_BPU         = 5,
+	PERF_COUNT_HW_CACHE_NODE        = 6,
+
+	PERF_COUNT_HW_CACHE_MAX,        /* non-ABI */
+};
+
+enum perf_hw_cache_op_id {
+	PERF_COUNT_HW_CACHE_OP_READ     = 0,
+	PERF_COUNT_HW_CACHE_OP_WRITE        = 1,
+	PERF_COUNT_HW_CACHE_OP_PREFETCH     = 2,
+
+	PERF_COUNT_HW_CACHE_OP_MAX,     /* non-ABI */
+};
+
+enum perf_hw_cache_op_result_id {
+	PERF_COUNT_HW_CACHE_RESULT_ACCESS   = 0,
+	PERF_COUNT_HW_CACHE_RESULT_MISS     = 1,
+
+	PERF_COUNT_HW_CACHE_RESULT_MAX,     /* non-ABI */
+};
+
 /*
  * Hardware event_id to monitor via a performance monitoring event:
  */
@@ -218,8 +252,18 @@ struct perf_event_attr {
 #endif
 };
 
+struct hw_perf_event_extra {
+	unsigned long config;
+	unsigned int reg;
+	int idx;
+};
+
+
 struct mc_perf_event {
 	struct perf_event_attr		attr;
+	struct hw_perf_event_extra extra_reg;
+	unsigned long		hw_config;
+	unsigned long		hw_config_ext;
 	int 				cpu_id;
 	int 				counter_id;	// counter_id
 	unsigned long			count;		// counter_value
diff --git a/test/perf_event/Makefile b/test/perf_event/Makefile
new file mode 100644
index 00000000..b9347dae
--- /dev/null
+++ b/test/perf_event/Makefile
@@ -0,0 +1,3 @@
+all: perf_test
+perf_test: perf_test.o perftool.o
+perftool.o: perftool.c perftool.h
diff --git a/test/perf_event/go_perf_test.sh b/test/perf_event/go_perf_test.sh
new file mode 100755
index 00000000..01f79b80
--- /dev/null
+++ b/test/perf_event/go_perf_test.sh
@@ -0,0 +1,66 @@
+#!/bin/sh
+
+MCEXEC="mcexec"
+
+PERF_HW_ID_MAX=9
+PERF_COUNT_HW_CACHE_MAX=6
+PERF_COUNT_HW_CACHE_OP_MAX=2
+PERF_COUNT_HW_CACHE_RESULT_MAX=1
+
+PERF_TYPE_HARDWARE=0
+PERF_TYPE_HW_CACHE=3
+
+echo "【PERF_TYPE_HARDWARE all spase】, "
+for id in `seq 0 ${PERF_HW_ID_MAX}`
+do
+	${MCEXEC} ./perf_test 0 1 ${PERF_TYPE_HARDWARE} ${id}
+done
+
+echo "【HW_CACHE no exclude】, "
+for i in `seq 0 ${PERF_COUNT_HW_CACHE_MAX}`
+do
+	for j in `seq 0 ${PERF_COUNT_HW_CACHE_OP_MAX}`
+	do
+		for k in `seq 0 ${PERF_COUNT_HW_CACHE_RESULT_MAX}`
+		do
+			${MCEXEC} ./perf_test 0 1 ${PERF_TYPE_HW_CACHE} `expr ${k} \* 65536 + ${j} \* 256 + ${i}`
+		done
+	done
+done
+
+echo "【HARDWARE exclude user space】, "
+for id in `seq 0 ${PERF_HW_ID_MAX}`
+do
+	${MCEXEC} ./perf_test 1 1 ${PERF_TYPE_HARDWARE} ${id}
+done
+
+echo "【HW_CACHE exclude user space】, "
+for i in `seq 0 ${PERF_COUNT_HW_CACHE_MAX}`
+do
+	for j in `seq 0 ${PERF_COUNT_HW_CACHE_OP_MAX}`
+	do
+		for k in `seq 0 ${PERF_COUNT_HW_CACHE_RESULT_MAX}`
+		do
+			${MCEXEC} ./perf_test 1 1 ${PERF_TYPE_HW_CACHE} `expr ${k} \* 65536 + ${j} \* 256 + ${i}`
+		done
+	done
+done
+
+echo "【HARDWARE exclude kernel space】, "
+for id in `seq 0 ${PERF_HW_ID_MAX}`
+do
+	${MCEXEC} ./perf_test 2 1 ${PERF_TYPE_HARDWARE} ${id}
+done
+
+echo "【HW_CACHE exclude kernel space】, "
+for i in `seq 0 ${PERF_COUNT_HW_CACHE_MAX}`
+do
+	for j in `seq 0 ${PERF_COUNT_HW_CACHE_OP_MAX}`
+	do
+		for k in `seq 0 ${PERF_COUNT_HW_CACHE_RESULT_MAX}`
+		do
+			${MCEXEC} ./perf_test 2 1 ${PERF_TYPE_HW_CACHE} `expr ${k} \* 65536 + ${j} \* 256 + ${i}`
+		done
+	done
+done
+
diff --git a/test/perf_event/no_mc_go.sh b/test/perf_event/no_mc_go.sh
new file mode 100755
index 00000000..24e5fb24
--- /dev/null
+++ b/test/perf_event/no_mc_go.sh
@@ -0,0 +1,66 @@
+#!/bin/sh
+
+MCEXEC=""
+
+PERF_HW_ID_MAX=9
+PERF_COUNT_HW_CACHE_MAX=6
+PERF_COUNT_HW_CACHE_OP_MAX=2
+PERF_COUNT_HW_CACHE_RESULT_MAX=1
+
+PERF_TYPE_HARDWARE=0
+PERF_TYPE_HW_CACHE=3
+
+echo "【PERF_TYPE_HARDWARE all spase】, "
+for id in `seq 0 ${PERF_HW_ID_MAX}`
+do
+	${MCEXEC} ./perf_test 0 1 ${PERF_TYPE_HARDWARE} ${id}
+done
+
+echo "【HW_CACHE no exclude】, "
+for i in `seq 0 ${PERF_COUNT_HW_CACHE_MAX}`
+do
+	for j in `seq 0 ${PERF_COUNT_HW_CACHE_OP_MAX}`
+	do
+		for k in `seq 0 ${PERF_COUNT_HW_CACHE_RESULT_MAX}`
+		do
+			${MCEXEC} ./perf_test 0 1 ${PERF_TYPE_HW_CACHE} `expr ${k} \* 65536 + ${j} \* 256 + ${i}`
+		done
+	done
+done
+
+echo "【HARDWARE exclude user space】, "
+for id in `seq 0 ${PERF_HW_ID_MAX}`
+do
+	${MCEXEC} ./perf_test 1 1 ${PERF_TYPE_HARDWARE} ${id}
+done
+
+echo "【HW_CACHE exclude user space】, "
+for i in `seq 0 ${PERF_COUNT_HW_CACHE_MAX}`
+do
+	for j in `seq 0 ${PERF_COUNT_HW_CACHE_OP_MAX}`
+	do
+		for k in `seq 0 ${PERF_COUNT_HW_CACHE_RESULT_MAX}`
+		do
+			${MCEXEC} ./perf_test 1 1 ${PERF_TYPE_HW_CACHE} `expr ${k} \* 65536 + ${j} \* 256 + ${i}`
+		done
+	done
+done
+
+echo "【HARDWARE exclude kernel space】, "
+for id in `seq 0 ${PERF_HW_ID_MAX}`
+do
+	${MCEXEC} ./perf_test 2 1 ${PERF_TYPE_HARDWARE} ${id}
+done
+
+echo "【HW_CACHE exclude kernel space】, "
+for i in `seq 0 ${PERF_COUNT_HW_CACHE_MAX}`
+do
+	for j in `seq 0 ${PERF_COUNT_HW_CACHE_OP_MAX}`
+	do
+		for k in `seq 0 ${PERF_COUNT_HW_CACHE_RESULT_MAX}`
+		do
+			${MCEXEC} ./perf_test 2 1 ${PERF_TYPE_HW_CACHE} `expr ${k} \* 65536 + ${j} \* 256 + ${i}`
+		done
+	done
+done
+
diff --git a/test/perf_event/perf_test.c b/test/perf_event/perf_test.c
new file mode 100644
index 00000000..e57d94d4
--- /dev/null
+++ b/test/perf_event/perf_test.c
@@ -0,0 +1,82 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <langinfo.h>
+#include <locale.h>
+#include <sys/mman.h>
+
+#include "perftool.h"
+
+void
+usage()
+{
+	printf("Usage: perf_test mode cntr_num [<type> <config>]...\n");
+	exit(1);
+}
+
+int
+main(int argc, char** argv)
+{
+	long fds[PERF_CNTR_NUM];
+	long long counts[PERF_CNTR_NUM];
+	int types[PERF_CNTR_NUM];
+	int configs[PERF_CNTR_NUM];
+
+	int cntr_num;
+	int mode;
+	int i;
+
+	if (argc < 3) {
+		printf("ERROR: Too few arguments.\n");
+		usage();
+	}
+	
+	mode = atoi(argv[1]);
+	cntr_num = atoi(argv[2]);
+
+	/* perse args */
+	if (argc < cntr_num * 2 + 2) {
+		printf("ERROR: Too few arguments.\n");
+		usage();
+	}
+
+	for (i = 0; i < cntr_num; i++) {
+		types[i] = atoi(argv[i * 2 + 3]);
+		configs[i] = atoi(argv[i * 2 + 4]);
+	}
+
+	/* perf_event_open */
+	for (i = 0; i < cntr_num; i++) {
+		fds[i] = pe_opener(-1, mode, types[i], configs[i]);
+	}
+
+
+	/* perf_start */
+	for (i = 0; i < cntr_num; i++) {
+		PERF_BEGIN(fds[i]);
+	}
+
+	memory_task();
+	calc_task();
+
+	/* perf_end and read */
+	for (i = 0; i < cntr_num; i++) {
+		PERF_END(fds[i]);
+		PERF_READ(fds[i], counts[i]);
+		close(fds[i]);
+	}
+
+	/* print result */
+	for (i = 0; i < cntr_num; i++) {
+		if (types[i] == PERF_TYPE_HARDWARE) {
+			printf("%s,%ld\n", hw_event_names[configs[i]], counts[i]);
+		}
+		else if (types[i] == PERF_TYPE_HW_CACHE) {
+			printf("%s_%s_%s,%ld\n",
+			        id_names[(configs[i] >> 0) & 0xff],
+			        op_id_names[(configs[i] >> 8) & 0xff],
+			        result_names[(configs[i] >> 16) & 0xff],
+			        counts[i]);
+		}
+	}
+	return 0;
+}
diff --git a/test/perf_event/perftool.c b/test/perf_event/perftool.c
new file mode 100644
index 00000000..ccf0fbd0
--- /dev/null
+++ b/test/perf_event/perftool.c
@@ -0,0 +1,99 @@
+#include <stdlib.h>
+#include <stdio.h>
+#include <unistd.h>
+#include <string.h>
+#include <sys/ioctl.h>
+#include <linux/perf_event.h>
+#include <asm/unistd.h>
+
+//#include "perftool.h"
+
+#define WORKSIZE (1024 * 1024 * 32)
+#define LOOPSIZE 1000000
+#define REP 1000
+
+long
+perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
+		int cpu, int group_fd, unsigned long flags)
+{
+	int ret;
+
+	ret = syscall(__NR_perf_event_open, hw_event, pid, cpu,
+		group_fd, flags);
+	return ret;
+}
+
+long
+pe_opener(long group_fd, int mode, int type, unsigned long config)
+{
+	struct perf_event_attr pe;
+	int fd;
+	long long perf_hw_cache_id;
+	long long perf_hw_cache_op_id;
+	long long perf_hw_cache_op_result_id;
+
+	memset(&pe, 0, sizeof(struct perf_event_attr));
+	pe.type = type;
+	pe.size = sizeof(struct perf_event_attr);
+	pe.config = config;
+	pe.disabled = 1;
+	if (mode & 0x01) {
+		pe.exclude_user = 1; // EXCLUDE EVENTS THAT HAPPEN IN USER-SPACE
+	} else {
+		pe.exclude_user = 0; // INCLUDE EVENTS THAT HAPPEN IN USER-SPACE
+	}
+	if (mode & 0x02) {
+		pe.exclude_kernel = 1; // EXCUDE EVENTS THAT HAPPEN IN KERNEL-SPACE
+	} else {
+		pe.exclude_kernel = 0; // INCUDE EVENTS THAT HAPPEN IN KERNEL-SPACE
+	}
+	pe.exclude_hv = 1;
+	if (group_fd > 0) {
+		pe.read_format = PERF_FORMAT_GROUP;
+	}
+
+	fd = perf_event_open(&pe, 0, -1, group_fd, 0);
+
+	return fd;
+}
+
+long long
+hw_cache_build(long long id, long long op_id, long long op_result_id)
+{
+	return (id) | (op_id << 8) | (op_result_id << 16);
+}
+
+void
+memory_task()
+{
+	char* work = malloc(WORKSIZE);
+	char* fromaddr;
+	char* toaddr;
+	double r;
+	int offset;
+	int i;
+
+	for (i = 0; i < LOOPSIZE; i++) {
+		r = drand48();
+		offset = (int)(r * (double)WORKSIZE);
+		fromaddr = work + offset;
+		r = drand48();
+		offset = (int)(r * (double)WORKSIZE);
+		toaddr = work + offset;
+		*toaddr = *fromaddr;
+	}
+}
+
+void
+calc_task()
+{
+	int i, j;
+	double tmp;
+
+	for (i = 0; i < REP; i++) {
+		for (j = 0; j < REP; j++) {
+			tmp = drand48() * drand48();
+		}
+	}
+}
+	
diff --git a/test/perf_event/perftool.h b/test/perf_event/perftool.h
new file mode 100644
index 00000000..2fdce3df
--- /dev/null
+++ b/test/perf_event/perftool.h
@@ -0,0 +1,99 @@
+#ifndef __PERFTOOL_H__
+#define __PERFTOOL_H__
+
+#include <sys/ioctl.h>
+#include <linux/perf_event.h>
+#include <asm/unistd.h>
+
+extern long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
+				  int cpu, int group_fd, unsigned long flags);
+
+extern long perf_instructions();
+extern long perf_count_hw_cache_l1d_read_miss();
+extern long perf_count_hw_cache_dtlb_read_miss();
+extern long pe_opener(long group_fd, int mode, int type, unsigned long config);
+extern long long hw_cache_build(long long id, long long op_id, long long op_result_id);
+extern void memory_task();
+extern void calc_task();
+
+#define PERF_CNTR_NUM 4
+
+#define PERF_BEGIN(fd) \
+	if (fd > 0) ioctl(fd, PERF_EVENT_IOC_RESET, 0);	\
+	if (fd > 0) ioctl(fd, PERF_EVENT_IOC_ENABLE, 0);
+
+#define PERF_END(fd) \
+	if (fd > 0) ioctl(fd, PERF_EVENT_IOC_DISABLE, 0);
+
+#define PERF_READ(fd, count) \
+	if (fd > 0) read(fd, &count, sizeof(long long)); else count = -1;
+
+
+char* hw_event_names[] = {
+	"CPU_CYCLES             ",
+	"INSTRUCTIONS           ",
+	"CACHE_REFERENCES       ",
+	"CACHE_MISSES           ",
+	"BRANCH_INSTRUCTIONS    ",
+	"BRANCH_MISSES          ",
+	"BUS_CYCLES             ",
+	"STALLED_CYCLES_FRONTEND",
+	"STALLED_CYCLES_BACKEND ",
+	"REF_CPU_CYCLES         ",
+};
+
+char* id_names[] = {
+	"L1D ",
+	"L1I ",
+	"LL  ",
+	"DTLB",
+	"ITLB",
+	"BPU ",
+	"NODE",
+};
+	
+char* op_id_names[] = {
+	"OP_READ    ",
+	"OP_WRITE   ",
+	"OP_PREFETCH",
+};
+
+char* result_names[] = {
+	"ACCESS",
+	"MISS  ",
+};
+
+long long ids[] = {
+	PERF_COUNT_HW_CACHE_L1D,
+	PERF_COUNT_HW_CACHE_L1I,
+	PERF_COUNT_HW_CACHE_LL,
+	PERF_COUNT_HW_CACHE_DTLB,
+	PERF_COUNT_HW_CACHE_ITLB,
+	PERF_COUNT_HW_CACHE_BPU,
+	PERF_COUNT_HW_CACHE_NODE,
+};
+
+long long op_ids[] = {
+	PERF_COUNT_HW_CACHE_OP_READ,
+	PERF_COUNT_HW_CACHE_OP_WRITE,
+	PERF_COUNT_HW_CACHE_OP_PREFETCH,
+};
+
+long long result_ids[] = {
+	PERF_COUNT_HW_CACHE_RESULT_ACCESS,
+	PERF_COUNT_HW_CACHE_RESULT_MISS,
+};
+
+int hw_configs[] = {
+	PERF_COUNT_HW_CPU_CYCLES,
+	PERF_COUNT_HW_INSTRUCTIONS,
+	PERF_COUNT_HW_CACHE_REFERENCES,
+	PERF_COUNT_HW_CACHE_MISSES,
+	PERF_COUNT_HW_BRANCH_INSTRUCTIONS,
+	PERF_COUNT_HW_BRANCH_MISSES,
+	PERF_COUNT_HW_BUS_CYCLES,
+	PERF_COUNT_HW_STALLED_CYCLES_FRONTEND,
+	PERF_COUNT_HW_STALLED_CYCLES_BACKEND,
+	PERF_COUNT_HW_REF_CPU_CYCLES,
+};
+#endif