diff --git a/src/ae_kqueue.c b/src/ae_kqueue.c
index 8666b03cf..e8454245f 100644
--- a/src/ae_kqueue.c
+++ b/src/ae_kqueue.c
@@ -36,8 +36,29 @@
 typedef struct aeApiState {
     int kqfd;
     struct kevent *events;
+
+    /* Events mask for merge read and write event.
+     * To reduce memory consumption, we use 2 bits to store the mask
+     * of an event, so that 1 byte will store the mask of 4 events. */
+    char *eventsMask; 
 } aeApiState;
 
+#define EVENT_MASK_MALLOC_SIZE(sz) (((sz) + 3) / 4)
+#define EVENT_MASK_OFFSET(fd) ((fd) % 4 * 2)
+#define EVENT_MASK_ENCODE(fd, mask) (((mask) & 0x3) << EVENT_MASK_OFFSET(fd))
+
+static inline int getEventMask(const char *eventsMask, int fd) {
+    return (eventsMask[fd/4] >> EVENT_MASK_OFFSET(fd)) & 0x3;
+}
+
+static inline void addEventMask(char *eventsMask, int fd, int mask) {
+    eventsMask[fd/4] |= EVENT_MASK_ENCODE(fd, mask);
+}
+
+static inline void resetEventMask(char *eventsMask, int fd) {
+    eventsMask[fd/4] &= ~EVENT_MASK_ENCODE(fd, 0x3);
+}
+
 static int aeApiCreate(aeEventLoop *eventLoop) {
     aeApiState *state = zmalloc(sizeof(aeApiState));
 
@@ -54,6 +75,8 @@ static int aeApiCreate(aeEventLoop *eventLoop) {
         return -1;
     }
     anetCloexec(state->kqfd);
+    state->eventsMask = zmalloc(EVENT_MASK_MALLOC_SIZE(eventLoop->setsize));
+    memset(state->eventsMask, 0, EVENT_MASK_MALLOC_SIZE(eventLoop->setsize));
     eventLoop->apidata = state;
     return 0;
 }
@@ -62,6 +85,8 @@ static int aeApiResize(aeEventLoop *eventLoop, int setsize) {
     aeApiState *state = eventLoop->apidata;
 
     state->events = zrealloc(state->events, sizeof(struct kevent)*setsize);
+    state->eventsMask = zrealloc(state->eventsMask, EVENT_MASK_MALLOC_SIZE(setsize));
+    memset(state->eventsMask, 0, EVENT_MASK_MALLOC_SIZE(setsize));
     return 0;
 }
 
@@ -70,6 +95,7 @@ static void aeApiFree(aeEventLoop *eventLoop) {
 
     close(state->kqfd);
     zfree(state->events);
+    zfree(state->eventsMask);
     zfree(state);
 }
 
@@ -120,15 +146,37 @@ static int aeApiPoll(aeEventLoop *eventLoop, struct timeval *tvp) {
     if (retval > 0) {
         int j;
 
-        numevents = retval;
-        for(j = 0; j < numevents; j++) {
-            int mask = 0;
+        /* Normally we execute the read event first and then the write event.
+         * When the barrier is set, we will do it reverse.
+         * 
+         * However, under kqueue, read and write events would be separate
+         * events, which would make it impossible to control the order of
+         * reads and writes. So we store the event's mask we've got and merge
+         * the same fd events later. */
+        for (j = 0; j < retval; j++) {
             struct kevent *e = state->events+j;
+            int fd = e->ident;
+            int mask = 0; 
 
-            if (e->filter == EVFILT_READ) mask |= AE_READABLE;
-            if (e->filter == EVFILT_WRITE) mask |= AE_WRITABLE;
-            eventLoop->fired[j].fd = e->ident;
-            eventLoop->fired[j].mask = mask;
+            if (e->filter == EVFILT_READ) mask = AE_READABLE;
+            else if (e->filter == EVFILT_WRITE) mask = AE_WRITABLE;
+            addEventMask(state->eventsMask, fd, mask);
+        }
+
+        /* Re-traversal to merge read and write events, and set the fd's mask to
+         * 0 so that events are not added again when the fd is encountered again. */
+        numevents = 0;
+        for (j = 0; j < retval; j++) {
+            struct kevent *e = state->events+j;
+            int fd = e->ident;
+            int mask = getEventMask(state->eventsMask, fd);
+
+            if (mask) {
+                eventLoop->fired[numevents].fd = fd;
+                eventLoop->fired[numevents].mask = mask;
+                resetEventMask(state->eventsMask, fd);
+                numevents++;
+            }
         }
     } else if (retval == -1 && errno != EINTR) {
         panic("aeApiPoll: kevent, %s", strerror(errno));