diff --git a/src/select.cpp b/src/select.cpp
index 0fb1c5fb..96b1dd7b 100644
--- a/src/select.cpp
+++ b/src/select.cpp
@@ -401,16 +401,36 @@ zmq::select_t::fds_set_t::fds_set_t ()
 
 zmq::select_t::fds_set_t::fds_set_t (const fds_set_t& other_)
 {
+#if defined ZMQ_HAVE_WINDOWS
+    // On Windows we don't need to copy the whole fd_set.
+    // SOCKETS are continuous from the beginning of fd_array in fd_set.
+    // We just need to copy fd_count elements of fd_array.
+    // We gain huge memcpy() improvement if number of used SOCKETs is much lower than FD_SETSIZE.
+    memcpy (&read,  &other_.read,  (char *) (other_.read.fd_array  + other_.read.fd_count ) - (char *) &other_.read );
+    memcpy (&write, &other_.write, (char *) (other_.write.fd_array + other_.write.fd_count) - (char *) &other_.write);
+    memcpy (&error, &other_.error, (char *) (other_.error.fd_array + other_.error.fd_count) - (char *) &other_.error);
+#else
     memcpy (&read, &other_.read, sizeof other_.read);
     memcpy (&write, &other_.write, sizeof other_.write);
     memcpy (&error, &other_.error, sizeof other_.error);
+#endif
 }
 
 zmq::select_t::fds_set_t& zmq::select_t::fds_set_t::operator= (const fds_set_t& other_)
 {
+#if defined ZMQ_HAVE_WINDOWS
+    // On Windows we don't need to copy the whole fd_set.
+    // SOCKETS are continuous from the beginning of fd_array in fd_set.
+    // We just need to copy fd_count elements of fd_array.
+    // We gain huge memcpy() improvement if number of used SOCKETs is much lower than FD_SETSIZE.
+    memcpy (&read,  &other_.read,  (char *) (other_.read.fd_array  + other_.read.fd_count ) - (char *) &other_.read );
+    memcpy (&write, &other_.write, (char *) (other_.write.fd_array + other_.write.fd_count) - (char *) &other_.write);
+    memcpy (&error, &other_.error, (char *) (other_.error.fd_array + other_.error.fd_count) - (char *) &other_.error);
+#else
     memcpy (&read, &other_.read, sizeof other_.read);
     memcpy (&write, &other_.write, sizeof other_.write);
     memcpy (&error, &other_.error, sizeof other_.error);
+#endif
     return *this;
 }
 
diff --git a/src/zmq.cpp b/src/zmq.cpp
index 2cdda058..aa7893c6 100644
--- a/src/zmq.cpp
+++ b/src/zmq.cpp
@@ -1110,10 +1110,14 @@ int zmq_poll (zmq_pollitem_t *items_, int nitems_, long timeout_)
 
         //  Wait for events. Ignore interrupts if there's infinite timeout.
         while (true) {
-            memcpy (&inset, &pollset_in, sizeof (fd_set));
-            memcpy (&outset, &pollset_out, sizeof (fd_set));
-            memcpy (&errset, &pollset_err, sizeof (fd_set));
 #if defined ZMQ_HAVE_WINDOWS
+            // On Windows we don't need to copy the whole fd_set.
+            // SOCKETS are continuous from the beginning of fd_array in fd_set.
+            // We just need to copy fd_count elements of fd_array.
+            // We gain huge memcpy() improvement if number of used SOCKETs is much lower than FD_SETSIZE.
+            memcpy (&inset,  &pollset_in , (char *) (pollset_in.fd_array  + pollset_in.fd_count ) - (char *) &pollset_in );
+            memcpy (&outset, &pollset_out, (char *) (pollset_out.fd_array + pollset_out.fd_count) - (char *) &pollset_out);
+            memcpy (&errset, &pollset_err, (char *) (pollset_err.fd_array + pollset_err.fd_count) - (char *) &pollset_err);
             int rc = select (0, &inset, &outset, &errset, ptimeout);
             if (unlikely (rc == SOCKET_ERROR)) {
                 errno = zmq::wsa_error_to_errno (WSAGetLastError ());
@@ -1121,6 +1125,9 @@ int zmq_poll (zmq_pollitem_t *items_, int nitems_, long timeout_)
                 return -1;
             }
 #else
+            memcpy (&inset,  &pollset_in,  sizeof (fd_set));
+            memcpy (&outset, &pollset_out, sizeof (fd_set));
+            memcpy (&errset, &pollset_err, sizeof (fd_set));
             int rc = select (maxfd + 1, &inset, &outset, &errset, ptimeout);
             if (unlikely (rc == -1)) {
                 errno_assert (errno == EINTR || errno == EBADF);