diff --git a/trunk/src/app/srs_app_rtmp_conn.cpp b/trunk/src/app/srs_app_rtmp_conn.cpp
index 3ba916879..9c60430e2 100644
--- a/trunk/src/app/srs_app_rtmp_conn.cpp
+++ b/trunk/src/app/srs_app_rtmp_conn.cpp
@@ -49,6 +49,7 @@ using namespace std;
 #include <srs_protocol_msg_array.hpp>
 #include <srs_protocol_amf0.hpp>
 #include <srs_app_recv_thread.hpp>
+#include <srs_core_performance.hpp>
 
 // when stream is busy, for example, streaming is already
 // publishing, when a new client to request to publish,
@@ -591,7 +592,7 @@ int SrsRtmpConn::do_playing(SrsSource* source, SrsQueueRecvThread* trd)
         // no message to send, sleep a while.
         if (count <= 0) {
             srs_verbose("sleep for no messages to send");
-            st_usleep(SRS_CONSTS_RTMP_PULSE_TIMEOUT_US);
+            st_usleep(SRS_PERF_SEND_MSGS_CACHE * 1000);
         }
 
         // reportable
diff --git a/trunk/src/core/srs_core_performance.hpp b/trunk/src/core/srs_core_performance.hpp
index ef207c3d5..f07e2bd37 100644
--- a/trunk/src/core/srs_core_performance.hpp
+++ b/trunk/src/core/srs_core_performance.hpp
@@ -64,8 +64,20 @@ CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 *       buffer=65536B, small=4096B, sleep=780ms
 *       that is, when got nread bytes smaller than 4KB, sleep(780ms).
 */
-#define SRS_PERF_MERGED_READ
 #undef SRS_PERF_MERGED_READ
+#define SRS_PERF_MERGED_READ
+
+/**
+* the send cache time in ms.
+* to improve send performance, cache msgs and send in a time.
+* for example, cache 500ms videos and audios, then convert all these
+* msgs to iovecs, finally use writev to send.
+* @remark this largely improve performance, from 3.5k+ to 7.5k+.
+*       the latency+ when cache+.
+* @remark the socket send buffer default to 185KB, it large enough.
+* @see https://github.com/winlinvip/simple-rtmp-server/issues/194
+*/
+#define SRS_PERF_SEND_MSGS_CACHE 500
 
 #endif