diff --git a/config.json b/config.json
index a8f6599..2d911fb 100644
--- a/config.json
+++ b/config.json
@@ -1,6 +1,7 @@
 {
-  "gop_cache": 0,
+  "gop_cache": true,
   "probe_timeout": 2000,
+  "mw_latency": 350,
 
   "rtmp": {
     "enable": true,
diff --git a/rtmp/rtmp_publisher.go b/rtmp/rtmp_publisher.go
index 2424a5c..7485fc7 100644
--- a/rtmp/rtmp_publisher.go
+++ b/rtmp/rtmp_publisher.go
@@ -30,10 +30,10 @@ func NewPublisher(sourceId string, stack *librtmp.Stack) *Publisher {
 
 func (p *Publisher) Init() {
 	//创建内存池
-	p.audioMemoryPool = stream.NewMemoryPool(48000 * (stream.AppConfig.GOPCache + 1))
-	if stream.AppConfig.GOPCache > 0 {
+	p.audioMemoryPool = stream.NewMemoryPool(48000 * 1)
+	if stream.AppConfig.GOPCache {
 		//以每秒钟4M码率大小创建内存池
-		p.videoMemoryPool = stream.NewMemoryPool(4096 * 1000 / 8 * stream.AppConfig.GOPCache)
+		p.videoMemoryPool = stream.NewMemoryPool(4096 * 1000)
 	} else {
 		p.videoMemoryPool = stream.NewMemoryPool(4096 * 1000 / 8)
 	}
@@ -77,7 +77,7 @@ func (p *Publisher) OnDeMuxStream(stream_ utils.AVStream) {
 func (p *Publisher) OnDeMuxPacket(packet utils.AVPacket) {
 	p.SourceImpl.OnDeMuxPacket(packet)
 
-	if stream.AppConfig.GOPCache > 0 {
+	if stream.AppConfig.GOPCache {
 		return
 	}
 
diff --git a/rtmp/rtmp_transtream.go b/rtmp/rtmp_transtream.go
index 273e853..797cc98 100644
--- a/rtmp/rtmp_transtream.go
+++ b/rtmp/rtmp_transtream.go
@@ -27,6 +27,7 @@ var nextOffset int
 
 func (t *TransStream) Input(packet utils.AVPacket) {
 	utils.Assert(t.TransStreamImpl.Completed)
+
 	var data []byte
 	var chunk *librtmp.Chunk
 	var videoPkt bool
@@ -47,6 +48,13 @@ func (t *TransStream) Input(packet utils.AVPacket) {
 		payloadSize += 5 + length
 	}
 
+	//即不开启GOP缓存又不合并发送. 直接使用AVPacket的预留头封装发送
+	if !stream.AppConfig.GOPCache && stream.AppConfig.MergeWriteLatency < 1 {
+		//首帧视频帧必须要
+	} else {
+
+	}
+
 	//payloadSize += payloadSize / t.chunkSize
 	//分配内存
 	t.memoryPool.Mark()
@@ -95,12 +103,12 @@ func (t *TransStream) Input(packet utils.AVPacket) {
 
 	rtmpData := t.memoryPool.Fetch()[:n]
 	ret := true
-	if stream.AppConfig.GOPCache > 0 {
+	if stream.AppConfig.GOPCache {
 		//ret = t.transBuffer.AddPacket(rtmpData, packet.KeyFrame() && videoPkt, packet.Dts())
 		ret = t.transBuffer.AddPacket(packet, packet.KeyFrame() && videoPkt, packet.Dts())
 	}
 
-	if !ret || stream.AppConfig.GOPCache < 1 {
+	if !ret || stream.AppConfig.GOPCache {
 		t.memoryPool.FreeTail()
 	}
 
@@ -197,7 +205,13 @@ func (t *TransStream) AddSink(sink stream.ISink) {
 
 	utils.Assert(t.headerSize > 0)
 	sink.Input(t.header[:t.headerSize])
+	if !stream.AppConfig.GOPCache {
+		return
+	}
 
+	//开启GOP缓存的情况下
+	//开启合并写的情况下:
+	// 如果合并写大小每满一次
 	// if stream.AppConfig.GOPCache > 0 {
 	// 	t.transBuffer.PeekAll(func(packet interface{}) {
 	// 		sink.Input(packet.([]byte))
@@ -238,9 +252,9 @@ func (t *TransStream) WriteHeader() error {
 	t.TransStreamImpl.Completed = true
 	t.header = make([]byte, 1024)
 	t.muxer = libflv.NewMuxer(audioCodecId, videoCodecId, 0, 0, 0)
-	t.memoryPool = stream.NewMemoryPoolWithRecopy(1024 * 1000 * (stream.AppConfig.GOPCache + 1))
-	if stream.AppConfig.GOPCache > 0 {
-		t.transBuffer = stream.NewStreamBuffer(int64(stream.AppConfig.GOPCache * 200))
+	t.memoryPool = stream.NewMemoryPoolWithRecopy(1024 * 4000)
+	if stream.AppConfig.GOPCache {
+		t.transBuffer = stream.NewStreamBuffer(200)
 		t.transBuffer.SetDiscardHandler(t.onDiscardPacket)
 	}
 
diff --git a/stream/config.go b/stream/config.go
index 055a006..287b107 100644
--- a/stream/config.go
+++ b/stream/config.go
@@ -1,5 +1,9 @@
 package stream
 
+const (
+	DefaultMergeWriteLatency = 350
+)
+
 type RtmpConfig struct {
 	Enable bool   `json:"enable"`
 	Addr   string `json:"addr"`
@@ -50,9 +54,11 @@ func init() {
 	AppConfig = AppConfig_{}
 }
 
+// AppConfig_ GOP缓存和合并写必须保持一致，同时开启或关闭. 关闭GOP缓存，是为了降低延迟，很难理解又另外开启合并写.
 type AppConfig_ struct {
-	GOPCache     int `json:"gop_cache"` //缓存GOP个数，不是时长
-	ProbeTimeout int `json:"probe_timeout"`
-	Rtmp         RtmpConfig
-	Hook         HookConfig
+	GOPCache          bool `json:"gop_cache"` //是否开启GOP缓存，只缓存一组音视频
+	ProbeTimeout      int  `json:"probe_timeout"`
+	MergeWriteLatency int  `json:"mw_latency"` //缓存指定时长的包，满了之后才发送给Sink. 可以降低用户态和内核态的交互，大幅提升性能.
+	Rtmp              RtmpConfig
+	Hook              HookConfig
 }
diff --git a/stream/memory_pool.go b/stream/memory_pool.go
index 45c9446..201b638 100644
--- a/stream/memory_pool.go
+++ b/stream/memory_pool.go
@@ -13,6 +13,10 @@ type MemoryPool interface {
 
 	Write(data []byte)
 
+	// Reserve 保留指定大小的内存空间
+	//主要是为了和实现和Write相似功能，但是不拷贝, 所以使用流程和Write一样.
+	Reserve(size int)
+
 	Allocate(size int) []byte
 
 	Fetch() []byte
@@ -117,6 +121,11 @@ func (m *memoryPool) Write(data []byte) {
 	copy(allocate, data)
 }
 
+func (m *memoryPool) Reserve(size int) {
+	utils.Assert(m.mark)
+	_ = m.allocate(size)
+}
+
 func (m *memoryPool) Allocate(size int) []byte {
 	utils.Assert(m.mark)
 	return m.allocate(size)
diff --git a/stream/sink.go b/stream/sink.go
index 7e69643..2b35cbb 100644
--- a/stream/sink.go
+++ b/stream/sink.go
@@ -66,11 +66,16 @@ type SinkImpl struct {
 	State_         SessionState
 	TransStreamId_ TransStreamId
 	disableVideo   bool
+
 	//Sink在请求拉流->Source推流->Sink断开整个阶段 是无锁线程安全
 	//如果Sink在等待队列-Sink断开，这个过程是非线程安全的
 	//SetState的时候，如果closed为true，返回false, 调用者自行删除sink
 	closed atomic.Bool
 
+	//HasSentKeyVideo 是否已经发送视频关键帧
+	//未开启GOP缓存的情况下，为避免播放花屏，发送的首个视频帧必须为关键帧
+	HasSentKeyVideo bool
+
 	DesiredAudioCodecId_ utils.AVCodecID
 	DesiredVideoCodecId_ utils.AVCodecID
 
diff --git a/stream/source.go b/stream/source.go
index a3cba46..be100ba 100644
--- a/stream/source.go
+++ b/stream/source.go
@@ -10,14 +10,18 @@ import (
 	"github.com/yangjiechina/live-server/transcode"
 )
 
-// SourceType Source 推流类型
+// SourceType 推流类型
 type SourceType byte
 
-// Protocol 输出协议
+// Protocol 输出的流协议
 type Protocol uint32
 
 type SourceEvent byte
 
+// SessionState 推拉流Session的状态
+// 包含握手和Hook授权阶段
+type SessionState uint32
+
 const (
 	SourceTypeRtmp  = SourceType(1)
 	SourceType28181 = SourceType(2)
@@ -35,11 +39,12 @@ const (
 	SourceEventPlayDone = SourceEvent(2)
 	SourceEventInput    = SourceEvent(3)
 	SourceEventClose    = SourceEvent(4)
-)
 
-// SessionState 推拉流Session状态
-// 包含, 握手阶段、Hook授权.
-type SessionState uint32
+	// TransMuxerHeaderMaxSize 传输流协议头的最大长度
+	// 在解析流分配AVPacket的Data时, 如果没有开启合并写, 提前预留指定长度的字节数量.
+	// 在封装传输流时, 直接在预留头中添加对应传输流的协议头，减少或免内存拷贝. 在传输flv以及转换AVCC和AnnexB格式时有显著提升.
+	TransMuxerHeaderMaxSize = 30
+)
 
 const (
 	SessionStateCreate           = SessionState(1)
@@ -167,6 +172,53 @@ func IsSupportMux(protocol Protocol, audioCodecId, videoCodecId utils.AVCodecID)
 	return true
 }
 
+// 分发每路Stream的Buffer给传输流
+// 按照时间戳升序发送
+func (s *SourceImpl) dispatchStreamBuffer(transStream ITransStream, streams []utils.AVStream) {
+	size := len(streams)
+	indexs := make([]int, size)
+
+	for {
+		min := int64(0xFFFFFFFF)
+
+		//找出最小的时间戳
+		for index, stream := range streams[:size] {
+			if s.buffers[stream.Index()].Size() == indexs[index] {
+				continue
+			}
+
+			pkt := s.buffers[stream.Index()].Peek(indexs[index]).(utils.AVPacket)
+			v := pkt.Dts()
+			if min == 0xFFFFFFFF {
+				min = v
+			} else if v < min {
+				min = v
+			}
+		}
+
+		if min == 0xFFFFFFFF {
+			break
+		}
+
+		for index, stream := range streams[:size] {
+			buffer := s.buffers[stream.Index()]
+			if buffer.Size() == indexs[index] {
+				continue
+			}
+
+			for i := indexs[index]; i < buffer.Size(); i++ {
+				packet := buffer.Peek(i).(utils.AVPacket)
+				if packet.Dts() > min {
+					break
+				}
+
+				transStream.Input(packet)
+				indexs[index]++
+			}
+		}
+	}
+}
+
 func (s *SourceImpl) AddSink(sink ISink) bool {
 	// 暂时不考虑多路视频流，意味着只能1路视频流和多路音频流，同理originStreams和allStreams里面的Stream互斥. 同时多路音频流的Codec必须一致
 	audioCodecId, videoCodecId := sink.DesiredAudioCodecId(), sink.DesiredVideoCodecId()
@@ -240,49 +292,8 @@ func (s *SourceImpl) AddSink(sink ISink) bool {
 		return false
 	}
 
-	if AppConfig.GOPCache > 0 && !ok {
-		indexs := make([]int, size)
-
-		for {
-			min := int64(0xFFFFFFFF)
-
-			for index, stream := range streams[:size] {
-				size := s.buffers[stream.Index()].Size()
-				if size == indexs[index] {
-					continue
-				}
-
-				pkt := s.buffers[stream.Index()].Peek(indexs[index]).(utils.AVPacket)
-				v := pkt.Dts()
-				if min == 0xFFFFFFFF {
-					min = v
-				} else if v < min {
-					min = v
-				}
-			}
-
-			if min == 0xFFFFFFFF {
-				break
-			}
-
-			for index, stream := range streams[:size] {
-				buffer := s.buffers[stream.Index()]
-				size := buffer.Size()
-				if size == indexs[index] {
-					continue
-				}
-
-				for i := indexs[index]; i < buffer.Size(); i++ {
-					packet := buffer.Peek(i).(utils.AVPacket)
-					if packet.Dts() > min {
-						break
-					}
-
-					transStream.Input(packet)
-					indexs[index]++
-				}
-			}
-		}
+	if AppConfig.GOPCache && !ok {
+		s.dispatchStreamBuffer(transStream, streams[:size])
 	}
 
 	return false
@@ -357,8 +368,8 @@ func (s *SourceImpl) OnDeMuxStream(stream utils.AVStream) (bool, StreamBuffer) {
 	}
 
 	//为每个Stream创建对应的Buffer
-	if AppConfig.GOPCache > 0 {
-		buffer := NewStreamBuffer(int64(AppConfig.GOPCache * 1000))
+	if AppConfig.GOPCache {
+		buffer := NewStreamBuffer(200)
 		//OnDeMuxStream的调用顺序，就是AVStream和AVPacket的Index的递增顺序
 		s.buffers = append(s.buffers, buffer)
 		return true, buffer
@@ -393,7 +404,7 @@ func (s *SourceImpl) OnDeMuxStreamDone() {
 }
 
 func (s *SourceImpl) OnDeMuxPacket(packet utils.AVPacket) {
-	if AppConfig.GOPCache > 0 {
+	if AppConfig.GOPCache {
 		buffer := s.buffers[packet.Index()]
 		buffer.AddPacket(packet, packet.KeyFrame(), packet.Dts())
 	}