Nagle算法簡介

Nagle算法主要是避免發送小的數據包，要求TCP連接上最多只能有一個未被確認的小分組，在該分組的確認到達之前不能發送其他的小分組。

在默認的情況下,Nagle算法是默認開啟的，Nagle算法比較適用于發送方發送大批量的小數據，并且接收方作出及時回應的場合，這樣可以降低包的傳輸個數。

但是如果你的程序是 write-write-read 模式，在使用了Nagle算法后，第二個 write 就會被推后一個RRT發送而造成一個很長的ack等待，從而產生一個延遲。為了避免這種情況，一般建議在應用層做緩沖，將兩個write合在一起，成為 write-read。

代碼分析

我們通過一個例子觀察下Nagle算法的延遲

• 服務器端代碼：recipes/tpc/nodelay_server.cc
• 客戶端代碼：recipes/tpc/nodelay.cc

客戶端代碼

// ...... 僅展示出部分代碼if (tcpnodelay){// 設置tcp TCP_NODELAYstream->setTcpNoDelay(true);printf("connected, set TCP_NODELAY\n");}else{stream->setTcpNoDelay(false);printf("connected\n");}for (int n = 0; n < num; ++n){printf("Request no. %d, sending %d bytes\n", n, len);if (buffering){std::vector<char> message(len + sizeof len, 'S');memcpy(message.data(), &len, sizeof len);int nw = stream->sendAll(message.data(), message.size());printf("%.6f sent %d bytes\n", now(), nw);}else{// 先發送頭 在發送數據 驗證Nagel算法stream->sendAll(&len, sizeof len);printf("%.6f sent header\n", now());usleep(1000); // prevent kernel merging TCP segmentsstd::string payload(len, 'S');int nw = stream->sendAll(payload.data(), payload.size());printf("%.6f sent %d bytes\n", now(), nw);}}

上面可以看出，正常下我們開啟Nagel，并發送header和data兩個數據包，我們還可以合并header和data，將其合成一個包發送，此外，我們還可以設置TCP_NODELAY選項。

測試

環境：兩臺橋聯的虛擬機
我們使用ping命令測試一下兩臺機器正常情況下的延遲

wang@localhost tpc]$ ping 192.168.1.104
PING 192.168.1.104 (192.168.1.104) 56(84) bytes of data.
64 bytes from 192.168.1.104: icmp_seq=1 ttl=64 time=4.30 ms
64 bytes from 192.168.1.104: icmp_seq=2 ttl=64 time=4.08 ms
64 bytes from 192.168.1.104: icmp_seq=3 ttl=64 time=4.26 ms
64 bytes from 192.168.1.104: icmp_seq=4 ttl=64 time=3.06 ms
64 bytes from 192.168.1.104: icmp_seq=5 ttl=64 time=3.77 ms
^C
--- 192.168.1.104 ping statistics ---
5 packets transmitted, 5 received, 0% packet loss, time 4016ms
rtt min/avg/max/mdev = 3.064/3.900/4.307/0.459 ms

大概在4ms左右

開啟Nagle發送兩個數據包

[wang@localhost tpc]$ ./nodelay 192.168.1.104 3210
connecting to 192.168.1.104:3210
connected
Request no. 0, sending 3210 bytes
1715777496.470708 sent header
1715777496.478341 sent 3210 bytes
Sent all 1 requests, receiving responses.
1715777496.483114 received 4 bytes, ack = 3210
total 0.013132 seconds

大概在13ms左右

關閉Nagle發送兩個數據包

[wang@localhost tpc]$ ./nodelay -D 192.168.1.104 3210
connecting to 192.168.1.104:3210
connected, set TCP_NODELAY
Request no. 0, sending 3210 bytes
1715777794.851212 sent header
1715777794.855121 sent 3210 bytes
Sent all 1 requests, receiving responses.
1715777794.856854 received 4 bytes, ack = 3210
total 0.006093 seconds

可以看見，關閉Nagel后，延遲大約6ms左右

開啟Nagel合并發送一個數據包

[wang@localhost tpc]$ ./nodelay -b 192.168.1.104 3210
connecting to 192.168.1.104:3210
connected
Request no. 0, sending 3210 bytes
1715778177.438170 sent 3214 bytes
Sent all 1 requests, receiving responses.
1715778177.441433 received 4 bytes, ack = 3210
total 0.004178 seconds

因為只發送一個，延遲跟ping差不多，大概4ms左右