diff --git a/docs/configuration.md b/docs/configuration.md
index f292bfbb7dcd6..984365748219f 100644
--- a/docs/configuration.md
+++ b/docs/configuration.md
@@ -864,6 +864,47 @@ Apart from these, the following properties are also available, and may be useful
between nodes leading to flooding the network with those.
+
+ spark.shuffle.io.mode |
+ NIO |
+
+ (Netty only) IO mode for channel implementation. If you set NIO, pure java socket is used for the channel.
+ Else if you set EPOLL, higher performance native socket is used through JNI.
+ |
+
+
+ spark.shuffle.io.backLog |
+ -1 |
+
+ (Netty only) Maximum length of the queue incomming connections.
+ If this value <= 0, the default length depending on your platform is used.
+ 200 is set for Windows, and 128 is set for others.
+ Especially, your platform has /proc/sys/net/core/somaxconn,
+ the value recorded in this file is preferred.
+ |
+
+
+ spark.shuffle.io.receiveBuffer |
+ -1 |
+
+ (Netty only) Receive buffer size for transportation.
+ If this value <= 0, the default value depending on your platform and
+ IO mode represented by spark.shuffle.io.mode is used.
+ The optimal size for receive buffer should be
+ latency * network_bandwidth. Assuming latency = 1ms, network_bandwidth = 10Gbps,
+ buffer size should be ~ 1.25MB.
+ |
+
+
+ spark.shuffle.io.sendBuffer |
+ -1 |
+
+ (Netty only) Send buffer size for transportation.
+ If this value <= 0, the default value is set as with spark.shuffle.io.receiveBuffer.
+ As well as receive buffer, the optimal size for send buffer should be
+ considered latency and network bandwidth.
+ |
+
spark.shuffle.io.preferDirectBufs |
true |