5.4. TCP Window Scaling

The TCP Window Scaling feature, which is enabled by default in Red Hat Enterprise Linux, is an extension of the TCP protocol that significantly improves the throughput.

For example, on a 1 Gbps connection with 1.5 ms Round Trip Time (RTT):

One of the features TCP provides is flow control. With flow control, a sender can send as much data as the receiver can receive, but no more. To achieve this, the receiver advertises a window value, which is the amount of data a sender can send.

TCP originally supported window sizes up to 64 KiB, but at high Bandwidth Delay Products (BDP), this value becomes a restriction because the sender cannot send more than 64 KiB at a time. High-speed connections can transfer much more than 64 KiB of data at a given time. For example, a 10 Gbps link with 1 ms of latency between systems can have more than 1 MiB of data in transit at a given time. It would be inefficient if a host sends only 64 KiB, then pauses until the other host receives that 64 KiB.

To remove this bottleneck, the TCP Window Scaling extension allows the TCP window value to be arithmetically shifted left to increase the window size beyond 64 KiB. For example, the largest window value of 65535 shifted 7 places to the left, resulting in a window size of almost 8 MiB. This enables transferring much more data at a given time.

TCP Window Scaling is negotiated during the three-way TCP handshake that opens every TCP connection. Both sender and receiver must support TCP Window Scaling for the feature to work. If either or both participants do not advertise window scaling ability in their handshake, the connection reverts to using the original 16-bit TCP window size.

By default, TCP Window Scaling is enabled in Red Hat Enterprise Linux:

# sysctl net.ipv4.tcp_window_scaling
net.ipv4.tcp_window_scaling = 1

If TCP Window Scaling is disabled (0) on your server, revert the setting in the same way as you set it.