Category:Network Transmission Control Protocol (TCP) vistit wikipedia on TCP .

Congestion control

Is used to optimize the senders behaviour to the current network load.

other

Flow Control

links

Optimization

There are some parameters that can be set to greatly improve the Throughput in some networks such as high-bandwidth high-latency or high out-of-order-count . Bandwidth-delay product (BDP) and Windowsize are the keywords.

Most parameters can be set using ifconfig or editing /etc/sysctl.conf .

The linux kernel supports autotuning of the TCP-Window size. The variable must be 1 . The second parameter of tcp_wmem/tcp_rmem is adjustd by the autoconfig.

 #cat /proc/sys/net/ipv4/tcp_moderate_rcvbuf

Parameters

List Overview:

• MTU
• txqueuelen / netdev_max_backlog (tcp_backlog)
• tcp_mem, tcp_rmem, tcp_wmem
• vm/min_free_kbytes
• rmem_max / wmem_max

Additional Settings could be done via ethtool.

description

tcp_mem

/net/ipv4/tcp_mem difines the memory behavior of the memory allocated to tcp-Stacks. There are 3 values to be set, measured in memorypages (4kBytes on our hardware):

1. Low value: If the actual memory usage is below this point, nothing happens.
2. Start preasure: If usage climbs over this value the kernel starts to lower tcp-buffer sizes untill the Low value is reached.
3. Max value: Tcp streams get dropped no growth beneath this point.

tcp_rmem / tcp_wmem

/net/ipv4/tcp_rmem and /net/ipv4/tcp_wmem control the sizes of the recive and send buffers of a TCP-socket, measured in Bytes. Three values have to be set each. Tune tcp_mem accordingly to prevent a too aggressive memory-policy.

1. Minimal buffer size. Even under heavy load.
2. Default size. Overrides /proc/sys/net/core/rmem_default value . This value ist used to calculate the TCP-windowsize. In general this value ist autotuned by the kernel.
3. Maximum size that can be allocated to one socket.

configuration

Modify Parameters:

 #echo "4096 16384 4194304" > /proc/sys/net/ipv4/tcp_wmem

or add the following line to /etc/sysctl.conf for boot-time configuration:

 net/ipv4/tcp_wmem = 4096 16384 4194304

VLANs+Channelbonding

Refers to this special networklayout. As NFS uses TCP the nodes have to be optimized for maximum TCP-Throughput. As a result of round-robin the number of TCP-out-of-order-deliveries is higher than in general and as a packetloss won't happen that often, the TCP-window-size should be increased to prevent retransmissions of packets, which would slow down the connection.

websites

• TCP-tuning@wikipedia
• very good paper on high-bandwidth high-latency
• short overview
• Acticle on Gigabit tuning
• Some more parameters