User`s guide

Networking

oriented, pseudo-RAM device such as Xen Transcendent Memory (tmem) or in-kernel compressed

memory (zmem). (3.5)

• Safe swapping is supported using network block devices (NBDs) or NFS. (3.6)

1.1.10 Networking

• TCP controlled delay management (CoDel) is a new active queue management algorithm that is

designed to handle excessive buffering across a network connection (bufferbloat). The algorithm

is based on for how long packets are buffered in the queue rather than the size of the queue. If the

minimum queuing time rises above a threshold value, the algorithm discards packets and reduces the

transmission rate of TCP. (3.5)

• TCP connection repair implements process checkpointing and restart, which allows a TCP connection

to be stopped on one host and restarted on another host. Container virtualization can use this feature to

move a network connection between hosts. (3.5)

• TCP and STCP early retransmit allows fast retransmission (under certain conditions) to reduce the

number of duplicate acknowledgements. (3.5)

• TCP fast open (TFO) can speed up the opening of successive TCP connections between two endpoints

by eliminating one round time trip (RTT) from some TCP transactions. A performance improvement of

between 4 and 41% has been measured for web page loading.

TFO is not enabled by default. To enable it, use the following command:

# sysctl -w net.ipv4.tcp_fastopen=1

To make the change persist across system reboots, add the following entry to /etc/sysctl.conf:

net.ipv4.tcp_fastopen = 1

Applications that want to use TFO must notify the system using appropriate API calls, such as the

TCP_FASTOPEN option to setsockopt() on the server side or the MSG_FASTOPEN flag with sendto()

on the client side. (client side 3.6, server side 3.7)

• The TCP small queue algorithm is another mechanism intended to help deal with bufferbloat. The

algorithm limits the amount of data that can be queued for transmission by a socket. The limit is set by

/proc/sys/net/ipv4/tcp_limit_output_bytes, where the default value is 128 KB. To reduce

network latency, specify a lower value for this limit. (3.6)

1.1.11 Performance

• The slub slab allocator now implements wider lockless operations for most paths on CPU architectures

that support CMPXCHG (compare and exchange) instructions. This change can improve the performance

of slab intensive workloads. (3.1)

• The perf report --gtk command launches a simple GTK2-based performance report browser. (3.4)

• The perf annotate command now allows you to use the Enter key to trace recursively through

function calls in the TUI interface. (3.4)

• The perf record -b command supports a new hardware-based, branch-profiling feature on some

CPUs that allows you to examine branch execution. (3.4)

• Uprobes allow you to place a performance probe at any memory address in a user application so that

you can collect debugging and performance information non-disruptively. (3.5)