search

L4STeam / linux

Kernel tree containing patches for TCP Prague and the dualpi2 qdisc
Other
52 stars 16 forks source link

readme

Linux kernel tree with L4S patches

This linux kernel repository contains the various patches developed in the context of the L4S experiment.

Namely:

Installation (debian derivatives)

wget https://github.com/L4STeam/linux/releases/download/testing-build/l4s-testing.zip
unzip l4s-testing.zip
sudo dpkg --install debian_build/*
sudo update-grub  # This should auto-detect the new kernel
# You can optionally set newly installed kernel as the default, e.g., editing GRUB_DEFAULT in /etc/default/grub
# You can now reboot (and may have to manually select the kernel in grub)
# Be sure that the newly installed kernel is successfully used, e.g., checking output of
uname -r
# Be sure to ensure the required modules are loaded before doing experiments, e.g.,
sudo modprobe sch_dualpi2
sudo modprobe tcp_prague

This branch (testing)

This branch accumulates all patches into a single kernel tree, in order to ease up testing.

You can grab a pre-built debian archive of the kernel image and headers through the latest actions artifacts. The tip of the master branch is also always build/packaged (alongside iproute2) and attached as pre-release artifact for the testing-build tag.

Compilation

Compile it as any kernel, enabling the dualpi2 AQM and TCP Prague in the config.

Assuming you compile this on a similar machine that where you intend to run the kernel (e.g., architecture, distribution, ...):

# Try to use existing kernel config
if [ -f /proc/config.gz ]; then
    zcat proc/config.gz > .config
    make olddefconfig
else if [ -f "/boot/config-$(uname -r)" ]; then
    cp "/boot/config-$(uname -r)" .config
    make olddefconfig
else
    make defconfig
fi

# Enable TCP Prague and dualpi2
scripts/config -m TCP_CONG_PRAGUE
scripts/config -m NET_SCH_DUALPI2
# Optionally enable DCTCP and BBR v2
scripts/config -m TCP_CONG_DCTCP
scripts/config -m TCP_CONG_BBR2

# Build the kernel
make -j$(nproc) LOCALVERSION=-prague-1
# Alternatively, you can generate *.deb with
# BUILD_NUMBER=${BUILD_NUMBER:-1} make \
#   -j$(nproc) bindeb-pkg \
#   LOCALVERSION=-prague-${BUILD_NUMBER} \
#   KDEB_PKGVERSION=1
# see the output of `make help` to generate rpms/...

# Install it on the current system if applicable
make install
make modules_install

# Update your bootloader to list the new kernel
update-grub
# You may then want to udpate the GRUB_DEFAULT variable
# in /etc/default/grub to the newly installed kernel

If you intend to use non-default parameters for dualpi2, make sure to also build the patched iproute2 , e.g.,

git clone https://github.com/L4STeam/iproute2.git && cd iproute2
./configure
make
tc/tc qdisc replace dev eth0 root dualpi2 ...
# You can optionally install (!potentially overwrite) the new
# iproute2 utils with `make install`

Performing experiments

While dualpi2 can work with DCTCP, DCTCP suffers from a few unfortunate interactions with GSO/pacing/..., resulting in under-utilization.

As a result, we advice you to use tcp_prague which currently has basic fixes to those limitations. Note that this might still under-perform in heavily virtualized settings, as scheduling becomes less reliable.

# some preparations for having better paced traffic and reduce bursts for each network interface $NETIF that sends L4S traffic
# Avoid processing 64K packets in the kernel, which will send those packets in a burst independent of the pacing (lro only for newer NICS and kernels that support it):
sudo ethtool -K $NETIF tso off gso off gro off lro off
# fq qdisc needs to be configured on clients and server NICS (instead of fq_codel; fq is the only one that supports the pacing)
sudo tc qdisc replace dev $NETIF root handle 1: fq limit 20480 flow_limit 10240
# Enable Accurate ECN (only needed for BBR2 and DCTCP, not needed for Prague)
sysctl -w net.ipv4.tcp_ecn=3
# set Prague congestion control system wide (or in the application with socket options)
sysctl -w net.ipv4.tcp_congestion_control=prague

Accurate ECN negotiation

Prague attempts to negotiate Accurate ECN automatically. Note that, at the moment, Accurate ECN must be enabled on both ends of a connection in order it with DCTCP or BBR v2.

Among 3 different congestion control algorthms (net.ipv4.tcp_congestion_control=prague/bbr2/cubic) and 4 differnet ECN configurations (net.ipv4.tcp_ecn=3/1/2/0), the negotiated ECN modes between server and client are as follows: Note that the negotiated ECN modes can be either Accurate ECN, Classic ECN, or Non-ECN. image

In addition, the following congestion control algorithm and ECN fields of IP header (ECN(1)/ECN(0)/Not ECT) is used when server as Data Sender. Note thtat prague-reno is the current fallback mode of TCP Prague. image

Also, when client is Data Sender, the following congestion cotnrol algorithm and ECN fields of IP header is applied. image