facebookincubator/katran

Katran is a C++ library and BPF program to build high-performance layer 4 load balancing forwarding plane. Katran leverages XDP infrastructure from the kernel to provide an in-kernel facility for fast packet's processing.

Katran's key features

1. Blazing fast (especially w/ XDP in driver mode).
2. Performance scaling linearly with a number of NIC's RX queues.
3. RSS friendly encapsulation.

See the detailed features' description below :

Documentation's sections

1. README (this file) - a generic overview of katran.
2. USAGE - the detailed description of how katran's library can be used, configured or changed.
3. DEVELOPING - description of how developing process looks like: how to build and test katran.
4. EXAMPLE - output of how to run provided examples (thrift and gRPC services, which leverage katran library)

Examples of usage

We provide simple examples of katran library usage w/ thrift and gRPC endpoints. Please refer to Examples for more detailed information.

Installation

Current distribution tested: Ubuntu 20.04

Requirements:

• Recent linux kernel (5.6+)
• Recent version of clang compiler (6.0+) For ubuntu if unsure run sudo apt install build-essential

build.sh

This script is a simple wrapper around getdeps.py (see next section). It requiers python3.6 or later in your PATH.

To install dependencies run: ./build.sh install

To build: ./build.sh <params>

If no flags are added then a default build is executed using the --allow-system-packages flag: ./build.sh

To do more complex stuff or understand how everything is build you can read the getdeps.py section and script.

getdeps.py

This script is used by many of Meta's OSS tools. It will download and build all of the necessary dependencies first, and will then invoke cmake etc to build katran. This will help ensure that you build with relevant versions of all of the dependent libraries, taking into account what versions are installed locally on your system.

It's written in python so you'll need python3.6 or later on your PATH. It works on Linux, macOS and Windows.

The settings for katran's cmake build are held in its getdeps manifest build/fbcode_builder/manifests/katran, which you can edit locally if desired.

This script will download and build all of the necessary dependencies first, and will then invoke cmake etc to build katran. This will help ensure that you build with relevant versions of all of the dependent libraries, taking into account what versions are installed locally on your system.

Currently requires python 3.6+ to be on your path.

python3 ./build/fbcode_builder/getdeps.py --allow-system-packages build katran

It puts output in its scratch area:

• installed/katran/lib/

You can also specify a --scratch-path argument to control the location of the scratch directory used for the build. You can find the default scratch install location from logs or with python3 ./build/fbcode_builder/getdeps.py show-inst-dir

There are also --install-dir and --install-prefix arguments to provide some more fine-grained control of the installation directories. However, given that katran provides no compatibility guarantees between commits we generally recommend building and installing the libraries to a temporary location, and then pointing your project's build at this temporary location, rather than installing katran in the traditional system installation directories. e.g., if you are building with CMake you can use the CMAKE_PREFIX_PATH variable to allow CMake to find katran in this temporary installation directory when building your project.

If you want to invoke cmake again to iterate, there is a helpful run_cmake.py script output in the scratch build directory. You can find the scratch build directory from logs or with python3 ./build/fbcode_builder/getdeps.py show-build-dir

build_katran.sh (to be deprecated)

To build and install katran library and thrift/gRPC examples - you need to run build_katran.sh script.

It should take care of all the required dependencies. If you need to build it for other Linux distributions, you need to make sure that you have installed: - folly - recent version of clang compiler (6.0+) - glog/gtest/gflags/elf libraries

Additionally, if you want to build examples, fbthrift and gRPC must be installed, as well!

While we run Meta's CI on CentOS, we do our best to support OSS build on recent Ubuntu versions. If you have any issues with build or need older Ubuntu versions - open github issue or better send us PR :)

Motivation behind layer 4 load balancing

Layer 4 load balancer (lb) enables to easily scale out Layer7 load balancers (the ones which terminate TCP sessions). Benefits of L4 lb over other techniques for scaling L7 lb is that it is :

1. compared to DNS it doesn't need to wait for TTL to redirect traffic from failed L7 lb.

2. compared to anycast-based solution, L4 lbs are more resilient to the networking related issues (which triggers mass ECMP reshuffle)/ enable better handling of adding/removing of L7 lbs from the pool as well as provide better support for unequal load balancing.

Environment requirements for katran to run

There are special requirements for katran to work properly. However, we do believe that most of the potential users of the library can easily satisfy them:

1. katran works only in DSR (direct service response) mode.

2. Network topology should be L3 based (everything above the top of the rack switch should be routed). This is because we are 'offloading' routing decision for sending packets to the real server to first routing device (by unconditionally sending all packets from katran there.)

3. katran doesn't support fragmentation (it cannot forward the fragmented packet, nor it can fragment them by itself if resulting packet size is bigger then MTU). This could be mitigated either by increasing MTU inside your network or by changing advertising TCP MSS from L7 lbs (this is recommended even if you have increased MTU, as it will prevent fragmentation related issues towards some of the client. For example, if instead of default TCP MSS 1460 (for ipv4) you will advertise 1450 - it will help clients behind PPPoE connection).

4. katran doesn't support packets w/ IP options set.

5. Maximum packet size cannot be bigger than 3.5k (and 1.5k by default).

6. katran is built with the assumption that it's going to be used in a "load balancer on a stick" scenario: where single interface would be used both for traffic "from user to L4 lb (ingress)" and "from L4 lb to L7 lb (egress)."

L4 load balancing network topology

Steps:

1. katran receives packet

2. Checks if the destination of the packet is configured as a VIP (virtual IP address - IP address of the service).

3. For an incoming packet toward a VIP - katran is checking if it saw packet from the same session before, and if it has - it sends the packet to the same real (actual server/l7 lb which then processes/terminates the TCP session).

4. If it's a new session - from 5 tuples in the packet, calculate a hash value.

5. Using this hash value - pick a real server.

6. Update session table with this lookup information so that katran can simply lookup this information for the next packet in the session and not calculate the hash again.

7. Encapsulate packet in another IP packet and send to the real.

L4 load balancing failure scenario

As we use only the data from the packet's headers to calculate a hash value, which is then used to pick a real server, different L4 lbs are consistent in real server selection, even w/o explicit state sharing amongst each other. This feature allows us to restart/drain single L4 lb w/o affecting TCP sessions, going to the L7 lbs.

katran's features description

1. Fast : katran uses XDP for packet forwarding, which allows to run packet handling routines right after packet has been received by network interface card (NIC) and before kernel had any chance to run (when XDP is working in "driver mode", katran supports "generic XDP" mode of operation as well (with some performance degradation compare to "driver mode")).

2. Performance scales linearly with a number of NIC's RX queues : The way XDP works is that it invokes BPF program on every received packet, and if your NIC has multiple queues, for each of them BPF program will be invoked independently. As katran is completely lockless and uses per-cpu versions of BPF maps - it scales linearly.

3. RSS friendly encapsulation : katran is uses ipip encapsulation for packet forwarding from L4 lb to L7 lb. However, to be able to work in conjunction with RSS on L7 lb receive side, instead of using the same source for every ipip packet, katran crafts a special one, in such a way, that different flows will have different outer (ipip) source IP, but packets in same flow will always have the same.

4. Fixed size (size is configurable on start) connection tracking table w/ LRU strategy for eviction of old entries.

5. Modified Maglev hashing for connections : It provides us good resiliency in case of failure and excellent load balancing features. The hashing was modified to be able to support unequal weights for backend (L7 lbs) servers

6. No need for busylooping on receive path : Your load balancer will barely consume any CPU if there is no traffic to serve.

7. katran (and XDP in general) allows you to run any application w/o any performance penalties on the same server (compare to some of other "kernel bypass" technologies)

Trivia

katran is named (it's actually an alternative name) after spiny dogfish shark.

License

katran is GPL-2.0 licensed, as found in the LICENSE file.